ARTIFICIAL INTELLIGENCE IN ASSET MANAGEMENT

CFA INSTITUTE RESEARCH FOUNDATION / LITERATURE REVIEW

SÖHNKE M. BARTRAM,

JÜRGEN BRANKE, AND

MEHRSHAD MOTAHARI

ARTIFICIAL INTELLIGENCE

IN ASSET MANAGEMENT

BARTRAM, BRANKE, AND MOTAHARILITERATURE REVIEW / ARTIFICIAL INTELLIGENCE IN ASSET MANAGEMENT

ARTIFICIAL INTELLIGENCE

IN ASSET MANAGEMENT

Söhnke M. Bartram, Jürgen Branke,

and Mehrshad Motahari

Research

Foundation

Literature Review

Statement of Purpose

The CFA Institute Research Foundation is a not-

for-proﬁt organization established to promote

the development and dissemination of relevant

research for investment practitioners worldwide.

Neither CFA Institute Research Foundation, CFA Institute, nor the

publication’s editorial staﬀ is responsible for facts and opinions

presented in this publication. This publication reﬂects the views of

the author(s) and does not represent the oﬃcial views of CFA Institute

Research Foundation.

CFA®, Chartered Financial Analyst®, and GIPS® are just a few of the trademarks owned

by CFA Institute. To view a list of CFA Institute trademarks and the Guide for the Use of

CFA Institute Marks, please visit our website at www.cfainstitute.org.

No part of this publication may be reproduced, stored in a retrieval system, or transmitted,

in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise,

without the prior written permission of the copyright holder.

is publication is designed to provide accurate and authoritative information in regard

to the subject matter covered. It is sold with the understanding that the publisher is not

engaged in rendering legal, accounting, or other professional service. If legal advice or other

expert assistance is required, the services of a competent professional should be sought.

Cover photo credit: nuchao / iStock / Getty Images Plus

ISBN 978-1-952927-02-7

Acknowledgements

Helpful comments and suggestions by Florian Bardong (SysAMI Advisors),

Gurvinder Brar (Macquarie), Marie Brière (Amundi), Charles Cara

(Absolute Strategy), Carmine De Franco (Ossiam), Giuliano De Rossi

(Goldman Sachs), Marco Dion (Qube Research and Technologies), Kevin

Endler (ACATIS), Daniel Giamouridis (Bank of America Merrill Lynch),

Alex Gracian (Resolute Investments), Farouk Jivraj (Barclays), Bryan Kelly

(AQR), Petter Kolm, Alexei Kondratyev (Standard Chartered), Christos

Koutsoyannis (Atlas Ridge Capital), Jörg Ladwein (Allianz Investment

Management), Ke Lu, Jon Lukomnik (Sinclair Capital), Spyros Mesomeris

(UBS), Matt Monach (Aberdeen Standard Investments), Andreas Neuhierl,

Raghavendra Rau, Berkan Sesen, Maximilian Stroh (Invesco), Scott Taylor

(AIG), Simon Taylor, Argyris Tsiaras, Nir Vulkan (Oxford Man Institute),

Markos Zachariadis, Riccardo Zecchinelli (UK Cabinet Oce), and seminar

participants at 13th Financial Risks International Forum, 2020 CERF in the

City Conference, 2020 WBS Investment Challenge, Barclays Quantitative

Investment Strategies (QIS) group, Cambridge Judge Business School,

the 13th Financial Risks International Forum, and the 2020 Paris Conference

on FinTech and Cryptonance are gratefully acknowledged. Söhnke Bartram

gratefully acknowledges the warm hospitality of Cambridge University,

Fudan University, and Oxford University.

The CFA Institute

Research Foundation

Board of Trustees

2019–2020

Chair

Ted Aronson, CFA

AJO

Heather Brilliant, CFA

Diamond Hill

Margaret Franklin, CFA

CFA Institute

Bill Fung, PhD

Aventura, FL

Daniel Gamba, CFA

BlackRock

JT Grier, CFA*

Virginia Retirement

System

Vice Chair

Joanne Hill

CBOE Vest Financial

Roger Ibbotson*

Yale School of Management

Joachim Klement, CFA

Independent

Vikram Kuriyan, PhD, CFA

GWA and Indian School

of Business

Aaron Low, CFA

LUMIQ

Mauro Miranda, CFA

Panda Investimentos

AAI Ltda.

Lotta Moberg, PhD, CFA

William Blair

Sophie Palmer, CFA

Jarislowsky Fraser

Dave Uduanu, CFA

Sigma Pensions Ltd

Ofcers and Directors

Executive Director

Bud Haslett, CFA

CFA Institute

Gary P. Brinson Director of Research

Laurence B. Siegel

Blue Moon Communications

Associate Research Director

Luis Garcia-Feijóo, CFA, CIPM

Coral Gables, Florida

Secretary

Jessica Lawson

CFA Institute

Treasurer

Kim Maynard

CFA Institute

Research Foundation Review Board

*Emeritus

William J. Bernstein

Efﬁcient Frontier

Advisors

Elroy Dimson

London Business School

Stephen Figlewski

New York University

William N. Goetzmann

Yale School of

Management

Elizabeth R. Hilpman

Barlow Partners, Inc.

Paul D. Kaplan, CFA

Morningstar, Inc.

Robert E. Kiernan III

Advanced Portfolio

Management

Andrew W. Lo

Massachusetts Institute

of Technology

Alan Marcus

Boston College

Paul O’Connell

FDO Partners

Krishna Ramaswamy

University of

Pennsylvania

Andrew Rudd

Advisor Software, Inc.

Stephen Sexauer

Allianz Global Investors

Solutions

Lee R. Thomas

Paciﬁc Investment

Management Company

Contents

1. Introduction ........................................................................................ 2

2. Trends in Artiﬁcial Intelligence ............................................................ 4

3. Portfolio Management ....................................................................... 8

3.1. Alpha and Sigma .......................................................................... 8

3.2. Portfolio Optimization ................................................................. 12

4. Trading ................................................................................................ 14

4.1. Algorithmic Trading ..................................................................... 15

4.2. Transaction Cost Analysis ............................................................ 17

4.3. Trade Execution ........................................................................... 18

5. Portfolio Risk Management ................................................................ 20

5.1. Market Risk ................................................................................... 20

5.2. Credit Risk .................................................................................... 22

6. Robo-Advisors .................................................................................... 24

7. Artiﬁcial Intelligence Risks and Challenges: What Can Go Wrong? ..... 26

8. Conclusion .......................................................................................... 29

Appendix A. Basic Artiﬁcial Intelligence Concepts and Techniques ....... 30

A.1. Artiﬁcial Intelligence and Machine Learning ............................... 30

A.1.1. Origin and Deﬁnition ............................................................ 30

A.1.2. Supervised Learning ............................................................. 31

A.1.3. Unsupervised Learning ......................................................... 32

A.1.4. Reinforcement Learning ....................................................... 32

A.2. Overview of Common Artiﬁcial Intelligence Techniques ............ 32

A.2.1. Least Absolute Shrinkage and Selection Operator

Regression ................................................................................. 32

A.2.2. Artiﬁcial Neural Networks and Deep Learning .................... 34

A.2.3. Decision Trees and Random Forests .................................... 35

A.2.4. Support Vector Machines ..................................................... 36

A.2.5. Cluster Analysis .................................................................... 37

A.2.6. Evolutionary (Genetic) Algorithms ....................................... 37

A.2.7. Natural Language Processing ............................................... 38

A.2.8. Comparisons of AI Techniques ............................................. 39

Appendix B. Trends and Patterns in Finance Research Using AI ............. 41

References .............................................................................................. 45

is publication qualies for 1.5 PL credits under the guide-

lines of the CFA Institute Professional Learning Program.

Articial Intelligence in Asset Management

Söhnke M. Bartram

Research Fellow, Centre for Economic Policy Research, and Professor of Finance,

University of Warwick, Warwick Business School, Department of Finance

Jürgen Branke

Professor of Operational Research and Systems, University of Warwick,

Warwick Business School

Mehrshad Motahari

Research Associate, Cambridge Centre for Finance and Cambridge Endowment

for Research in Finance, University of Cambridge, Cambridge Judge Business School

1. Introduction

Articial intelligence (AI) is one of the hottest topics of current times because

it has disrupted most industries in recent years, and the nancial services sec-

tor is no exception. With the advent of ntech, which has a particular empha-

sis on AI, the sector has experienced a revolution in some of its core practices.

Probably the most aected area is asset management, which is expected to

suer the largest number of job cuts in the near future (Buchanan 2019).

A sizable proportion of asset management companies are now using AI and

statistical models to run trading and investment platforms. e increased use

of AI across a range of tasks in asset management calls for a more systematic

examination of the various techniques and applications involved, as well as

the concomitant opportunities and challenges they bring to the sector.

is study provides a comprehensive overview of a wide range of existing

and emerging applications of AI in asset management, highlighting the key

topics of debate. We focus on three major areas: portfolio management, trad-

ing, and portfolio risk management. Portfolio management entails making

asset allocation decisions to construct a portfolio with specic risk and return

characteristics. AI techniques can contribute to this process by facilitating

fundamental analysis through quantitative or textual data analysis and gen-

erating novel investment strategies. AI techniques can also help improve the

shortcomings of classical portfolio construction techniques. In particular, AI

can produce better asset return and risk estimates and solve portfolio opti-

mization problems with complex constraints, yielding portfolios with better

out-of-sample performance compared with traditional approaches.

Trading is another popular area for AI applications. Considering the

growing speed and complexity of trades, AI techniques are becoming an

essential part of trading practice. A particularly attractive feature of AI is

its ability to process large amounts of data to generate trading signals.

Algorithms can be trained to automatically execute trades based on these sig-

nals, which has given rise to the industry of algorithmic (or algo) trading. In

addition, AI techniques can reduce transaction costs by automatically analyz-

ing the market and subsequently identifying the best time, size, and venue

for trades.

AI also has vast implications for portfolio risk management. Since the

2008 global nancial crisis, risk management and compliance have been at

the forefront of asset management practices. With nancial assets and global

markets becoming increasingly complex, traditional risk models may no lon-

ger be sucient for risk analysis. At the same time, AI techniques that learn

1. Introduction

and evolve by using data can provide additional tools for monitoring risk.

Specically, AI assists risk managers in validating and backtesting risk mod-

els. AI approaches can also extract information more eciently from various

sources of structured or unstructured data and generate more accurate fore-

casts of bankruptcy and credit risk, market volatility, macroeconomic trends,

nancial crises, and so on than traditional techniques.

Furthermore, robo-advising has gained signicant public interest in

recent years. Robo-advisors are computer programs that provide digital nan-

cial investment advice based on mathematical rules or algorithms tailored to

investors’ needs and preferences. e popularity of robo-advisors stems from

their success in democratizing investment advisory services by making them

cheaper and more accessible to unsophisticated individual investors. Robo-

advisors are particularly attractive to young and tech-savvy investors, such

as Generation Y (millennials). AI is the backbone of typical robo-advising

algorithms, which rely heavily on the application of AI across all dimensions

of asset management.

We also discuss a number of possible disadvantages of using AI in asset

management. AI models are often opaque and complex, making them dif-

cult for managers to monitor and scrutinize. e models’ reliance on and

sensitivity to data can introduce a considerable source of risk. AI models can

be improperly trained as a result of using poor-quality or insucient data.

Ineective human supervision might lead to systematic crashes, an inability

to identify inference errors, and a lack of understanding of investment prac-

tices and performance attribution by investors. Lastly, whether the benets

associated with AI can justify its considerable development and implementa-

tion cost is unclear.

e remainder of the piece is organized as follows. Section 2 provides

an overview of trends in AI and of the most common AI techniques used

in asset management. AI applications in portfolio management, trading, and

portfolio risk management are discussed in Sections 3, 4, and 5, respectively.

Section 6 covers the use of AI in robo-advising, and Section 7 discusses some

of the risks and concerns associated with AI. Section 8 concludes with a sum-

mary of the main takeaways.

2. Trends in Articial Intelligence

In recent years, the popularity of AI in general—and of machine learning

(ML) specically—has surged in both practice and academia. Consequently,

the number of research papers published with the keywords “articial intel-

ligence” and “machine learning” has increased dramatically in the past ve

years (Figure 1). AI is a broader concept than ML, because it refers to the

general use of computers to imitate human cognitive functions. ML is eec-

tively a subset of AI, in which machines are able to decide and perform

actions based on past experiences. To date, AI applications in nance mostly

make use of ML techniques, such as statistical learning, and thus the AI label

applies only in a very broad sense (e.g., Gu, Kelly, and Xiu 2020). Moreover,

a large part of what is branded as AI (or ML) in nance is not new but has

existed in the form of statistical or econometric modeling techniques for a

long time.

e recent hype about AI can be attributed to three developments that

are not necessarily related to the science of AI itself (Giamouridis 2017).

First, computer processing and storage capacity have improved remarkably in

the past decade, making the use of some longstanding AI techniques feasible.

Figure 1. Number of Published Research Papers by Topic over Time, 1996–2018

Number of Papers

200

180

160

140

120

100

96 18020098 04 06 10 12 1608 14

Artificial Intelligence

Machine Learning

Notes: e gure presents the number of published papers with specic keywords by year, as

reported by the Scopus database. e sample starts in 1996, ends in 2018, and includes papers hav-

ing either “articial intelligence” or “machine learning” in their abstract, title, or keyword section.

2. Trends in Articial Intelligence

Second, the volume and breadth of data that can be used to train AI models

have increased substantially. Lastly, AI algorithms have been improved and

become widely accessible, allowing for their use in many cases without the

need for expert computer science knowledge. All these factors have contrib-

uted to the popularity of AI and ML as research topics in social sciences.

Although AI is a broad eld that entails a range of approaches devel-

oped over time, the recent interest in AI is almost entirely centered on ML,

which is by far the most popular AI approach to date. ML is concerned with

using data progressively to adapt the parameters of statistical, probabilistic,

and other computing models. It essentially automates one or several stages of

information processing. Although an extensive list of techniques can accom-

plish this automation, most ML applications in asset management, and even

in nance more generally, rely on a number of major (classes of) techniques

(Figure 2). ese include articial neural networks (ANNs), cluster analy-

sis, decision trees and random forests, evolutionary (genetic) algorithms, least

absolute shrinkage and selection operator (LASSO), support vector machines

(SVMs), and natural language processing (NLP). Appendix A provides a

detailed, more technical description of each of these techniques.

e interest of academic research in specic AI techniques has steadily

increased in the past two decades, as illustrated by the number of published

papers on the subject (Figure 3). Some of these techniques, such as evolu-

tionary algorithms or neural networks, were established research topics long

before ML gained popularity. On the other hand, SVMs and NLP have

gained interest more recently. Neural network, random forest, and NLP tech-

niques have experienced the sharpest increase in their mention in published

papers during the past ve years. Appendix B provides a more detailed view

of the use of AI techniques in nance research based on analyzing all work-

ing papers posted on SSRN. e following sections discuss these techniques

and their applications in the context of asset management.

Articial Intelligence in Asset Management

Figure 2. Summary of Key AI/ML Techniques

• Ordinary regression model with an additional penalty term that ensures

choosing the smallest necessary subset of explanatory variables

• Reduces spurious coefficient estimates to zero, which significantly

enhances the out-of-sample performance of the model

• Typical application: Forecasting

LASSO

• Clusters data into groups so that the units in each group have similar

characteristics

• The number of clusters can be defined by the user or determined

automatically by the algorithm

• Typical application: Asset classification

Cluster

Analysis

• Optimization technique capable of searching through large, complex,

nonlinear sets of solutions, identifying those that are preferred

• Process inspired by natural evolution

• Typical application: Variants of portfolio optimization that cannot be

solved with classical optimization algorithms

Evolutionary

(Genetic)

Algorithms

• Nonlinear regression model

• Network of connected nodes that loosely model neurons in a brain

• Receives a training set of input and desired output data pairs and is able

to learn the relationship between them

• Can then be used to predict the output of previously unseen inputs

• Typical application: Forecasting

Artificial

Neural

Networks

• A decision tree classifies units based on their features

• Classification is done by traversing a logical tree from root to leaves, at

each branch moving left or right depending on the unit's features; such

trees can be interpreted by humans

• Constructed automatically based on training set of input and desired

output pairs

• Random forests simply average the outputs of several decision tree

models in order to produce more reliable forecasts

• Typical application: Classification and forecasting

Decision

Trees and

Random

Forests

• Can be used for classification or regression

• Can handle nonlinear relationships by mapping the inputs to a

higher-dimensional space

• Faster to train than artificial neural networks

• Typical application: Forecasting

Support

Vector

Machines

• Range of techniques used to process natural language data (e.g., textual, audio)

• Particularly useful for extracting information from textual media

(e.g., social media, websites, news articles)

• Typical application: Automatic analysis of corporate annual reports

and news articles

Natural

Language

Processing

Note: e gure lists and describes major AI techniques commonly used in asset management.

2. Trends in Articial Intelligence

Figure 3. Number of Published Papers by AI Technique over Time, 1996–2018

Neural Network or Deep Learning

Support Vector Machine

Cluster Analysis Random Forest or Decision Tree

Genetic (or Evolutionary) Algorithm LASSO

Natural Language Processing

Number of Papers

200

180

160

140

120

100

020098 04 06 10 12 1608 14

Notes: e gure presents the number of published papers with specic keywords by year. It is

based on the number of published papers listed on Scopus starting in 1996 and ending in 2018.

e papers have “nance” and/or “asset management” keywords together with at least one of the

following keywords: “cluster analysis,” “genetic algorithm” or “evolutionary algorithm,” “lasso,”

“natural language processing,” “neural network” or “deep learning,” “random forest” or “decision

tree,” and “support vector machine.”

3. Portfolio Management

AI techniques can be used to perform sophisticated fundamental analysis,

including the use of text analysis, and to optimize asset allocations in nan-

cial portfolios. Amid various challenges of conventional portfolio optimiza-

tion approaches, AI techniques often provide better estimates of returns and

covariances than more conventional methods do. ese estimates can then be

used within traditional portfolio optimization frameworks. Moreover, AI can

be used directly for asset allocation decisions to construct portfolios that meet

performance targets more closely than portfolios created using traditional

methods (Figure 4).

3.1. Alpha and Sigma

Fundamental analysis can be considered the cornerstone of portfolio man-

agement and can be facilitated signicantly by AI (Table 1). Arguably the

most signicant application of AI in fundamental analysis is textual analy-

sis (Das 2014; Kearney and Liu 2014; Fisher, Garnsey, and Hughes 2016).

Figure 4. AI in Portfolio Management

Output

Portfolio

Portfolio Optimization

- Genetic algorithms can solve

optimization problems under

complex constraints (e.g.,

cardinality, additional objectives)

- Neural networks can be used

to produce optimal portfolios

directly or portfolios that mimic

an index with a small set of assets

Expected Returns

- AI approaches can produce

more accurate estimates of

expected returns (e.g., LASSO,

neural networks, support

vector machines)

Variances/Covariances

- AI can provide better

estimates of variances and

covariances (e.g., neural

networks, support vector

machines)

- The covariance matrix

structure can be replaced

with a tree structure using

hierarchical clustering

Notes: e gure presents a summary of how AI can be incorporated into portfolio construction.

AI approaches can provide the inputs (i.e., expected returns, variance/covariance, and asset views)

and use them in asset allocation to meet portfolio managers’ targets.

3. Portfolio Management

Table 1. AI and Fundamental Analysis

Techn ique Study Sample/Data

ANNs Atsalakis and Valavanis

(2009)

No empirical work; surveys other studies

Lam (2004) Financial data for 364 S&P 500 Index companies

from 1985 to 1995

Ballings, Van den Poel,

Hespeels, and Gryp (2015)

Financial data for 5,767 listed European rms from

2009 to 2010

Cluster

Analysis

Ballings et al. (2015) Financial data for 5,767 listed European rms from

2009 to 2010

Decision

Trees

Ballings et al. (2015) Financial data for 5,767 listed European rms from

2009 to 2010

Bryzgalova, Pelger, and

Zhu (2019)

Financial data for all US rms available on CRSP

from 1964 to 2016

Genetic

Algorithms

Hu, K. Liu, Zhang, Su,

Ngai, and M. Liu (2015)

No empirical work; surveys other studies

Hybrid/

Ensemble

Li, Huang, Deng, and Zhu

(2014)

Stock data for all HKEX-listed rms in year 2001

Huang (2012) Financial data for 200 stocks listed on the Taiwan

Stock Exchange from 1996 to 2010

LASSO Feng et al. (2017) NYSE, AMEX, and NASDAQ stock data from

1976 to 2017

NLP Leung and Ton (2015) Stock data for 2,000 rms listed on the Australian

Securities Exchange (ASX) from 2003 to 2008

Sprenger et al. (2014) 400,000 stock-related Twitter messages and S&P

500 stock prices for 2010

Schumaker and Chen

(2006)

9,211 nancial news articles and 10,259,042 stock

quotes for a ve-week period in 2005

SVMs Han and Chen (2007) Financial statement data for 251 stocks listed on

the Shanghai Stock Exchange and Shenzhen Stock

Exchange

Fan and Palaniswami

(2001)

Financial data for stocks listed on the ASX from

1992 to 2000

 Ballings et al. (2015) Financial data for 5,767 listed European rms from

2009 to 2010

Note: e table presents a list of frequently cited studies that use one or several major AI techniques

(hybrid or ensemble approaches) for fundamental analysis.

Articial Intelligence in Asset Management

NLP approaches are capable of extracting economically meaningful informa-

tion from various sources of text, such as corporate annual reports (Azimi and

Agrawal 2019), news articles (Schumaker and Chen 2006; Ke, Kelly, and Xiu

2019), and Twitter posts (Sprenger, Sandner, Tumasjan, and Welpe 2014).

Unlike more traditional textual analysis techniques, such as dictionary-based

approaches that extract information only from individual words in the text,

AI approaches can also interpret context and sentence structure.

LASSO regression can automatically select the factors with the highest

explanatory power for future returns from a large set of return-predictive sig-

nals documented in the literature (Feng, Giglio, and Xiu 2017; Freyberger,

Neuhierl, and Weber 2018). e LASSO framework can also be used to nd

lead–lag relationships between asset groups or markets. For example, one can

investigate which domestic industry or market returns are the most signicant

predictors of returns among all other markets or industries (Rapach, Strauss,

and Zhou 2013; Rapach, Strauss, Tu, and Zhou 2019). More-generalized ver-

sions of LASSO regression, known as “elastic nets,” complement LASSO’s

variable selection feature by also ensuring that estimated coecients are not

disproportionately large (e.g., Gu et al. 2020). In addition, AI models can be

used to identify stocks expected to outperform or underperform, using a range

of economic or rm-level variables. e results of these analyses can then be

incorporated into the portfolio optimization process by allocating more (less)

weight to assets with high (low) alpha. Beyond using historical data, training

AI using actual experts’ stock buy or sell recommendations (Bew, Harvey,

Ledford, Radnor, and Sinclair 2019; Papaioannou and Giamouridis, forth-

coming) has also been successful.

Across AI techniques available for return prediction, ANNs have been

found to perform best compared with ordinary least squares regression, elas-

tic nets, LASSO regressions, random forests, and gradient boosted regression

trees (Gu et al. 2020). In fact, the out-of-sample predictions of an ANN with

three hidden layers were almost 30% more accurate than those generated by a

gradient boosted regression tree, which was the second best-performing tech-

nique among the six. Note that these results might be highly task- and data-

specic. Nevertheless, the success of neural networks in this case is largely

attributed to their ability to capture complex nonlinear relationships. In addi-

tion, these models stand apart because they are highly versatile and because

a large number of functional forms and structures are available that allow

neural networks to learn from data more eectively than other techniques.

Recent studies have also introduced methods of interpreting neural networks

statistically using condence intervals and by ranking the importance of input

variables and interaction eects (Dixon and Polson 2019).

3. Portfolio Management

Not surprisingly, neural networks are therefore one of the most popular

AI techniques for predicting stock returns (Vui, Sim, Soon, On, Alfred, and

Anthony 2013; Abe and Nakayama 2018), company fundamentals (Alberg

and Lipton 2017), and returns of other asset classes such as bonds (Bianchi,

Büchner, and Tamoni 2019). However, evidence is also available that indi-

cates vector machines can be better at predicting the rst two moments of

asset returns than ANNs can, provided they are tuned appropriately (Huang,

Nakamori, and Wang 2005; Chen, Shih, and Wu 2006; Arrieta-ibarra and

Lobato 2015). Consequently, a popular implementation consists of using the

average prediction across various AI techniques. is “ensemble” approach has

been shown to produce better predictions than any individual AI technique

(Rasekhschae, Christian, and Jones 2019; Borghi and De Rossi, forthcom-

ing). Recent ndings indicate that AI signals generate signicant prots in

both short and long positions (0.78% abnormal returns per month for a long-

only, value-weighted portfolio) and that these prots remain statistically and

economically signicant even in the post-2001 period, during which a global

decay is seen in abnormal returns (Avramov, Cheng, and Metzker 2019).

Modeling and predicting asset prices becomes a particularly challenging

exercise when derivatives are involved. As a result, constructing optimal port-

folios that include derivatives is dicult, because their prices and payos are

not well dened and are contingent on other assets. Most conventional deriv-

ative pricing approaches rely heavily on theoretical models, such as Black–

Scholes, that are based on somewhat restrictive assumptions. is is, again, a

realm where AI can play a role. For example, ANNs can be used for pricing

and hedging using nonparametric option pricing frameworks that perform

better than the Black–Scholes model in terms of delta hedging (Hutchinson,

Lo, and Poggio 1994) and forecasting future option prices (Yao, Li, and

Tan 2000). Recent studies also extend the deep learning framework to price

exotic (Becker, Cheridito, and Jentzen 2019a) and American-style (Becker,

Cheridito, and Jentzen 2019b) options.

Lastly, AI can be used for improving estimates of variance–covariance

matrices in the Markowitz framework. To illustrate, hierarchical cluster anal-

ysis can replace the covariance structure of asset returns with a tree structure

(de Prado 2016). is approach uses all the information contained in the covari-

ance matrix but requires fewer estimates and thus leads to more stable and robust

portfolio weights. Empirical evidence using simulated return observations sug-

gests that a minimum variance portfolio under this approach has a 31.3% higher

Sharpe ratio than that under the classical Markowitz (1952) framework.

Ultimately, the jury is still out as to whether AI implementations are gen-

erally superior to more traditional implementations in stock selection, factor

Articial Intelligence in Asset Management

investing, or asset allocation. More evidence would be desirable to conrm

that the benets of AI models, including their ability to capture nonlineari-

ties, outweigh the costs and potential data issues, such as collinear variables.

e additional evidence will only grow more important because many asset

managers have recently started using AI, potentially leading to the superior

performance of AI-based investment strategies being arbitraged away in the

near future. Moreover, other reasons to be cautious also exist. For example,

some research advocating the use of AI in portfolio management has exam-

ined only small samples of assets or emerging markets that lack liquidity

and eciency. Another challenging aspect of using AI is selecting relevant

variables from the raw data and transforming them into appropriate for-

mats for AI models to function properly, also known as “feature engineer-

ing.” is constitutes an essential and time-consuming part of alpha research

(Rasekhschae et al. 2019).

3.2. Portfolio Optimization

A portfolio manager’s decision entails allocating funds among a (large) set of

assets such that the target portfolio satises an objective (e.g., mimicking an

index, maximizing the Sharpe ratio), given certain constraints. e mean–

variance framework of Markowitz typically oers the theoretical founda-

tion. However, two main challenges arise in practice (Michaud and Michaud

2008; Kolm, Tütüncü, and Fabozzi 2014). First, the optimal asset weights

are highly sensitive to estimates of expected returns. Considering that esti-

mates of future expected returns are often uncertain, the optimization exer-

cise can yield unstable weights that perform poorly out of sample. In fact, the

noise in return estimates can erode any diversication benets. For example,

DeMiguel, Garlappi, and Uppal (2009) show that an equally weighted port-

folio has a higher out-of-sample Sharpe ratio than the optimal Markowitz

portfolio and a range of other optimal portfolios.

Second, estimating the variance–covariance matrix, which is at the heart

of Markowitz’s theory, requires a large time series of data and the assumption

of stable correlations between asset returns. Moreover, the matrix becomes

unstable when asset correlations increase, which happens at times when diver-

sication is most important and yet more dicult to achieve (de Prado 2016).

AI addresses these challenges in two ways. First, it can produce return

and risk estimates that are more accurate than those produced by other meth-

ods and that can be used within traditional portfolio construction frame-

works. Second, AI techniques can provide alternative portfolio construction

approaches to generate more accurate portfolio weights and produce opti-

mized portfolios with better out-of-sample performance than those generated

3. Portfolio Management

by traditional linear techniques. Although empirical evidence is limited,

interest seems to be growing among both academics and practitioners.

In particular, ANNs can be trained to make asset allocation decisions

subject to complex constraints that are often not straightforward to integrate

into the mean–variance framework. For example, a neural network can select

portfolios according to a learning criterion that maximizes returns subject

to value-at-risk constraints (Chapados and Bengio 2001). ANNs can also

solve complex multi-objective optimization problems. To illustrate, a neu-

ral network–based methodology can construct a mean–variance-skewness

optimal portfolio in a fast and ecient manner (Yu, Wang, and Lai 2008).

Furthermore, ANNs can incorporate views about the future asset perfor-

mance into the portfolio optimization using a Black and Litterman (1992)

framework, generating higher out-of-sample Sharpe ratios than the market

portfolio (Zimmermann, Neuneier, and Grothmann 2002).

Another popular AI technique in portfolio construction is evolution-

ary algorithms that have the exibility to accommodate more complex asset

allocation problems. For example, evolutionary algorithms solve optimiza-

tion problems under cardinality constraints (restricting the number of assets

in the portfolio) and maximum or minimum holding thresholds (Branke,

Scheckenbach, Stein, Deb, and Schmeck 2009). Evolutionary algorithms are

also able to incorporate additional objectives. For example, one can incor-

porate model risk (i.e., the risk of failing to produce accurate estimates of

asset returns and volatilities as a result of model mis-specications) into the

optimization problem to reduce forecasting error (Skolpadungket, Dahal, and

Harnpornchai 2016). Optimal portfolios using this approach have better-

realized Sharpe ratios by approximately 10% than those that consider only

return and volatility in their objective functions.

e ability of ANNs to capture nonlinear relationships between assets

without any prior knowledge about the underlying structure of the data can

be useful in synthetic replication—that is, replicating a benchmark portfolio

such as an index by holding a fraction of the constituents while minimizing

tracking error by matching some of the benchmark’s risk factors. For exam-

ple, ANNs can approximate the Financial Times Stock Exchange (FTSE)

100 Index with only seven stocks (Lowe 1994), resulting in lower transaction

costs from portfolio rebalancing as well as reduced portfolio management and

monitoring costs. is framework has promising out-of-sample performance

and is exible enough to generate target portfolios with other specied char-

acteristics. For example, one can nd the best strategy (i.e., the strategy with

the lowest risk or cost) to construct a portfolio that outperforms a specic

index by 1% on an annual basis (Heaton, Polson, and Witte 2017).

4. Trading

Algorithms can play a role in all stages of the trading process (Nuti,

Mirghaemi, Treleaven, and Yingsaeree 2011). e trading process can be

broken down into pre-trade analysis, trade execution, and post-trade analy-

sis (Figure 5). Pre-trade analysis entails using data to analyze properties of

nancial assets with the objective of forecasting not only their future perfor-

mance but also the risks and costs involved in trading them. Insights from

this analysis ultimately lead to the execution of trades. Pre-trade analysis

Figure 5. Algorithmic Trading with AI

Pre-trade

- AI uses data to generate a

provisional trading list

- Risks and costs involved

in trading are estimated to

select feasible trades

Execution

- Strategies generated in the

previous stage are executed

- AI uses data to determine

optimal execution strategies,

minimizing transaction costs

Post-trade

- Realized trade and market

outcome data are analyzed

- Risk in trading positions is

monitored continuously

Note: e gure presents the three stages of algorithmic trading and summarizes the applications

of AI in each stage.

4. Trading

can be a manual stage, meaning it involves some form of human supervision,

given that asset managers might want to consider results from pre-trade anal-

yses together with risk assessments and client preferences. In high-frequency

or fully automated systems, however, pre-trade analysis does not involve any

human intervention. Trade execution implements trades while ensuring low

transaction costs. Actual trading outcomes are evaluated during post-trade

analysis to monitor performance and improve the trading system. Post-trade

analysis often involves some form of human supervision or overlay. In con-

trast, pre-trade analysis and trade execution are handled mostly by algorithms

because they require timely and complex analyses.

4.1. Algorithmic Trading

AI plays a role in trading by facilitating algorithmic trading—dened as algo-

rithms that automate one or more stages of the trading process. Algorithmic

trading has experienced a growing presence in asset management thanks to

three recent phenomena (Kirilenko and Lo 2013). First, developments in

computing power, data science, and telecommunication have led to structural

changes in the way nancial markets operate. Computers are now capable

of collecting and analyzing large amounts of data and of executing trades

in milliseconds without any human intervention. Second, breakthroughs in

quantitative nance and ML have provided the necessary tools for comput-

ers to conduct insightful nancial analysis faster and more eciently than

human beings. ird, the increasing speed, complexity, and scale of nancial

markets, together with the breadth of new structural products, have made

keeping track of markets and making real-time trading decisions dicult if

not impossible for humans, whereas complex AI techniques such as ANNs

can now be implemented in close to real time (Leshik and Cralle 2011).

Strategies used in algorithmic trading are often based on technical analy-

sis, which uses past stock and market data to predict future asset returns.

Although performing fundamental analysis is also possible, algorithmic

trades are often of a high frequency, so that analyzing lower-frequency data

such as rm fundamentals is typically less eective. In addition, evidence

is available that indicates that technical indicators dominate fundamental

ones in generating protable trading signals using AI (Borghi and De Rossi,

forthcoming). erefore, AI-based approaches have established a more active

presence in technical analysis (Table 2).

e main inputs to traditional forms of technical analysis are past price

and trading volume data. Strategies based on prices often model trends, such

as momentum or reversal, and cycles using historical data to forecast future

returns. On the other hand, volume-based strategies predict future returns

Articial Intelligence in Asset Management

Table 2. AI and Technical Trading Rules

Techn ique Study Sample/Data

ANNs Dixon, Klabjan, and Bang

(2017)

Five-minute mid-prices for 43 CME-listed

commodity and foreign exchange futures from

1991 to 2014

Choudhry, McGroarty,

Peng, and Wang (2012)

JPY/USD, DM/USD, and USD/EUR exchange

rates from 1998 and 1999

Gradojevic and Yang (2006) CAD/USD exchange rates from 1990 to 2000

Atsalakis and Valavanis

(2009)

No empirical work; surveys other studies

Dunis, Laws, and

Sermpinis (2010)

Daily EUR/USD exchange rates from 1999 to 2007

Fischer and Krauss (2018) Daily stock data for the constituents of the S&P 500

from 1992 until 2015

Cluster

Analysis

Liao and Chou (2013) 30 industrial indices from TAIEX, Shanghai Stock

Exchange/Shenzhen Stock Exchange, and the Hang

Seng Index from 2008 to 2011

Decision

Trees

Booth, Gerding, and

McGroarty (2014)

Stock data for 30 rms from the DAX stock index

from 2000 to 2013

Coqueret and Guida (2018) Financial data for a sample of between 305 and 599

large US rms from 2002 to 2016

Genetic

Algorithms

Hu et al. (2015) No empirical work; surveys other studies

Allen and Karjalainen

(1999)

S&P 500 Index daily prices from 1928 to 1995

Manahov, Hudson, and

Gebka (2014)

One-minute quote data for six major currency pairs

from 2012 to 2013

Berutich, López, Luna,

and Quintana (2016)

Stock data for 21 rms listed on the Spanish market

from 2000 to 2013

Hybrid/

Ensemble

Cheng, Chen, and Wei

(2010)

TAIEX stock index data from 2000 to 2005

Tan, Quek, and Cheng

(2011)

Stock data for more than 20 rms from the US

market from 1994 to 2006

Tsai, Lin, Yen, and Chen

(2011)

Stock data for a subset of the Taiwan stock market

companies from 2002 to 2006

Nuij, Milea, Hogenboom,

Frasincar, and Kaymak

(2014)

Stock data for all FTSE 350 stocks from January to

April 2007

Geva and Zahavi (2014) Stock data for 72 S&P 500 rms from 2006 to 2007

(continued)

4. Trading

based on recent investor trading activity. Modern technical analysis also

incorporates information from other sources, including fund ows, investor

trades, and textual data from news articles or online sources. AI techniques

using NLP can be particularly useful with respect to these new, unstructured

sources of data.

4.2. Transaction Cost Analysis

Analyzing transaction costs is an essential part of pre-trade analysis that

indicates whether the costs of trading are small enough for a trading signal

to generate prots net of implementation costs. Transaction costs have three

main components: bid–ask spreads, market impact costs, and trading com-

missions. Among these three, market impact costs—dened as the adverse

eect of a trade on market prices—are the only costs that are not observ-

able before the trade is initiated. Nevertheless, having an estimate of market

impact costs is crucial because they represent a signicant portion of trans-

action costs: Market impact absorbs as much as two-thirds of trading gains

made by systematic funds (Financial Stability Board 2017).

AI approaches complement traditional market impact models by pro-

viding additional insights. e nonparametric structure of AI techniques,

together with their ability to capture nonlinear dynamics, are particularly use-

ful for predicting market impact, and various AI techniques have been tested

for this purpose. Performance-weighted random forests are found to outper-

form linear regression, ANNs, and SVMs in predicting the market impact

of a market order by 20% out of sample (Booth, Gerding, and McGroarty

2015). On the other hand, SVMs do not seem to perform particularly well

when forecasting market impact, whereas ANNs do well if they are properly

dened and estimated (Park, Lee, and Son 2016).

Techn ique Study Sample/Data

LASSO Chinco, Clark-Joseph,

and Ye (2019)

One-minute returns of NYSE stocks from 2005

to 2012

NLP Renault (2017) Dataset of stocks with messages published on

StockTwits from 2012 to 2016

 Hagenau, Liebmann, and

Neumann (2013)

Stock data for a subset of German and British rms

from 1997 to 2011

Note: e table presents a list of frequently cited studies that use one or several major AI techniques

(hybrid or ensemble approaches) to devise technical trading rules used in algorithmic trading.

Table 2. AI and Technical Trading Rules (continued)

Articial Intelligence in Asset Management

Although these nonparametric techniques perform well in estimating

market impact, they have two major shortcomings. First, the majority of

approaches have no economic intuition for the drivers of price impact. As

a result, they are prone to capturing noise rather than relevant information.

Second, these techniques cannot distinguish between permanent and tempo-

rary market impact, which would require additional variables, including trade

direction and liquidity (Farmer, Gerig, Lillo, and Mike 2006). To address

these two issues, a parametric approach such as LASSO regression can be

used alongside nonparametric techniques. With LASSO regression, the most

informative variables capturing information related to the order book and

other sources are selected to predict price impact. Empirical evidence indi-

cates that trade sign, market order size, and liquidity based on best limit order

prices are the most important variables for forecasting market impact (Zheng,

Moulines, and Abergel 2013). A Bayesian network model is another approach

for estimating market impact while providing intuition on the main drivers.

Unlike most other ML techniques, this approach can also account for vari-

ables with data availability issues and model them as latent variables using

Bayesian inference. anks to this feature, other important variables can be

identied (e.g., net order ow imbalance) and added to the model to improve

the forecast (Briere, Lehalle, Nefedova, and Raboun 2019).

Another useful application of AI consists of estimating the market impact

of trades in assets that lack sucient (or any) historical trading data, given that

using traditional approaches to estimate the market impact costs is almost

impossible in this case. A cluster analysis approach can tackle this problem by

identifying comparable assets with similar behavior and using their histori-

cal data instead. For example, cluster analysis can allocate bonds into clusters

based on their duration, maturity, or value outstanding and measure their

similarity according to these variables. Within each cluster, the information

of other bonds is used for bonds without sucient data. Bloomberg’s liquidity

assessment tool notably uses this technique to provide liquidity information

for various assets.

4.3. Trade Execution

Executing large trades often involves signicant market impact costs.

erefore, such trades are typically broken up into a sequence of smaller

orders, which are easier and cheaper to execute. is approach is known as the

execution strategy that requires determining the timing and size of smaller

orders using some form of execution model. e objective of such models is to

minimize transaction costs while completing the transaction within a speci-

ed period. Classical modeling approaches for this problem use stochastic

4. Trading

control techniques to determine optimal execution strategies (a methodology

that goes back to Bertsimas and Lo 1998). Classical models, however, often

rely on restrictive assumptions regarding asset price dynamics and the func-

tional form of market impact (Kearns and Nevmyvaka 2013).

In contrast, AI approaches facilitate trade execution modeling by actively

learning from real market microstructure data when determining optimal

execution strategies. Recent studies advocate reinforcement learning tech-

niques (i.e., algorithms that receive vectors of microstructure and order book

variables, such as bid–ask spread, volume imbalances between the buy and

sell sides of limit order book, and signed transaction volume) as input and

return optimal execution strategies as output (e.g., Nevmyvaka, Feng, and

Kearns 2006; Kearns and Nevmyvaka 2013; Hendricks and Wilcox 2014;

Kolm and Ritter, forthcoming). e algorithms essentially learn to map each

combination of input variables, known as a “state,” to trading actions such

that transaction costs are minimized (Kearns and Nevmyvaka 2013).

e advantage of AI-based approaches is that they rely on data rather

than normative assumptions to determine market impact costs, price move-

ments, and liquidity. ey therefore have the exibility to adapt as market

conditions change and new data become available. ese models are often

dicult to train and understand, however, especially for large portfolios that

benet the most from a reduction in transaction costs. In addition, systematic

execution strategies run the risk of cascading into a systemic event aecting

the whole market. A famous precedent for this phenomenon is the so-called

ash crash of 2010 (Kirilenko, Kyle, Samadi, and Tuzun 2017).

5. Portfolio Risk Management

AI also has applications in risk management, with regard to both market risk

and credit risk (Financial Stability Board 2017; Aziz and Dowling 2019).

Market risk refers to the likelihood of loss resulting from aggregate market

uctuation, and credit (or counterparty) risk is the risk of a counterparty not

fullling its contractual obligations, which results in a loss in value (Figure 6).

Although AI has broader uses in risk management, these two categories are

the most important in asset management.

5.1. Market Risk

Market risk analysis involves modeling, assessing, and forecasting risk factors

that aect the investment portfolio. AI can play a role in this area in three

ways: (1) making use of qualitative data for risk modeling, (2) validating and

Figure 6. AI Applications in Risk Management

Artificial Intelligence in

Risk Management

Incorporating qualitative data

in risk modeling (e.g., news

articles, annual reports,

social media)

Validating and backtesting

risk models

Producing forecasts of financial

or economic variables used in

risk management (e.g.,

bankruptcy probability, value at

risk, interest rates,

exchange rates)

Note: e gure presents a summary of three areas in which AI can play a role in risk management.

5. Portfolio Risk Management

backtesting risk models, and (3) producing more accurate forecasts of aggre-

gate nancial or economic variables (Figure 6).

One area of application for AI in market risk management relates to

extracting information from textual or image data sources. Textual data

sources, including news articles, online posts, nancial contracts, central

bank minutes and statements, and social media, can contain valuable infor-

mation for managing market risk (Groth and Muntermann 2011). Satellite

images are analyzed to predict sales at supermarkets or future crop harvests

(Katona, Painter, Patatoukas, and Zeng 2018). e information provided by

these sources is, in many cases, not captured by other quantitative variables.

For example, AI approaches that use textual information have been shown

to generate better predictions of market crashes (Manela and Moreira 2017),

interest rates (Hong and Han 2002), and other major macroeconomic out-

comes (Cong, Liang, and Zhang 2019) than those using information cap-

tured by other data sources. ese approaches can also extract information

from corporate disclosures with the aim of determining rms’ systematic risk

proles (e.g., Groth and Muntermann 2011; Bao and Datta 2014; Cong et al.

2019). All these applications have triggered an interest among central banks

in incorporating methods of AI-based text mining in macroprudential analy-

ses (Bholat, Hansen, Santos, and Schonhardt-Bailey 2015). To date, empiri-

cal implementations and evidence in this area are scarce.

AI can also help risk managers validate and backtest risk models

(Financial Stability Board 2017). Regulators and nancial supervisory insti-

tutions emphasize this important part of model risk management (Board of

Governors of the Federal Reserve System 2011). Unsupervised AI approaches

can detect anomalies in risk model output by evaluating all projections gener-

ated by the model and automatically identifying any irregularities. Risk man-

agers can also use supervised AI techniques to generate benchmark forecasts

as part of model validation practice. Comparing model results and bench-

mark forecasts will indicate whether the risk model is producing predictions

that dier signicantly from those generated by AI. A signicant disagree-

ment between AI forecasts and standard risk model outputs can highlight

potential problems and trigger a more thorough investigation.

Depending on the exposure of the assets in a portfolio to the underlying

risk factors, various nancial or economic variables can aect its performance.

erefore, modeling future trends in these factors, especially macroeconomic

variables, is important (Elliott and Timmermann 2008; Ahmed, Atiya, El

Gayar, and El-Shishiny 2010), and ANNs are particularly popular in this

context. For example, empirical evidence suggests that variants of ANNs per-

form signicantly better than linear autoregressive approaches in forecasting

Articial Intelligence in Asset Management

47 monthly macroeconomic variables of the G7 economies (Teräsvirta, van

Dijk, and Medeiros 2005). Using ANNs entails the risk of producing implau-

sible forecasts at long horizons, however. Nonetheless, ANNs have been par-

ticularly successful in forecasting interest rates (e.g., Kim and Noh 1997; Oh

and Han 2000) and exchange rates (e.g., Kaashoek and van Dijk 2002; Majhi,

Panda, and Sahoo 2009).

ANNs can also be used to devise systematic risk factors. ese models

can capture nonlinearities and interactions of covariates, including rm char-

acteristics and macroeconomic variables (e.g., Bryzgalova et al. 2019, Chen,

Pelger, and Zhu 2020; Gu, Kelly, and Xiu 2019; Feng, Polson, and Xu 2020).

Such factors can better account for risk premia and distinguish between non-

diversiable and diversiable (idiosyncratic) risk than conventional linear fac-

tors can. LASSO regressions can also be useful in determining systematic

factor structures. ese models are able to select the most relevant system-

atic risk factors from a subset of factors or market indices (Giamouridis and

Paterlini 2010).

AI techniques can also predict market volatility and nancial crises, espe-

cially ANNs and SVMs, whose ability to capture nonlinear dynamics gives

them an advantage over traditional generalized autoregressive conditional

heteroskedasticity (GARCH) models. ANNs can predict market volatility

either directly (Hamid and Iqbal 2004) or in combination with a variant of

GARCH (Donaldson and Kamstra 1997; Fernandes, Medeiros, and Scharth

2014). Some researchers, however, found SVMs to be superior to ANNs in

this context (Chen, Hardle, and Jeong 2009). In addition to volatility model-

ing, ANNs and SVMs are used to predict nancial crises. Models performing

this forecasting task are often referred to as early warning systems. Almost all

major nancial institutions use a form of early warning system to monitor

systemic risk. ANNs and SVMs have been shown to predict currency crises

(e.g., Lin, Khan, Chang, and Wang 2008; Sevim, Oztekin, Bali, Gumus, and

Guresen 2014), banking crises (e.g., Celik and Karatepe 2007; Ristolainen

2018), and recessions generally (e.g., Yu, Wang, Lai, and Wen 2010; Ahn,

Oh, T.Y. Kim, and D.H. Kim 2011; Gogas, Papadimitriou, Matthaiou, and

Chrysanthidou 2015) with reasonable accuracy. Nevertheless, crises are rare

nancial events, so in the absence of a sucient number of such events in

the sample, one could question the ability of AI models to accurately predict

future crises.

5.2. Credit Risk

e objective of credit risk management is to ensure that the failure of any

counterparty to meet its obligations does not have a negative eect on the

5. Portfolio Risk Management

portfolio beyond specic limits. Asset managers need to monitor the credit

risk of the entire portfolio as well as of individual positions and transactions.

is practice involves modeling the solvency risk associated with institutions

issuing nancial products, including equities, bonds, swaps, and options.

An extensive range of approaches exists for modeling solvency or bankruptcy

risk. Multivariate discriminant analysis, logit, and probit models are among

the most common traditional methods used (Bellovary, Giacomino, and

Akers 2007).

Credit risk modeling is one of the rst areas of nance to consider the

application of AI techniques. e two most widely used techniques are ANNs

and SVMs. In fact, ANNs have become mainstream bankruptcy modeling

techniques since the early 1990s (Tam 1991). e popularity of ANNs stems

largely from their higher success in forecasting bankruptcy and determin-

ing credit ratings compared with traditional techniques (e.g., Zhang, Hu,

Patuwo, and Indro 1999; Tsai and Wu 2008). More-recent studies, however,

advocate the use of SVMs (e.g., Auria and Moro 2008; Ribeiro, Silva, Chen,

Vieira, and das Neves 2012) because they yield slightly more accurate bank-

ruptcy forecasts than ANNs do (Huang, H. Chen, Hsu, W.-H. Chen, and

Wu 2004). Moreover, SVMs are less likely to face some of the issues common

with ANNs, such as overtting. ANNs and SVMs also perform particularly

well when estimating loss given default (dened as the economic loss when

default occurs), which the Basel II Accord requires nancial institutions to

model in addition to the default probability for regulatory capital monitoring

purposes (Loterman, Brown, Martens, Mues, and Baesens 2012).

Beyond SVMs and ANNs, a wide range of other AI approaches—

including genetic algorithms (Varetto 1998)—can be used for credit risk

modeling (Kumar and Ravi 2007; Peña, Martinez, and Abudu 2011). Because

each of the modeling techniques has its own specic advantages and disad-

vantages, an ensemble technique that uses various approaches separately and

then combines the resulting predictions should be considered for achieving

the best performance (Verikas, Kalsyte, Bacauskiene, and Gelzinis 2010).

6. Robo-Advisors

Robo-advisors are computer programs that provide customized advice to

assist individual investors in investment activities. ese programs have

gained signicant attention recently because of their success in reducing bar-

riers to entry for retail investors. Academic interest in researching how to

enhance robo-advisors using AI is growing (Figure 7). e primary focus is

on devising algorithms known as recommender systems that produce optimal

portfolios catered to investors’ risk appetites (e.g., Xue, Q. Liu, Li, X. Liu, Ye,

Wang, and Yin 2018). However, robo-advising can integrate all types of AI

Figure 7. Robo-Advising with AI

Robo-Advisor

Investor Information

- Attitude toward risk

- Behavioral questionnaire

- Financial goals

Artificial Intelligence

- Wide range of ML

techniques to analyze

market data

- NLP to incorporate textual

data and provide chatbots

Big Data

- Access to large volumes of

financial and nonfinancial

data sources

Financial and

Investment Advice

- Financial advice (e.g., banking

products, insurance policies)

- Investment advice (e.g., portfolio

of assets calibrated to investor

goals and risk tolerance)

Advantages

- Efficient delivery of financial advice and

investment recommendations

- Supports less educated or affluent investors

- Ability to perform complex analyses on

large datasets

- Not prone to human biases and mistakes

Disadvantages

- Limited view of risk tolerance

- Ignores taxes and inflation

- Ineffective advice during crises

- Shifting responsibility from institutions

to retail investors

Note: e gure illustrates the structure of robo-advisor systems that incorporate AI and summa-

rizes the advantages and disadvantages of these systems.

6. Robo-Advisors

applications into portfolio management, trading, and portfolio risk manage-

ment. By building on the success of AI in these elds, robo-advisors can not

only produce portfolios with better out-of-sample performance for investors

but also rebalance portfolios, automatically managing the portfolio’s risks and

minimizing transaction costs. Because robo-advising is less expensive than

working with a human advisor and can be performed through a simplied

interface, investing via a robo-advisor is ultimately both more benecial and

more accessible for retail investors.

Robo-advisors are also less prone to behavioral biases, mistakes, and ille-

gal practices. In fact, robo-advising has been shown to appeal most to inves-

tors who fear being victims of investment fraud (Brenner and Meyll 2019).

More sophisticated institutional investors can benet from robo-advisors’

ability to eciently process a wide range of nancial data. Although reducing

behavioral biases when making investment decisions is benecial to all types

of investors (D’Acunto, Prabhala, and Rossi 2017), less sophisticated inves-

tors particularly benet from robo-advice in terms of enhancing portfolio

performance, increasing diversication, and reducing volatility. At the same

time, because robo-advisors have trade execution services integrated into

them, they often encourage investors to trade more. is increased trading

can be both a benet, in terms of encouraging investors to rebalance positions

more often, and a pitfall, because it can lead to excessive trading that benets

robo-advising systems through commissions at the expense of investors. To

be able to use robo-advisors and benet from their advantages, an investor

needs a certain minimum level of technological understanding and nancial

sophistication.

Not all robo-advisors necessarily use new, sophisticated methods. An

analysis of 219 international robo-advisors shows that Markowitz’s portfolio

theory is the most prevalent approach, although some systems do not disclose

their techniques (Beketov, Lehmann, and Wittke 2018). More-sophisticated

robo-advisors rely on proprietary algorithms and do not divulge the details

of their approach to analyzing portfolios and making recommendations.

Nevertheless, an examination of the industry indicates that the most success-

ful robo-advisors rely heavily on AI to conduct investment and trading analy-

ses (Sabharwal 2018). After all, robo-advising and ntech in general derive

most of their success from collecting and analyzing data, and AI is an integral

part of this process (Dhar and Stein 2017).

7. Articial Intelligence Risks and

Challenges: What Can Go Wrong?

Although many studies of AI in nance highlight the technology’s advan-

tages and benets for various applications, AI users should also be aware

of some of its actual or perceived risks and downsides with respect to asset

management. ese potential negative issues are often related to complexity,

opacity, and dependence on data integrity (Figure 8).

Understanding and explaining the inferences made by most AI models

is dicult, if not impossible. As the complexity of the task or the algorithm

grows, opacity can render human supervision ineective, thereby becom-

ing an even more signicant problem. is issue might have repercussions

for asset managers in three ways. First, the diculty in predicting how AI

models will respond to major surprises or “black swan” events could lead to

systematic crashes. Even in the absence of major events, AI algorithms may

make the same errors at the same time, introducing the risk of cascading mar-

ket crashes. Indeed, the considerable cost of producing AI algorithms has led

to most asset management companies using the same tools and algorithms.

As a result, AI-driven crashes could be much more likely than other cascad-

ing algorithmic crashes we have experienced. Cascading algorithmic crashes

are not specic to AI systems and may arise from even simple widespread

quantitative approaches, such as value investing. What makes AI dierent,

however, is that its opacity may prevent such risks from being properly mod-

eled and monitored.

Second, AI can make wrong decisions based on incorrect inferences

that have captured spurious or irrelevant patterns in the data. For example,

ANNs that are trained to pick stocks with high expected returns might select

illiquid, distressed stocks (Avramov et al. 2019). ird, attributing invest-

ment performance can become more challenging when using AI models. For

example, the widely used Barra Risk Factor Analysis, based on linear factor

models, might not suit AI-based strategies that capture nonlinear relation-

ships between characteristics and returns. Consequently, in cases of poor

fund performance, explaining to investors how and why the investment

strategy failed can be dicult, which could undermine investors’ trust in the

fund or even in the industry. To better understand the behavior of AI models,

some people approximate an AI model’s prediction behavior by construct-

ing an additional, simpler, and interpretable “surrogate model.” Shapley val-

ues from game theory can be used to understand how much dierent feature

7. Articial Intelligence Risks and Challenges

values contribute to a prediction. A good overview on these and many other

approaches to explaining AI models can be found in Molnar (2020).

Moreover, the black box character of many AI systems raises the issue of

responsibility and makes regulation challenging (Zetzsche, Arner, Buckley,

and Tang 2020).

Figure 8. AI Areas of Concern

Opacity & Complexity

- Systematic crashes

- Incorrect inference

- Performance attribution

difficulty

Data Integrity & Sufficiency

- Heavy reliance on data quality

- Requirement for large

amounts of data

- Past data not fully

representing the future

Note: e gure summarizes major potential sources of risk introduced by adopting AI in asset

management.

Articial Intelligence in Asset Management

Data quality and suciency can be other major sources of concern. Like

other empirical models, AI models rely on the integrity and availability

of data. Poor data quality can easily trigger what is famously described as

“garbage in, garbage out.” Data quality and suciency become particularly

important because AI outputs are often taken at face value. erefore, iden-

tifying data-related issues by evaluating the model outcomes might not be

a straightforward exercise. Furthermore, AI models require large amounts

of data during the learning phase, often more than are available. is lack

of data might lead to improper calibration caused by the input data’s poor

signal-to-noise ratio, especially in the case of low-frequency nancial data

with numerous missing observations. Imputation, a preprocessing step in

which statistical values are used as substitutes for missing observations (e.g.,

Kofman and Sharpe 2003), may help, but obviously only to a certain extent.

Some argue that past data in general might not fully represent the future.

is shortcoming can become particularly prominent when the short time

series of available nancial data misses certain important extreme events in

the past, increasing the likelihood that AI models will fail during a crash

or crisis (Patel and Lincoln 2019). As a side eect, AI’s growing presence

in the investment industry and asset managers’ reliance on it for day-to-day

tasks might further increase the asset managers’ cybersecurity risk (Board of

Governors of the Federal Reserve System 2011).

Overall, whether the benets of AI outweigh the considerable costs of

investing in the required software, hardware, human resources, and data sys-

tems is not yet clear. After all, limited resources are available for asset man-

agers to develop and test new strategies, so that investment in AI must be

considered alongside mutually exclusive, competing research projects. Once

the current AI hype has dissipated, investors may become less keen to invest

in AI-driven funds, which would make breaking even on investments in AI

infrastructure even harder. us, asset managers will need to carefully con-

sider both the benets and costs of AI (Patel and Lincoln 2019; Buchanan

2019), if only not to get cold feet when the next AI winter comes.

8. Conclusion

e use of AI in asset management is an emerging eld of interest among

both academics and practitioners. AI has vast applications for portfolio man-

agement, trading, and portfolio risk management that enable the industry to

be more ecient and compliant. It also serves at the heart of new practices

and activities, such as algorithmic trading and robo-advising. Nevertheless,

AI is still far from replacing humans completely. Indeed, most of its opera-

tions within asset management are conned and controlled by some form of

human supervision. Consequently, a better way to describe AI is as a collec-

tion of techniques that automate or facilitate (often small) parts of the practice

of asset management, from the capacity to solve portfolio optimization prob-

lems with specic conditions to fully automated algorithmic trading systems.

e success of AI in asset management is linked to its three key, inher-

ent capabilities. First, AI models are objective, highly ecient in conducting

repetitive tasks, and able to identify patterns in high dimensional data that

may not be perceptible by humans. AI can also analyze data with minimal

knowledge of the data’s structure or the relation between input and output,

including nonlinear relations. is feature is especially useful for forecast-

ing, yielding more accurate estimates because AI does not rely on restrictive

assumptions inherent in more traditional methods. Second, AI can extract

information from unstructured data sources, such as news articles, online

posts, reports, and images. As a result, a tremendous amount of informa-

tion can be incorporated into nancial analysis without manual processing

and intervention. ird, AI algorithms, unlike other statistical techniques,

are often designed to improve themselves by readjusting in accordance with

the data. is ability means that the manual reconguration or parameter

re-estimation that is essential for traditional models is unnecessary with AI.

Finally, AI’s greatest strength—its ability to process data with minimal

theoretical knowledge or supervision—can also be its greatest weakness.

Indeed, a popular saying asserts that AI will always generate a result, even

when one should not exist. is tendency causes problems when data quality

is poor, when the task being performed is too complex for humans to monitor

or understand, and when cascading systemic failures could occur as a result of

several AI algorithms reacting to each other. Asset managers must bear such

issues in mind as the role of AI becomes more pervasive and signicant.

Appendix A. Basic Articial Intelligence

Concepts and Techniques

A.1. Articial Intelligence and Machine Learning

A.1.1. Origin and Denition. AI is widely believed to have started at the

Dartmouth Summer Research Project on Articial Intelligence, a workshop

organized by John McCarthy in the summer of 1956 at Dartmouth College.

Many prominent mathematicians and scientists, including Marvin Minsky

and Claude Shannon, attended this six-week brainstorming workshop. e

workshop proposal introduced the term “articial intelligence” and stated the

following objectives:

e study is to proceed on the basis of the conjecture that every aspect of

learning or any other feature of intelligence can in principle be so precisely

described that a machine can be made to simulate it. An attempt will be

made to nd how to make machines use language, form abstractions and

concepts, solve kinds of problems now reserved for humans, and improve

themselves. (McCarthy, Minsky, Rochester, and Shannon 2006, p. 12)

In recent years, the original denition of AI and what it should encom-

pass has evolved. Russell and Norvig (2010) distinguish the following four

dierent dimensions, or schools of thought, that determine the objective

of AI.

1. Acting Humanly: From this point of view, AI refers to the challenge of

creating computers capable of performing tasks in ways that are similar

to how humans perform them. An example of this is the Turing Test,

proposed by Alan Turing. e test poses a challenge in which a human

interrogator presents questions and receives responses from either another

human or a machine. e machine passes the test if the interrogator is

unable to distinguish the human’s answers from the machine’s.

2. Acting Rationally: is dimension aims to build agents that act ratio-

nally, (i.e., that aim to achieve the best outcome or, when uncertainty

exists, the best expected outcome).

3. inking Humanly: is perspective refers to the replication of human

thinking processes. e eld of cognitive science is a major manifesta-

tion of this approach to AI. It uses computer programs and insights from

experimental psychology to emulate the human mind.

Appendix A. Basic Articial Intelligence Concepts and Techniques

4. inking Rationally: inking rationally refers to using rules for reach-

ing logical conclusions based on premises assumed to be true.

Until a few decades ago, most research in AI fell into the category of

thinking rationally, represented by expert systems. Such systems have large

knowledge bases and an inference mechanism that allows the deduction of

new knowledge by logically deriving it through rules. For example, knowing

that all men are mortal and that Socrates was a man, an expert system could

infer that Socrates was mortal. Expert systems were highly popular in the

1970s and 1980s, but the need for building large, complex knowledge bases

and the systems’ deterministic nature have led to them falling out of favor.

e idea to have the machine learn through observations (i.e., ML) eventually

turned out to be more applicable in practice, and it is the predominant AI

technique behind most modern applications.

Problems studied under ML are of three main types—supervised learn-

ing, unsupervised learning, and reinforcement learning (Alpaydin 2010;

Murphy 2012)—each of which have common applications. Although many

of these techniques have existed for decades, the sudden surge in their popu-

larity and application has resulted from performance improvements thanks to

technological progress that has enabled computers to train ML models on a

scale that was not possible even a few years ago.

A.1.2. Supervised Learning. Consider the problem of determining the

sales price of a house based on a set of attributes, such as interior square footage,

geographic location, and number of oors. In supervised learning, an algorithm

establishes a mathematical relation between the feature data (square footage,

location, and number of oors) and the response data (sales price). Rather than

explicitly programming the model, a supervised learning algorithm is given a

set of training data. It then adjusts its model so as to minimize the prediction

error on the training data. Once the model has been established, it can be used

to infer a response from features that have not been observed before. Often the

training is iterative (i.e., a relation is rst guessed randomly) and subsequently

adjusted based on how erroneous the guess was. Over time, increasingly accu-

rate relations are produced until, ideally, a “best” relation is found. Eectively,

a machine learns how to relate the feature data to the response data. Because

a training set of correctly classied response data is used to guide the learning

process, the learning is deemed supervised. Supervised learning is currently

the most common learning approach in practice.

Supervised learning has two main applications: classication and regres-

sion. Predicting the sales prices of houses is an example of regression, because

the response data are quantitative and continuous. In classication, one

Articial Intelligence in Asset Management

is interested in determining a response that falls into one of a few catego-

ries, such as whether or not a credit card transaction is fraudulent, based on

observed features such as the distance of the transaction from the cardholder’s

residence, the amount of the transaction, and the object purchased.

A.1.3. Unsupervised Learning. Unsupervised learning is used to iden-

tify structures in data without access to labels. e most popular example is

clustering (i.e., the categorization of data into dierent groups wherein the

elements of each group have similar characteristics). is approach is useful,

for example, in marketing, where customers can be separated into dierent

groups, and dierent marketing strategies can be developed for each group.

Other applications are the detection of regularities (e.g., people who buy X

also tend to buy Y) and the compression of data.

A.1.4. Reinforcement Learning. e premise of reinforcement learn-

ing is that an agent (e.g., a program, a robot, a control system) learns how

to act appropriately in an environment based on reward signals it receives in

response to its actions. In each iteration, the agent observes the state of the

environment, decides how to act, and then receives a reward and information

about the next state of the environment. Reinforcement learning is the core

technology behind Google’s AlphaZero, an algorithm that learned to beat

the best human players in the board game Go simply by playing against itself

many times. ese algorithms can also be used in nance to solve dynamic

optimization problems, including portfolio optimization and trading in the

presence of transaction costs (Kolm and Ritter, forthcoming).

A.2. Overview of Common Articial Intelligence Techniques

Several AI techniques are widely used in asset management. ese include

ANNs, cluster analysis, decision trees, evolutionary (genetic) algorithms,

LASSO regression, SVMs, and NLP. is section briey characterizes these

techniques, discusses their strengths and weaknesses, and notes their areas of

application (Table A.1).

A.2.1. Least Absolute Shrinkage and Selection Operator

Regression. Linear regression is a common and relatively simple way to t

a model to data to make predictions or to estimate missing values. It seeks

to nd the coecients of explanatory (or predictor) variables that contrib-

ute to the value of the dependent (or predicted) variable. To nd the best

model, the most common approach is to minimize the sum of squared errors,

which are the dierence between observed values and the values predicted

by the model. As model complexity increases with the number of regressors,

Appendix A. Basic Articial Intelligence Concepts and Techniques

Table A.1. Summary of AI Techniques

Techn ique Strengths Weaknesses Areas of Application

ANNs — Complex and nonlin-

ear relationships

— Data and computa-

tionally intensive

— Image processing

and recognition

— Incremental and

transfer learning

— Predictions not

explainable

— Speech recognition

and synthesis

— Can generalize well — Possible overtting — Forecasting

Cluster

Analysis

— Labels unnecessary — Clusters may be

intertwined

— Data analysis

— Helps to understand

data

— May require cluster

count

— Anomaly detection

— Choosing attributes

can be dicult

— Recommendations

Decision

Trees

— Classications are

explainable

— Possible overtting — Decision making

— Complex and nonlin-

ear relationships

— Complex trees possible — Classication

— Poor at predicting

continuous variables

Evolutionary

(Genetic)

Algorithms

— Ability to handle high-

dimensional spaces

— Varies on initial

conditions

— Parameter

optimization

— Finds novel solutions — Computationally

intensive

— Portfolio

optimization

LASSO

Regressions

— Identify most relevant

features

— Model can be unstable

and hard to interpret

— Forecasting and

robust regression

analysis

— Flexible and fairly

simple

— Perform poorly when

independent variables

are correlated

— Sparse solutions

NLP — Analyzes and gener-

ates text and speech

— Currently primitive

and unable to fully

understand text

— Search engines and

news ltering

— Finds information in

large textual datasets

— Text classication

and summarization

SVMs — Structure of data can

be unknown

— Dicult to interpret — Classication

 — Can generalize with

less overtting risk

— Kernel dicult to

choose for nonlinear

classication

— Regression

Notes: e table summarizes the key characteristics of major AI techniques and the branch of tex-

tual analysis approaches known as NLP. For each technique, the table details the key strengths,

weaknesses, and areas of application.

Articial Intelligence in Asset Management

however, the variability of predictions can also increase. Furthermore, with

too many parameters, the model can overt to the data, modeling noise

rather than the underlying trend and leading to poor generalization to unseen

data. e LASSO method (Tibshirani 1996; James, Witten, Hastie, and

Tibshirani 2017) improves over standard linear or nonlinear models by addi-

tionally penalizing model complexity. By aiming to set some of the model’s

parameters to zero, LASSO regression automatically identies the most rel-

evant data features.

A.2.2. Articial Neural Networks and Deep Learning. ANNs are

inspired by biological brains (Aggarwal 2018; Haykin 2009). Similar to

what occurs in biological neurons, in each node of an ANN, input signals

are aggregated and processed, and the result is forwarded to other nodes.

ANNs are often arranged in feedforward layers (Figure A.1), with input data

applied to an input layer and further processed by a number of hidden layers

before arriving at an output layer. ANNs with many hidden layers are called

“deep neural networks.” ANNs are trained by altering the weights of the con-

nections so that the errors between the predicted and desired data labels are

minimized. Once trained, the ANN can be used to predict the output of

previously unseen input data.

Figure A.1. Feedforward ANN

Input

Layer

Hidden

Layers

Output

Layer

Note: e gure illustrates a fully connected feedforward neural network. Output and input vari-

ables are linked through several layers of interconnected nodes.

Appendix A. Basic Articial Intelligence Concepts and Techniques

A fundamental concern among AI practitioners and researchers is the

interpretability of trained ANNs. Presently, determining how a given neural

network settles on its predictions is not easy. For some elds, this is not an

issue. In others, however, such as medical data analysis, doctors are reluctant

to use a neural network without a complete understanding of the mecha-

nisms by which the network arrives at a prediction. ANNs also require fairly

large datasets to train properly, which may not be available for all assets or

markets.

A.2.3. Decision Trees and Random Forests. A decision tree sequen-

tially splits a dataset into increasingly small subsets typically based on

a single feature value. For the leaf nodes, the predicted class is determined

(Figure A.2). Besides classication, decision trees can also be used for piece-

wise linear regression. Decision trees are a form of supervised learning, and the

tree is constructed to replicate the labels in the training data as best as possible.

A random forest (Breiman 2001) is an ensemble of classication and

regression methods that consists of many decision trees. Each decision tree

makes predictions and contributes to the random forest’s predictions via an

averaging procedure. e assumption is that many decision tree models, each

with a slightly dierent perspective, make better predictions than a single

Figure A.2. Decision Tree Classier

Annual

Income

<=$30,000

<=25 >25 <=600 >600

High School University

>$30,000

Age

High Risk Low Risk

Low Risk High Risk Education Low Risk

Credit Score

Note: Shown is a decision tree to classify credit applications into low and high risk. Starting from

the top, applications move down the tree based on their characteristics. For example, an applica-

tion of someone with an income less than or equal to $30,000 annually and who is more than

25 years old would be classied as high risk.

Articial Intelligence in Asset Management

decision tree. Individual trees are generated from a distinct sampling of the

training set while using a random subset of the attributes for nodes. Consensus

voting on tree classications determines the prediction (Figure A.3). A key

advantage decision trees have over other AI techniques, such as ANNs, is

that the rules they use to classify data are human readable, and thus the rea-

sons that lead to a particular classication can be easily traced.

A.2.4. Support Vector Machines. SVMs are supervised algorithms

that are typically used for classication (Vapnik 2000; James et al. 2017).

ese algorithms learn boundaries that partition the feature space into two or

more classes. Once dened, the boundaries can be used to classify new data.

SVMs are powerful, accurate tools for both classication and regression, and

they are resistant to overtting their training data. ey are computationally

intensive, however, and thus do not scale well to large datasets.

Figure A.4, Panel A, presents an example of linearly separable training

data. e separation of the data into two classes is dened by the solid line,

which has been chosen to maximize its distance from the nearest data point

in each class (indicated by the dashed lines). In this example, the SVM has

learned the linear boundary that most eectively separates the data. In prac-

tice, training data are not always distributed in a way that permits separa-

tion by a linear boundary. In such instances, kernel methods are used to nd

more complex separation boundaries (Cortes and Vapnik 1995). Figure A.4,

Figure A.3. Random Forest Voting Scheme for Classication

Instance

...

Random Forest

Majority Voting

Final Class

Tree-1

Class-A

Tree-2

Class-B

Tree-n

Class-B

Note: e gure illustrates a random forest voting scheme that makes a prediction. e random

forest combines the output of n decision trees and yields a nal output that the majority of trees

agree on.

Appendix A. Basic Articial Intelligence Concepts and Techniques

Panel B, shows a nonlinear boundary learned by an SVM with the radial

basis function kernel.

A.2.5. Cluster Analysis. Cluster analysis is an unsupervised learn-

ing technique that seeks to partition a dataset into groups or clusters (Tan,

Steinbach, Karpatne, and Kumar 2018; Aggarwal and Reddy 2014). Once the

clusters are identied and labeled, the model can be used to classify new data.

e same data may yield many dierent possible sets of clusters, depending

on the number of clusters desired, how the data are distributed, and the clus-

tering algorithm. Applications in asset management include cluster analysis

of markets, companies, nancial instruments, time series, and documents.

e most popular clustering algorithm is K-means clustering, which

requires the user to specify the desired number of clusters, K. e method

starts by randomly choosing some cluster centroids and allocating each data

point to the closest centroid. It subsequently alternates between moving each

centroid to the center of its data cluster and reassigning the data points.

A.2.6. Evolutionary (Genetic) Algorithms. An evolutionary algorithm

(often also called a genetic algorithm) is an optimization algorithm based on

Darwin’s theory of evolution by natural selection (Eiben and Smith 2015;

Simon 2013). Its basic operations are depicted in Figure A.5. e evolution-

ary algorithm starts with an initial population of candidate solutions, usually

Figure A.4. SVM Examples

A. SVM with Linear Kernel

B. SVM with Nonlinear Kernel

Notes: e gure illustrates an example of using an SVM to separate data into two groups. e

dashed lines are the support vectors. SVMs often transform data by adding an additional dimen-

sion, making the classication of the data points easier. ese transformations are called kernels.

Panel A uses a linear kernel, and Panel B uses a nonlinear one.

Articial Intelligence in Asset Management

generated randomly. In each iteration, new candidate solutions are created by

selecting a pair of better performing solutions in the population, merging the

information from these two solutions into one (recombination), and intro-

ducing small random perturbations (mutation). e new solutions replace

older, worse-performing solutions in the population. Via an iterative process

of varying previously found well-performing solutions and selectively keeping

better ones, increasingly better solutions are found over time.

Because evolutionary algorithms do not require a mathematical formula-

tion of the problem and do not make assumptions such as convexity or linear-

ity about the objective function, they can also be applied to complex problems

in which other optimization algorithms fail. For example, evolutionary algo-

rithms have been used to solve mean–variance portfolio selection problems

under cardinality constraints (restricting the number of assets in the portfolio).

A.2.7. Natural Language Processing. NLP is a group of computa-

tional methods that can process or generate human natural language such as

text or speech (Manning and Schütze 1999; Mitkov 2014). ese methods

include voice recognition, which converts spoken language to text; speech

generation, which converts text to speech; natural language understanding,

which extracts meaning from spoken or written text; and natural language

generation, which produces natural language data from other data sources.

Owing to the ubiquity of social media and the countless natural language

samples it provides, the amount of literature on NLP has increased exponen-

tially in recent years (Xing, Cambria, and Welsch 2018). Various SVMs and

neural network architectures, such as deep learning or recurrent networks,

are frequently used in NLP.

Figure A.5. Iteration of an Evolutionary Algorithm

Population

Offspring

Parents

Select

Recombine

Mutate

Replace

Note: Illustrated is one iteration of an evolutionary algorithm: Better-performing solutions are

selected as parents, then used to create new solutions (ospring) by means of recombination and

mutation, and nally, the ospring replaces solutions in the population.

Appendix A. Basic Articial Intelligence Concepts and Techniques

In nance, much attention is focused on natural language nancial fore-

casting, which extracts information such as sentiment from nancial news or

social media data and incorporates it into models that predict the movement of

nancial data. Most of the literature applies NLP to stock and foreign exchange

rate prediction because of the accessibility of information for these markets.

A.2.8. Comparisons of AI Techniques. AI algorithms and techniques

dier by the types of data on which they best operate; the kinds of predictions

they can make; the means by which they learn to t their data; the computa-

tional power required for training, testing, and deployment; and the ease with

which they can be scaled.

Data may be low dimensional, as in the case of housing price data that

depend on fewer than a dozen features, or high dimensional, as in the case of

an image consisting of millions of pixels that have values independent from

all others. Data that reside in just a few dimensions allow using simpler tech-

niques (e.g., LASSO, K-means clustering, K-nearest neighbors) for classica-

tion or regression. On the other hand, data that live in higher dimensions

suer from the curse of dimensionality: e more dimensions the machine

learning problem has, the more samples are required to make meaningful

predictions, and the more computationally intensive analyzing and modeling

the data become. Autoencoder neural networks are an example of an algo-

rithm that can project high-dimensional data into low dimensions, so that the

data can be modeled and predicted more eectively.

If the data are labeled, a wide variety of supervised learning algorithms

can be used to learn a good mapping from input to predicted output. If data

are unlabeled, unsupervised learning techniques, such as cluster analysis, can

be used to identify patterns in the data.

Some algorithms are versatile, whereas others are relatively restricted. For

instance, K-means clustering is an algorithm designed to partition data into

K subsets in an unsupervised fashion only. Neural networks, on the other

hand, are more broadly applicable. ey can be used for unsupervised and

supervised learning, classication, and regression, and they can be applied to

images, text, and time series.

e interpretability of a mathematical model varies widely. Linear regres-

sion models, logistic regression models, decision trees, and K-nearest neigh-

bor classiers are examples of algorithms with readily understood learned

behavior. In contrast, a precise statistical description of the learning and pre-

diction decisions of neural networks is still an unresolved problem.

Some algorithms are dicult to deploy on large datasets as a result of

their computational complexity. Specialized hardware or graphic cards may

Articial Intelligence in Asset Management

speed up computations, such as for neural networks. Note that although

training is sometimes very time consuming, deploying a trained model for

predictions on unseen data is usually fast.

e business logic of asset management determines the structure of data-

sets and the type of predictions to be obtained. Consequently, the business

needs ultimately determine the choice of AI algorithm.

Appendix B. Trends and Patterns

in Finance Research Using AI

Developments in academic research related to AI can be analyzed by study-

ing trends and patterns in working papers in nance that use AI techniques.

To capture the most recent trends and those related to the area of nance,

we downloaded all working papers that have AI-related keywords in their

title, abstract, or listed keywords and that have been posted in the Financial

Economics Network (FEN) on the Social Sciences Research Network

(SSRN) between 1996 and 2018. e primary keywords are “articial intel-

ligence,” “machine learning,” “cluster analysis,” “genetic algorithm” or “evolu-

tionary algorithm,” “lasso,” “natural language processing,” “neural network”

or “deep learning,” “random forest” or “decision tree,” and “support vector

machine.”

On SSRN, 1,814 working papers include at least one of our keywords.

By far the most popular AI-related keyword, “machine learning” appears in

29% of all papers in our sample. Among AI techniques, “neural network” (or

“deep learning”), with 38% of the papers, and “cluster analysis,” with 16%, are

the two most popular keywords (Panel A of Figure B.1). Download numbers

magnify the relative popularity of neural networks even further, consider-

ing that papers with “neural network”/“deep learning” keywords account for

almost half of the total downloads of all papers related to the AI techniques

(Panel B of Figure B.1). “Random forest” (or “decision tree”), with only 16%,

has the second highest proportion of downloads. Finally, “natural language

processing” is associated with only 4% of the papers and 3% of downloads,

the lowest among all techniques.

Interestingly, no working papers with a “machine learning” keyword

appear until 2003. Because the terms “machine learning” and “articial intel-

ligence” have gained popularity more recently, we sum the number of papers

for all AI-related keywords. In 1996, no working paper with any AI-related

keyword was uploaded to SSRN. Signicant growth has occurred since,

however, with 410 papers posted in 2018. e total number of papers on FEN

also increased during the same period, but to a much smaller degree: Papers

with AI-related keywords accounted for 3% of all papers in 2018.

Among the AI techniques, the popularity of “neural network” is consis-

tent over the years. “Lasso” and “support vector machine” techniques seem to

have gained popularity more recently, however. e same applies to “natural

Articial Intelligence in Asset Management

Figure B.1. Number and Downloads of Finance Papers Using AI, 1996–2018

169, 16%

101, 9%

133, 12%

46, 4%

415, 38%

155, 14%

72, 7%

22,602, 7%

26,070, 8%

32,309, 9%

11,804, 3%

171,517, 49%

55,545, 16%

27,208, 8%

A. Number of Papers

B. Number of Downloads

Cluster Analysis

Genetic (or Evolutionary) Algorithm

LASSO

Natural Language Processing

Support Vector Machine

Random Forest or Decision Tree

Neural Network or Deep Learning

Notes: e gure shows the number of papers (Panel A) and downloads (Panel B) associated with

various AI techniques for the sample of SSRN FEN working papers. e sample includes papers

having one of the following keywords in their abstract, title, or keyword section: “articial intel-

ligence,” “machine learning,” “cluster analysis,” “genetic algorithm” or “evolutionary algorithm,”

“lasso,” “natural language processing,” “neural network” or “deep learning,” “random forest” or

“decision tree,” and “support vector machine.”

Appendix B. Trends and Patterns in Finance Research Using AI

language processing”: No papers with this keyword appear on SSRN until

2008, but 13 working papers with this keyword were uploaded in 2018.

Among countries, the United States, the United Kingdom, and Germany

are the top three producers of papers with AI-related keywords (Figure B.2).

ese three countries account for more than half of all AI papers, followed

by Switzerland, India, France, China, and Italy. Not surprisingly, these coun-

tries also show high productivity in terms of nance papers generally.

Among institutions, Cornell University, Humboldt University of Berlin,

the University of Chicago, Stevens Institute of Technology, and ETH Zurich

are the top ve contributors to AI research papers (Table B.1). A large num-

ber of nonacademic institutions are also in the full list, illustrating the inter-

est in conducting research related to this topic in the nance industry and

central banks.

Figure B.2. Number of Papers Using AI by Country, 1996–2018

Notes: e gure shows a heat map based on the number of papers with AI-related keywords

for each country. A paper belongs to a specic country if at least one of its authors is aliated

with institutions located in that country. e sample includes papers having one of the following

keywords in their abstract, title, or keyword section: “articial intelligence,” “machine learning,”

“cluster analysis,” “genetic algorithm” or “evolutionary algorithm,” “lasso,” “natural language pro-

cessing,” “neural network” or “deep learning,” “random forest” or “decision tree,” and “support

vector machine.”

Articial Intelligence in Asset Management

Table B.1. Number of Papers by Institution, 1996–2018

Institution/University Sum of AI Papers

Independent 75

Aliation not provided to SSRN 67

Cornell University 64

Humboldt University of Berlin 47

University of Chicago 42

Stevens Institute of Technology 41

ETH Zurich 39

New York University 30

Imperial College London 26

Harvard University 22

University of St. omas 21

London School of Economics and Political Science 19

Other Institutions 2,937

Notes: e table presents the number of SSRN FEN working papers by institution. A paper belongs

to an institution if at least one of its authors has an aliation with that institution. e sample

includes papers having one of the following keywords in their abstract, title, or keyword section:

“articial intelligence,” “machine learning,” “cluster analysis,” “genetic algorithm” or “evolutionary

algorithm,” “lasso,” “natural language processing,” “neural network” or “deep learning,” “random

forest” or “decision tree,” and “support vector machine.” Institutions with fewer than 15 papers are

combined and reported as “Other Institutions.”

References

Abe, Masaya, and Hideki Nakayama. 2018. “Deep Learning for Forecasting

Stock Returns in the Cross-Section.” In Advances in Knowledge Discovery

and Data Mining, Part 1, edited by Dinh Phung, Vincent S. Tseng, Georey

I. Webb, Bao Ho, Mohadeseh Ganji, and Lida Rashidi, 273–84. Cham,

Switzerland: Springer International Publishing.

e authors examine the performance of deep learning for one-month-

ahead forecasting of stock returns in the cross-section in the Japanese stock

market. Deep neural networks (DNNs) are found to perform better than

shallow neural networks and other ML techniques, such as support vector

regression (SVR) and random forests. An ensemble of SVR, random for-

ests, and DNN can provide a further small gain in performance.

Aggarwal, Charu C. 2018. Neural Networks and Deep Learning. Cham,

Switzerland: Springer. https://doi.org/10.1007/978-3-319-94463-0.

is textbook, aimed at graduate students, researchers, and practitioners,

is devoted to studying neural networks, or “deep learning,” as it became

known in ML literature. It starts with background material on neural net-

work design and its relationship to traditional ML algorithms. e rest

of the textbook provides a more detailed discussion of the fundamentals

of neural networks (common neural architectures, shallow neural net-

works and DNNs, radial basis function networks, restricted Boltzmann

machines) and the recent advances in the eld (recurrent and convolutional

neural networks, deep reinforcement learning, neural Turing machines,

generative adversarial networks).

Aggarwal, Charu C., and Chandan K. Reddy, eds. 2014. Data Clustering:

Algorithms and Applications. Boca Raton, FL: CRC Press.

Data clustering is partitioning data into multiple groups where each obser-

vation shares some characteristics with other data points within its group.

Data clustering issues have been studied in dierent elds, and this book

attempts to bridge the gap by uniting the knowledge gained in ML and

data mining literature. e rst part of the book discusses data clustering

methods (probabilistic, distance based, density based, grid based, and spec-

tral clustering, which clusters high-dimensional and big data). e second

part is devoted to the application of data clustering in dierent elds and

situations (clustering categorical, text, multimedia, time-series, biological,

Articial Intelligence in Asset Management

network, and uncertain data). Finally, the book focuses on the appropri-

ateness of a particular clustering method through evaluation techniques

available from dierent variations of cluster analysis (visual, supervised,

ensemble, and validation-based solutions).

Ahmed, Nesreen K., Amir F. Atiya, Neamat El Gayar, and Hisham

El-Shishiny. 2010. “An Empirical Comparison of Machine Learning Models

for Time Series Forecasting.” Econometric Reviews 29 (5–6): 594–621. https://

doi.org/10.1080/07474938.2010.481556.

e authors conduct a comparative analysis of the forecasting performance

of several ML methods (multilayer perceptron, Bayesian neural networks,

radial basis functions, generalized regression neural networks, K-nearest

neighbor regression, classication and regression trees, SVR, and Gaussian

process) for time-series data. Using a monthly time-series data sample, the

results show that multilayer perceptron and the Gaussian process regres-

sion outperform the other ML methods considered, indicating that the

rarely used Gaussian process method may oer performance improvement

in forecasting problems.

Ahn, Jae Joon, Kyong Joo Oh, Tae Yoon Kim, and Dong Ha Kim. 2011.

“Usefulness of Support Vector Machine to Develop an Early Warning System

for Financial Crisis.” Expert Systems with Applications 38 (4): 2966–73. https://

doi.org/10.1016/j.eswa.2010.08.085.

Existing studies on early warning systems for nancial crises are based on

a belief system that considers nancial crises to be self-fullling as a result

of investor herding behavior. In this article, the authors take the alterna-

tive view that banking crises result from deteriorating long-term economic

fundamentals. With this new hypothesis, the authors nd that SVM is an

eective technique for this classication problem.

Alberg, John, and Zachary C. Lipton. 2017. “Improving Factor-Based

Quantitative Investing by Forecasting Company Fundamentals,” version 2.

arXiv.org/abs/1711.04837v2.

Fundamental nancial data of publicly traded companies, such as oper-

ating income, revenue, assets, dividends, and debt, provide a picture of

those companies’ nancial performance. If the future fundamentals can be

forecasted, then an investment strategy can build portfolios based on the

predicted fundamentals that can outperform a standard factor approach.

Using data on stocks for publicly traded US companies, the authors

develop and forecast features from the fundamentals and evaluate the

References

out-of-sample forecasting performance using two classes of DNNs (mul-

tilayer perceptrons and recurrent neural networks). is lookahead factor

model with DNNs signicantly outperforms the naive prediction (i.e.,

fundamentals remain unchanged). In addition, a simulated hypothetical

stock portfolio using an industry-grade stock portfolio simulator shows a

performance improvement over traditional factor models in terms of com-

pounded annual returns.

Allen, Franklin, and Risto Karjalainen. 1999. “Using Genetic Algorithms to

Find Technical Trading Rules.” Journal of Financial Economics 51 (2): 245–71.

https://doi.org/:10.1016/S0304-405X(98)00052-X.

In this article, the authors examine the S&P 500 composite index from

1928 to 1995 to generate trading rules using genetic algorithms. Similar

to what past ndings have revealed, trading rules created by genetic algo-

rithms do not make money after controlling for transaction costs. Genetic

algorithm-based trading rules can correctly identify periods of high return

and low volatility and periods with opposite conditions.

Alpaydin, Ethem. 2010. Introduction to Machine Learning, 2nd ed. Cambridge,

MA: MIT Press.

is introductory textbook on ML covers relevant algorithms and meth-

ods from computer science, neural computation, and statistics. e top-

ics include, but are not limited to, supervised learning, Bayesian decision

theory and Bayesian estimation, parametric and nonparametric methods,

multivariate methods, dimensionality reduction, decision trees, linear dis-

crimination, kernel machines, hidden Markov models, graphical models,

combining multiple learners, reinforcement learning (RL), and the design

and analysis of ML experiments.

Arrieta-ibarra, Imanol, and Ignacio N. Lobato. 2015. “Testing for

Predictability in Financial Returns Using Statistical Learning Procedures.”

Journal of Time Series Analysis 36 (5): 672–86. https://doi.org/10.1111/

jtsa.12120.

e authors examine and compare the out-of-sample forecasting perfor-

mance of three ML techniques (regression trees, neural networks, and

SVMs) with that of a traditional GARCH(1,1) model in predicting daily

stock market returns. ey conclude that in terms of predictive accuracy,

the three ML techniques do not outperform naive historical sample aver-

age returns when only past values of the series are considered. Adding other

variables (interest rates, commodity prices, or exchange rates) to the model

Articial Intelligence in Asset Management

may lead to model overtting and insignicant results. However, SVMs

and neural networks oer potential improvements over a GARCH(1,1)

model only for forecasting squared returns.

Atsalakis, George S., and Kimon P. Valavanis. 2009. “Surveying Stock

Market Forecasting Techniques – Part II: Soft Computing Methods.”

Expert Systems with Applications 36 (3): 5932–41. http://doi.org/10.1016/j.

eswa.2008.07.006.

e authors provide a survey of the literature on stock return forecasting in

dierent markets, revealing that recently, soft computing techniques have

increasingly been used for predicting stock market behavior. Neural net-

works, fuzzy logic, and genetic algorithm are some examples of soft com-

puting techniques. In the prior literature on stock return forecasting, soft

computing techniques generally outperform traditional models in terms of

forecast accuracy.

Auria, Laura, and Rouslan A. Moro. 2008. “Support Vector Machines

(SVM) as a Technique for Solvency Analysis.” DIW Berlin Discussion Paper

No. 811 (August). https://ssrn.com/abstract=1424949.

e authors propose the use of SVMs in predicting company solvency.

Using data on German companies, the SVM technique outperforms alter-

native approaches, such as logistic regression and discriminant analysis, in

out-of-sample accuracy.

Avramov, Doron, Si Cheng, and Lior Metzker. 2019. “Machine Learning

versus Economic Restrictions: Evidence from Stock Return Predictability.”

https://ssrn.com/abstract=3450322.

e researchers argue that ML methods, such as deep learning, often

do not clear economic restrictions the way that value-weighting returns,

excluding microcaps or distressed rms, do. When economic restrictions

are imposed, value-weighted portfolio returns decline relative to equal-

weighted portfolio returns. Upon further economic restrictions, the value-

weighted portfolio returns attenuate further once microcaps, non-rated

rms, and distressed rms are excluded. In the presence of transaction

costs, performance may further deteriorate. Nevertheless, ML methods are

useful, especially in long positions and because of their low downside risk.

Azimi, Mehran, and Anup Agrawal. 2019. “Is Positive Sentiment in

Corporate Annual Reports Informative? Evidence from Deep Learning.”

https://ssrn.com/abstract=3258821.

References

e authors use deep learning for text classication of qualitative informa-

tion found in corporate annual reports. e research ndings indicate that,

unlike ndings in prior literature, both positive and negative sentiments

are important in predicting abnormal returns and abnormal trading vol-

ume around 10-K ling dates. For instance, positive (negative) sentiment

is associated with higher (lower) abnormal returns and with lower (higher)

abnormal trading volume around 10-K ling dates.

Aziz, Saqib, and Michael Dowling. 2019. “Machine Learning and AI for

Risk Management.” In Disrupting Finance: FinTech and Strategy in the 21st

Century, edited by eo Lynn, John G. Mooney, Pierangelo Rosati, and

Mark Cummins, 33–50. Cham, Switzerland: Palgrave Pivot. https://doi.

org/10.1007/978-3-030-02330-0_3.

e authors provide an overview of the application of ML and AI tech-

niques in the risk management domain. ey review the relevant literature

and discuss the particular techniques used in the areas of credit risk, mar-

ket risk, operational risk, and compliance.

Ballings, Michel, Dirk Van den Poel, Nathalie Hespeels, and Ruben Gryp.

2015. “Evaluating Multiple Classiers for Stock Price Direction Prediction.”

Expert Systems with Applications 42 (20): 7046–56. https://doi.org/10.1016/j.

eswa.2015.05.013.

In this study, the authors compare ensemble ML methods (Random Forest,

AdaBoost, Kernel Factory) with single classiers (Neural Networks,

Logistic Regression, Support Vector Machines, K-Nearest Neighbor) with

respect to their ability to predict the direction of stock price movements.

Based on AUC (areas under the receiver operating characteristic curve)

as a performance measure, the algorithms are ranked as follows: Random

Forests, Support Vector Machines, Kernel Factory, AdaBoost, Neural

Networks, K-Nearest Neighbor, and Logistic Regression. e study con-

cludes that ensemble models should not be neglected in predicting stock

price direction.

Bao, Yang, and Anindya Datta. 2014. “Simultaneously Discovering and

Quantifying Risk Types from Textual Risk Disclosures.” Management Science

60 (6): 1371–91. https://doi.org/10.1287/mnsc.2014.1930.

e authors use a latent Dirichlet allocation topic model to simultaneously

identify and measure the risk types from risk disclosure information in

10-K forms. Empirical results indicate that the presented textual analy-

sis method outperforms all competing methods. Further, the proposed

Articial Intelligence in Asset Management

technique identies that only one-third of risk types are informative.

Among the risk types that are found to be informative, only systematic

and liquidity-related risks will increase investors’ risk perceptions, whereas

nonsystematic risk types will actually decrease them.

Becker, Sebastian, Patrick Cheridito, and Arnulf Jentzen. 2019a. “Deep

Optimal Stopping.” Journal of Machine Learning Research 20: 1–25.

In this article, the authors develop an optimal stopping strategy using deep

learning. ey construct lower and upper bounds, point estimates, and

condence intervals for the price. e authors then test the approach in

three dierent examples: the pricing of a Bermudan max-call option, the

pricing of a callable multi-barrier reverse convertible, and the problem of

optimally stopping a fractional Brownian motion. e results show that in

all three exercises, the proposed method produces highly accurate results

with short computing times.

Becker, Sebastian, Patrick Cheridito, and Arnulf Jentzen. 2019b. “Pricing

and Hedging American-Style Options with Deep Learning.” https://arxiv.

org/abs/1912.11060.

In this article, the authors consider American-style options and develop

a DNN for optimal exercise behavior, pricing, and hedging. In the rst

stage, the neural network method is used to nd the optimal stopping

rule and derive a lower bound estimate of the price. In the next stage, it

estimates an upper bound and condence intervals for the price. Finally,

the deep learning method constructs a dynamic hedging strategy.

Exercises based on Bermudan max-call options show that the proposed

method can achieve accurate price predictions and provide dynamic

hedging strategies.

Beketov, Mikhail, Kevin Lehmann, and Manuel Wittke. 2018. “Robo

Advisors: Quantitative Methods Inside the Robots.” Journal of Asset

Management 19: 363–70. http://doi.org/10.1057/s41260-018-0092-9.

e authors explore the methods of robo-advisors around the world and

nd modern portfolio theory to be the predominant method used in robo-

advisory systems. Further, the volume of assets managed by robo-advisors

is likely to be higher when more sophisticated systems are used.

Bellovary, Jodi L., Don E. Giacomino, and Michael D. Akers. 2007. “A

Review of Bankruptcy Prediction Studies: 1930 to Present.” Journal of

Financial Education 33: 1–42.

References

For this study, the authors review bankruptcy prediction literature pub-

lished since the 1930s. Bankruptcy prediction studies in earlier years pri-

marily used simple ratio analysis, whereas discriminant analysis became

more prevalent in the 1960s and 1970s. In the 1980s and 1990s, logit and

neural networks methods gained more popularity in bankruptcy analysis.

Comparisons of dierent methods reveal that multivariate discriminant

analysis and neural networks deliver better bankruptcy prediction mod-

els. In addition, the authors note that a model’s accuracy may not always

increase with the number of factors used in the model.

Bertsimas, Dimitris, and Andrew W. Lo. 1998. “Optimal Control of

Execution Costs.” Journal of Financial Markets 1 (1): 1–50. https://doi.

org/10.1016/S1386-4181(97)00012-8.

e performance of investments is aected by the presence of trading or

execution costs, which include commissions, bid–ask spreads, and oppor-

tunity costs. In this article, the authors present a dynamic trading strategy

that seeks to minimize expected trading costs. e cost-minimizing strat-

egy is a function of time and several state variables, so it can optimally

adapt to changing market conditions and price movements. Compared

with a naive trading strategy, the best-execution strategy reduces execution

costs by 25% to 40%.

Berutich, José Manuel, Francisco López, Francisco Luna, and David

Quintana. 2016. “Robust Technical Trading Strategies Using GP for

Algorithmic Portfolio Selection.” Expert Systems with Applications 46 (15):

307–15. https://doi.org/10.1016/j.eswa.2015.10.040.

e authors of this study examine a genetic algorithm approach for gen-

erating protable trading rules in the Spanish market. e article’s main

contribution is the introduction of a new random sampling method that

improves the out-of-sample results of the genetic algorithm compared with

traditional approaches. is method can also be generalized to cope with

dierent types of markets. Empirical tests of the strategies generated using

this approach conrm the superiority of the approach over a range of tra-

ditional alternatives.

Bew, David, Campbell R. Harvey, Anthony Ledford, Sam Radnor, and

Andrew Sinclair. 2019. “Modeling Analysts’ Recommendations via Bayesian

Machine Learning.” Journal of Financial Data Science 1 (1): 75–98.

Individual analysts focus on only a small number of thousands of compa-

nies in making recommendations on stock performance. In this study, the

Articial Intelligence in Asset Management

authors propose a novel approach for combining individual analyst recom-

mendations using an independent Bayesian classier combination (IBCC)

model that dynamically adjusts based on the length and quality of an ana-

lyst’s track record. e forecasts are obtained from the probabilistic IBCC

model using data from the pan-European region from 2004 to 2013. e

results indicate that the best outcome is achieved when the analyst/broker

recommendation is consistent with forecasts from the IBCC model.

Bholat, David M., Stephen Hansen, Pedro Santos, and Cheryl Schonhardt-

Bailey. 2015. “Text Mining for Central Banks.” http://dx.doi.org/10.2139/

ssrn.2624811.

Text mining can be used to extract meaning from texts. is handbook

illustrates the value of text mining methods to central banks. e rst

part of the handbook reviews the text mining literature, examining how

it could be applied to central bank policymaking. e second part explains

the implementation of popular text mining techniques, such as Boolean

and dictionary text mining, latent semantic analysis, latent Dirichlet allo-

cation, and descending hierarchical classication.

Bianchi, Daniele, Matthias Büchner, and Andrea Tamoni. 2019. “Bond Risk

Premia with Machine Learning.” https://doi.org/10.2139/ssrn.3400941.

e authors compare the performance of ML methods in predicting bond

returns. Compared with classical approaches, such as dense and sparse lin-

ear regression frameworks, DNNs perform signicantly better in forecast-

ing excess bond returns. Empirical exercises using both one-year-ahead

and one-month-ahead forecasts show that neural networks outperform

other classical approaches largely because of their ability to capture nonlin-

earities in the data.

Black, Fischer, and Robert Litterman. 1992. “Global Portfolio Optimization.”

Financial Analysts Journal 48 (5): 28–43. https://doi.org/10.2469/faj.v48.n5.28.

In this article, the authors discuss the limitations of modern portfolio

management theory in global portfolio management. e main limitation

of classic portfolio management theory is the diculty of its implemen-

tation and its unexpected poor behavior when used for global portfolios.

e authors suggest combining Markowitz’s mean–variance portfolio opti-

mization model with the CAPM of Sharpe and Lintner to improve the

portfolio allocation behavior.

Board of Governors of the Federal Reserve System. 2011. “Supervisory

Guidance on Model Risk Management.” Oce of the Comptroller of the

References

Currency (April 4). https://www.occ.treas.gov/news-issuances/bulletins/2011/

bulletin-2011-12a.pdf.

Financial institutions have been using models in many aspects of banking,

but models also carry risks and costs. is paper presents an overview of

model risk management, robust model development, model validation, and

good practices in model development, implementation, and application.

Booth, Ash, Enrico Gerding, and Frank McGroarty. 2014. “Automated

Trading with Performance Weighted Random Forests and Seasonality.”

Expert Systems with Applications 41 (8): 3651–61. https://doi.org/10.1016/j.

eswa.2013.12.009.

e authors of this study present an automated trading system based on

performance-weighted ensembles of random forests. e proposed method

uses seasonal trends to improve the predictability of stock returns. Using

data on the German stock index DAX, the proposed new technique shows

promising results for out-of-sample forecasts. Specically, it achieves better

prediction accuracy and higher protability than other ensemble methods.

Booth, Ash, Enrico Gerding, and Frank McGroarty. 2015. “Performance-

Weighted Ensembles of Random Forests for Predicting Price Impact.”

Quantitative Finance 15 (11): 1823–35. https://doi.org/10.1080/14697688.20

14.983539.

In this study, the researchers present a forecasting model based on an

ensemble of random forests. e empirical results show that the proposed

model outperforms other popular competing models—linear regression,

neural networks, and SVR—in out-of-sample forecasts. e ensemble of

random forests improves forecast accuracy by at least 15% compared with

the competing models.

Borghi, Riccardo, and Giuliano De Rossi. Forthcoming. “e Articial

Intelligence Approach to Picking Stocks.” In Machine Learning and Asset

Management. Cham, Switzerland: Springer.

e chapter’s authors explore ML techniques to select from a large set of

rm characteristics to predict one-month-ahead stock returns using data

for US and European stocks. ey use a proprietary alpha model, a global

stock selection model from Macquarie, as a benchmark for comparing the

results from traditional ML techniques (ordinary least squares, general-

ized linear model, LASSO, ridge, elastic net, neural networks, boosted

trees and random forest, and regression trees). e results indicate that

an ensemble of ML methods could outperform even a proprietary model

Articial Intelligence in Asset Management

with a successful track record, such as Macquarie’s Alpha Quant model, in

terms of generating attractive risk–return ratios. However, ML tend to be

less exposed to price momentum than alpha models are.

Branke, Juergen, Benedikt Scheckenbach, Michael Stein, Kalyanmoy

Deb, and Hartmut Schmeck. 2009. “Portfolio Optimization with an

Envelope-Based Multi-Objective Evolutionary Algorithm.” European

Journal of Operational Research 199 (3): 684–93. https://doi.org/10.1016/j.

ejor.2008.01.054.

e mean–variance-based portfolio optimization problem with linear

constraints can be solved eciently using parametric quadratic program-

ming. Many real-world constraints tend to be more complex and discrete,

however, such as when the number of stocks in a portfolio is limited. In

this study, the authors combine parametric quadratic programming with

a multi-objective evolutionary algorithm to nd the best solution to the

nonconvex portfolio selection. As the results demonstrate, the proposed

multi-objective evolutionary algorithm model signicantly outperforms

other existing evolutionary algorithms.

Breiman, Leo. 2001. “Random Forests.” Machine Learning 45 (1): 5–32.

https://doi.org/10.1023/A:1010933404324.

In this seminal work, the author introduces random forests, which are col-

lections of tree predictors used in a classication problem. As the number

of trees increase and given the law of large numbers, random forests do not

suer from overtting problems and are an eective forecasting tool used

in ML.

Brenner, Lukas, and Tobias Meyll. 2019. “Robo-Advisors: A Substitute for

Human Financial Advice?” https://ssrn.com/abstract=3414200.

With recent technological advances in the nancial services industry, the

use of automated nancial advisors (i.e., robo-advisors) has been increas-

ing. e authors investigate the determinants of investor decisions to opt

for robo-advisory services. Using representative investor survey data, the

authors show that a signicant substitution is taking place because inves-

tors are worried about potential conicts of interest with human nancial

advisors.

Briere, Marie, Charles-Albert Lehalle, Tamara Nefedova, and Amine

Raboun. 2019. “Modelling Transaction Costs When Trades May Be Crowded:

A Bayesian Network Using Partially Observable Orders Imbalance.” https://

ssrn.com/abstract=3420665.

References

e authors use a Bayesian network model to forecast transaction costs

measured as implementation shortfall. e Bayesian network approach can

account for missing data by using the most probable values instead. is

feature is particularly useful for integrating important latent variables, such

as the net order ow imbalance of investors to improve the transaction cost

forecasts. e ndings suggest that implementation shortfall forecasts are

more accurate when investors’ order size is larger or when stock volatility

is lower.

Bryzgalova, Svetlana, Markus Pelger, and Jason Zhu. 2019. “Forest rough

the Trees: Building Cross-Sections of Stock Returns.” https://ssrn.com/

abstract=3493458.

e authors describe a new method of building basis assets that capture

the underlying information in the cross-sections of asset returns. e novel

approach allows us to capture the complex information of cross-sectional

stock characteristics. e proposed method applies asset-pricing trees to

build portfolios that capture all relevant information, allow for nonlineari-

ties and interactions, act as building blocks for a stochastic discount fac-

tor, and provide test assets for asset pricing. e empirical results using

monthly equity returns for all CRSP securities show that, unlike the pro-

posed approach, conventional cross-sectional sorting methods fail to fully

capture the relevant information on stock characteristics.

Buchanan, Bonnie G. 2019. “Articial Intelligence in Finance.” http://doi.

org/10.5281/zenodo.2612537.

In this literature review, the author overviews AI, ML, and deep learning

(DL) and their applications in the nancial services sector. She rst dis-

cusses how rapidly growing AI is changing the nancial services industry by

using ML algorithms to analyze millions of data points to detect fraudulent

activity, by using “robo-advising” that avoids human interaction and conict

of interest, and by using algorithmic trading to make fast trading decisions.

e author then discusses the dierences between ML and econometrics

and the applications of each. Specically, ML and DL methods focus on

predictive accuracy, whereas econometric methods are mainly concerned

with inferential questions. ML and DL techniques have become powerful

tools for out-of-sample forecasts and for identifying the most useful predic-

tors. Further, the author briey reviews quantum computing and how it can

process data at speeds impossible for traditional computers. Finally, as AI

and ML are increasingly applied in the nancial services industry, regula-

tors are having diculty keeping pace with the new technology.

Articial Intelligence in Asset Management

Celik, Arzum Erken, and Yalcin Karatepe. 2007. “Evaluating and Forecasting

Banking Crises rough Neural Network Models: An Application for

Turkish Banking Sector.” Expert Systems with Applications 33 (4): 809–15.

https://doi.org/10.1016/j.eswa.2006.07.005.

e authors of this study forecast the ratios of nonperforming loans to total

loans, capital to assets, prots to assets, and equity to assets for the Turkish

banking sector. e results indicate that ANNs that use the Taguchi

approach to determine the optimal values of ANN parameters are an eec-

tive tool in forecasting banking crises.

Chapados, Nicolas, and Yoshua Bengio. 2001. “Cost Functions and Model

Combination for VaR-Based Asset Allocation Using Neural Networks.”

IEEE Transactions on Neural Networks 12 (4): 890–906. https://doi.

org/10.1109/72.935098.

e authors consider value at risk (VaR), typically used in quantifying the

risk associated with a portfolio, and extend it to the mean–VaR framework,

similar in spirit to Markowitz’s famous mean–variance model. Mean–

VaR seeks to nd the best portfolio allocation for a given VaR level. e

authors build neural networks to optimize the portfolio of assets given the

VaR constraint and transaction costs and show that the portfolio’s perfor-

mance is comparable to that of a forecasting model based on the classical

mean–variance portfolio management but better than that of the bench-

mark market index. However, the proposed method relies on fewer model

assumptions and takes transaction costs into account. Finally, both the

proposed and the classical forecasting models can signicantly outperform

the benchmark market index when a forecast combination technique is

applied.

Chen, Luyang, Markus Pelger, and Jason Zhu. 2020. “Deep Learning in

Asset Pricing.” https://ssrn.com/abstract=3350138.

e article’s authors use DNNs and a no-arbitrage constraint to estimate

individual stock returns. Specically, they combine three dierent neural

networks (feedforward network, long–short-term-memory network, gen-

erative adversarial network) with a no-arbitrage condition to estimate the

stochastic discount factor that explains stock returns. e stochastic dis-

count factor portfolio uses time-varying weights for traded assets, which

are functions of rm-specic and macroeconomic variables. e proposed

stochastic discount factor asset pricing model outperforms other bench-

mark models out of sample.

References

Chen, Shiyi, Wolfgang K. Hardle, and Kiho Jeong. 2009. “Forecasting

Volatility with Support Vector Machine-Based GARCH Model.” Journal of

Forecasting 29 (4): 406–33.

e study authors compare the volatility forecasts generated from the

traditional volatility forecasting methods (moving average, GARCH,

exponential generalized autoregressive conditional heteroskedasticity

[EGARCH], ANN-GARCH) with the forecasts obtained using an SVM

or ANN. Using a recursive forecasting framework, one-step-ahead fore-

casts of the volatilities of the British exchange rate and NYSE index are

evaluated based on the Diebold–Mariano forecast evaluation method. Of

all the models considered, SVM-GARCH models perform well in many

cases, whereas a standard GARCH model performs well when the sample

size is large and normally distributed. On the other hand, an EGARCH

model better predicts the one-step ahead volatility when the data are

highly skewed.

Chen, Wun-Hua, Jen-Ying Shih, and Soushan Wu. 2006. “Comparison

of Support-Vector Machines and Back Propagation Neural Networks in

Forecasting the Six Major Asian Stock Markets.” International Journal of

Electronic Finance 1 (1): 49–67. https://doi.org/10.1504/IJEF.2006.008837.

Most of the existing research in the application of data mining techniques

to nancial time series forecasting focuses on the US or European mar-

kets. e authors of this study examine the performance of SVM and

back propagation neural networks in forecasting six Asian stock markets.

Compared with naive autoregressive (1) time-series models, both proposed

models perform well.

Cheng, Ching-Hsue, Tai-Liang Chen, and Liang-Ying Wei. 2010. “A Hybrid

Model Based on Rough Sets eory and Genetic Algorithms for Stock Price

Forecasting.” Information Sciences 180 (9): 1610–29. https://doi.org/10.1016/j.

ins.2010.01.014.

Professional fund managers do not easily understand time-series fore-

casting methods and AI techniques used in forecasting stock returns.

erefore, they often rely on subjective judgments in predicting stock

prices based on some technical indicators. e authors develop a hybrid

forecasting framework that uses the tools of rough set theory and genetic

algorithms. Empirical results show that the proposed hybrid techniques

outperform the rough set theory and genetic algorithms in terms of fore-

cast accuracy and prots.

Articial Intelligence in Asset Management

Chinco, Alexander M., Adam D. Clark-Joseph, and Mao Ye. 2019. “Sparse

Signals in the Cross-Section of Returns.” Journal of Finance 74 (1): 449–92.

https://doi.org/10.1111/jo.12733.

Using LASSO, a penalized regression technique, the authors study the

out-of-sample t of one-minute-ahead return forecasts. Evidence sug-

gests that this improved performance mainly results from the identica-

tion of predictors that are unexpected, short lived, and sparse. Further, the

LASSO method is found to increase the forecast-implied Sharpe ratio.

Choudhry, Tauq, Frank McGroarty, Ke Peng, and Shiyun Wang. 2012.

“High-Frequency Exchange-Rate Prediction with an Articial Neural

Network.” Intelligent Systems in Accounting, Finance & Management 19 (3):

170–78. https://doi.org/10.1002/isaf.1329.

e authors study the eectiveness of ANN for forecasting high-frequency

exchange rates, ranging from one minute to a few minutes. ey nd that

bid and ask prices are signicant variables for exchange rate forecasting.

e study concludes that high-frequency trading strategies using ANNs

can generate positive prot beyond transaction costs.

Cong, Lin, Tengyuan Liang, and Xiao Zhang. 2019. “Textual Factors:

A Scalable, Interpretable, and Data-Driven Approach to Analyzing

Unstructured Information.” https://ssrn.com/abstract=3307057.

Increased access to unstructured data in the form of texts can be used to

complement information obtained from the traditional structured data.

Extracting insight from textual data is more complex, however, because

of the data’s intricate language structure and high dimensionality, and

because of the lack of data-driven approaches to analyzing textual data.

e authors develop an alternative method of analyzing textual data that is

more eective than the existing methods. ey use neural network models

for NLP to generate textual factors and show the procedure’s application in

analyzing nancial and macroeconomic data.

Coqueret, Guillaume, and Tony Guida. 2018. “Stock Returns and the

Cross-Section of Characteristics: A Tree-Based Approach.” https://ssrn.com/

abstract=3169773.

e authors use regression trees that iteratively split the sample into clus-

ters so they can investigate the eects of 30 classical rm characteristics

on future returns. Unlike linear regression models, tree-based regression

models take into account the conditional impact of rm attributes given

the value of other attributes. e ndings indicate that both technical

References

indicators (e.g., the Relative Strength Index) and past performance indica-

tors are important in stock pricing. e authors estimate that a portfolio

built using a short-term Relative Strength Index characteristic leads to an

annual gain of 2.4% relative to a naive, equally weighted portfolio.

Cortes, Corinna, and Vladimir Vapnik. 1995. “Support-Vector Networks.”

Machine Learning 20 (3): 273–97. https://doi.org/10.1007/BF00994018.

e support-vector network is an ML algorithm used in classication

problems with two groups. e authors generalize the support-vector net-

works for separable training data and extend the algorithm to nonseparable

training data.

D’Acunto, Francesco, Nagpurnanand Prabhala, and Alberto Rossi. 2017. “e

Promises and Pitfalls of Robo-Advising.” https://ssrn.com/abstract=3122577.

e authors investigate the use of robo-advising in portfolio manage-

ment. Investors who use robo-advising tend to have more assets, trade

more often, and have better risk-adjusted performance. e results show

that among investors who adopt robo-advising, diversication and stock

volatility decrease when the number of stocks held is less than ve. e

same is not true, however, when the number of stocks initially held is more

than 10. us, the study illustrates the possible heterogeneous impacts of

robo-advising on investors’ portfolio performance.

Das, Sanjiv Ranjan. 2014. “Text and Context: Language Analytics in

Finance.” Foundations and Trends in Finance 8 (3): 145–261. https://doi.

org/10.1561/0500000045.

In this monograph, the author provides a comprehensive survey of tech-

niques and applications of text analytics in the eld of nance. Specically,

he discusses text extraction steps using the statistical package R, how to

analyze the extracted text, the classication of texts and words, and the

application of text data in empirical studies.

DeMiguel, Victor, Lorenzo Garlappi, and Raman Uppal. 2009. “Optimal

versus Naive Diversication: How Inecient Is the 1/N Portfolio Strategy?”

Review of Financial Studies 22 (5): 1915–53. https://doi.org/10.1093/rfs/

hhm075.

e authors conduct an empirical exercise to evaluate the out-of-sample

performance of optimal asset allocation models. In terms of Sharpe ratio,

certainty equivalent return, and turnover, none of these models fared better

than the naive 1/N portfolio. e authors attribute this nding to model

Articial Intelligence in Asset Management

estimation error that more than osets the gains from optimal portfolio

diversication.

Dhar, Vasant, and Roger Stein. 2017. “FinTech Platforms and Strategy.” MIT

Sloan Research Paper No. 5183-16. https://ssrn.com/abstract=2892098.

Briey, ntech is described as nancial sector innovations that have revo-

lutionized most aspects of nancial services. e article’s authors discuss

the current status of ntech in the internet era, future directions of change,

and possible strategies to pursue. e ntech market has yet to achieve its

highest potential, because developing countries continue to enter the digi-

tal era and new platforms for nancial services are still developing.

Dixon, Matthew, Diego Klabjan, and Jin Hoon Bang. 2017. “Classication-

Based Financial Markets Prediction Using Deep Neural Networks.”

Algorithmic Finance 6 (3–4): 67–77.

DNNs have been successfully applied to speech transcription and image

detection. In this study, the authors describe the application of DNNs to

the classication of direction of movement in nancial data.

Dixon, Matthew, and Nicholas G. Polson. 2019. “Deep Fundamental Factor

Models.” https://arxiv.org/abs/1903.07677.

e authors extend fundamental factor models using neural networks to

allow for nonlinearity, interaction eects, and nonparametric shocks. e

proposed technique can be used under heteroskedastic errors, provides

interpretability, and ranks the importance of factors (e.g., current enterprise

value, price-to-book ratio, price-to-sales ratio, price-to-earnings ratio, log

market cap) and interaction eects. e neural network factor model pre-

dicts S&P 500 stock returns with smaller out-of-sample mean-squared

errors than generalized linear regressions do. In addition, neural network

factor models are found to produce information ratios that are three times

higher than those seen with linear models.

Donaldson, R. Glen, and Mark Kamstra. 1997. “An Articial Neural

Network-GARCH Model for International Stock Return Volatility.” Journal

of Empirical Finance 4 (1): 17–46. https://doi.org/10.1016/S0927-5398(96)

00011-4.

e authors examine the performance of stock return volatility forecast-

ing models using daily returns data from London, New York, Tokyo, and

Toronto. An ANN-GARCH model is found to generally outperform its

traditional competing models—GARCH, EGARCH, and Sign-GARCH

models—in both in-sample and out-of-sample periods.

References

Dunis, Christian L., Jason Laws, and Georgios Sermpinis. 2010. “Modelling

and Trading the EUR/USD Exchange Rate at the ECB Fixing.” European

Journal of Finance 16 (6): 541–60. https://doi.org/10.1080/13518470903037771.

e authors explore the application of neural networks to the exchange

rate forecasting problem. Unlike previous studies on neural networks that

select the network’s inputs on an ad hoc basis, the neural network tech-

niques in this study use autoregressive terms as inputs. e performance

of several such neural network techniques (including a higher-order neu-

ral network, a Psi Sigma network, and a neural network with multilayer

perceptron) on the daily EUR/USD exchange rate are compared with

the performance of traditional methods. In simple simulated exercises,

the multilayer perceptron technique signicantly outperforms the other

models considered. For sophisticated trading strategies, the higher-order

neural network technique achieves higher annualized returns than other

neural network techniques.

Eiben, Agoston E., and Jim E. Smith. 2015. Introduction to Evolutionary

Computing, 2nd ed. New York: Springer. https://doi.org/10.1007/978-3-662-

44874-8.

e authors of this textbook provide a comprehensive introduction to evo-

lutionary algorithms, from history to major design decisions to parameter

tuning and control. Advanced topics include multi-objective optimization,

the optimization of dynamic and noisy functions, constraint handling, and

hybridization with other techniques.

Elliott, Graham, and Allan Timmermann. 2008. “Economic Forecasting.”

Journal of Economic Literature 46 (1): 3–56. https://doi.org/10.1257/jel.46.1.3.

is study’s authors present economic forecasting problems in a unied

framework, including forecasters’ loss functions and types of economic

forecasting methods. e authors then explore the various forecasting

methodologies available to researchers, such as optimal point forecasts, a

classical approach to forecasting, a Bayesian approach to forecasting, den-

sity forecasts, and forecast combination techniques.

Fan, Alan, and Marimuthu Palaniswami. 2001. “Stock Selection Using

Support Vector Machines.” In International Joint Conference on Neural

Networks. Proceedings (Cat. No. 01CH37222), vol. 3, 1793–98. Washington,

DC: IEEE.

e authors choose stocks trading on the Australian Securities Exchange

using SVM. Compared with the 71% return for the benchmark portfolio,

Articial Intelligence in Asset Management

an equally weighted portfolio of stocks selected by SVM produced a 208%

return over a ve-year period.

Farmer, J. Doyne, Austin Gerig, Fabrizio Lillo, and Szabolcs Mike. 2006.

“Market Eciency and the Long-Memory of Supply and Demand: Is Price

Impact Variable and Permanent or Fixed and Temporary?” Quantitative

Finance 6 (2): 107–12. https://doi.org/10.1080/14697680600668048.

e long-memory of demand and supply means that future buying and

selling behavior, and therefore price movements, should be predictable. In

reality, however, price movements are essentially uncorrelated. e authors

aim to explain market eciency given these two apparently contradictory

facts. ey revisit earlier studies on the topic to demonstrate that market

eciency is maintained by liquidity imbalances rather than mean-reverting

price changes. To show this dynamic, they introduce transaction time into

the model and demonstrate that liquidity covaries in time with the long-

memory of supply and demand.

Feng, Guanhao, Stefano Giglio, and Dacheng Xiu. 2017. “Taming the Factor

Zoo: A Test of New Factors.” Fama-Miller Working Paper; Chicago Booth

Research Paper No. 17-04. https://ssrn.com/abstract=2934020.

e authors present a variant of the LASSO method to identify the most

relevant factors among many potential candidates to explain asset returns.

e results indicate that only a few recently introduced factors, including

protability, have signicant incremental explanatory power.

Feng, Guanhao, Nick Polson, and Jianeng Xu. 2020. “Deep Learning in

Characteristics-Sorted Factor Models.” https://ssrn.com/abstract=3243683.

Unlike the researchers of previous literature, who related rm character-

istics (inputs) to security returns (outputs), the authors of this paper intro-

duce an intermediate channel involving risk factors (intermediate features).

With the goal of reducing pricing errors, this bottom-up approach trains

a neural network that generates risk factors using rm characteristics to

explain security returns. e authors provide an alternative dimension

reduction framework on security sorting and factor generation.

Fernandes, Marcelo, Marcelo C. Medeiros, and Marcel Scharth. 2014.

“Modeling and Predicting the CBOE Market Volatility Index.” Journal of

Banking & Finance 40: 1–10. https://doi.org/10.1016/j.jbankn.2013.11.004.

e Chicago Board Options Exchange (CBOE) reports the volatil-

ity index (VIX) based on the 30-calendar day S&P 500 index option.

References

is implied volatility, a key indicator of overall market condition, is used

in many trading strategies. e authors study the time series properties of

the VIX series, showing that the VIX is negatively correlated with S&P

500 returns but has a positive relationship with S&P 500 volume. Further,

these two series do not appear to have a nonlinear relationship.

Financial Stability Board. 2017. “Articial Intelligence and Machine Learning

in Financial Services.” http://www.fsb.org/2017/11/articial-intelligence-and-

machine-learning-in-nancial-service.

is report by the Financial Stability Board examines the implications

arising from the application of AI and ML methods in nancial services.

It presents some background on the recent rise in the use of AI and ML in

various nance applications and services, discusses possible eects on the

nancial system, and assesses risks to its stability. e authors conclude

that in the absence of “audibility” and interpretability, AI and ML tech-

niques may pose macro-level risks and should therefore be monitored by

microprudential supervisors.

Fischer, omas, and Christopher Krauss. 2018. “Deep Learning with Long

Short-Term Memory Networks for Financial Market Predictions.” European

Journal of Operational Research 270 (2): 654–69. https://doi.org/10.1016/j.

ejor.2017.11.054.

In this study, the authors use long short-term memory (LSTM) networks

in predicting out-of-sample movements for S&P 500 stocks. LSTM is

a type of sequence learning method specically designed to learn long-

term dependencies. LSTM clearly outperforms memory-free classication

methods, such as deep neural nets and a logistic regression, in terms of

predictive accuracy. It also beats random forests, except during the global

nancial crisis.

Fisher, Ingrid E., Margaret R. Garnsey, and Mark E. Hughes. 2016. “Natural

Language Processing in Accounting, Auditing and Finance: A Synthesis of

the Literature with a Roadmap for Future Research.” Intelligent Systems in

Accounting, Finance & Management 23 (3): 157–214. https://doi.org/10.1002/

isaf.1386.

In this article, the authors synthesize the literature that applies NLP in

accounting, auditing, and nance. NLP is used to analyze textual data

(e.g., corporate nancial performance, management’s assessment of rm

performance, regulations), detect fraud, make inferences, and predict stock

prices. e review of large literature in the three elds reveals that of all

Articial Intelligence in Asset Management

the ML techniques applied in NLP, SVMs are the most popular tool,

followed by Naive Bayes, hierarchical clustering, statistical methods, and

term-frequency-inverse document frequency weighting.

Fletcher, Tristan, and John Shawe-Taylor. 2013. “Multiple Kernel Learning

with Fisher Kernels for High Frequency Currency Prediction.” Computational

Economics 42: 217–40. https://doi.org/10.1007/s10614-012-9317-z.

e authors construct kernels based on EUR/USD exchange rate data

from limited order book volumes, technical analysis, and market micro-

structure models. With these multiple kernels, SVMs are used to predict

the direction of price movements. SVMs based on multiple kernels signi-

cantly outperform individual SVMs and achieve 55% forecast accuracy, on

average.

Freyberger, Joachim, Andreas Neuhierl, and Michael Weber. 2018.

“Dissecting Characteristics Nonparametrically.” University of Chicago,

Becker Friedman Institute for Economics Working Paper No. 2018-50.

https://ssrn.com/abstract=3223630.

e authors propose a nonparametric method for choosing rm charac-

teristics that have incremental predictive power for the cross-section of

stock returns. Using an adaptive LASSO nonparametric method for model

selection and stock return forecast, they nd that of 62 rm characteristics,

only 9–16 have incremental predictive power. Compared with linear model

selection methods, LASSO performs relatively well in both model selec-

tion and expected return predictors.

Geva, Tomer, and Jacob Zahavi. 2014. “Empirical Evaluation of an

Automated Intraday Stock Recommendation System Incorporating Both

Market Data and Textual News.” Decision Support Systems 57: 212–23. https://

doi.org/10.1016/j.dss.2013.09.013.

e authors examine the eectiveness of incorporating textual news data

for stock return forecasting. To this end, various textual data types are

added sequentially to numerical market data to study the incremental gain

in forecasting performance. Overall, integrating textual data with market

data and using neural networks signicantly improve forecast accuracy.

Giamouridis, Daniel. 2017. “Systematic Investment Strategies.” Financial

Analysts Journal 73 (4): 10–14. https://doi.org/10.2469/faj.v73.n4.10.

e author oers some thoughts on the latest research on systematic invest-

ment from both academic and practical perspectives, also highlighting

References

areas where further research can improve the practice more considerably.

e article’s topics include data science and ML, factor investing, market

timing, and coordinated investing and holding.

Giamouridis, Daniel, and Sandra Paterlini. 2010. “Regular(ized) Hedge

Fund Clones.” Journal of Financial Research 33 (3): 223–47. https://doi.org/

10.1111/j.1475-6803.2010.01269.x.

e authors show the attractiveness of using constrained portfolio optimi-

zation methods for constructing ecient hedge fund clones. e ndings

indicate that the proposed hedge fund portfolio is competitive relative to

hedge fund clones obtained by commercially available proprietary con-

struction techniques as well as by benchmark hedge fund indices. e

replicated hedge fund portfolio exhibits a very high correlation with the

benchmark indices.

Gogas, Periklis, eophilos Papadimitriou, Maria Matthaiou, and Efthymia

Chrysanthidou. 2015. “Yield Curve and Recession Forecasting in a Machine

Learning Framework.” Computational Economics 45: 635–45. https://doi.

org/10.1007/s10614-014-9432-0.

e authors demonstrate the superior forecasting performance of SVMs

in forecasting economic recessions in the United States. Compared with

logit and probit models, the SVM technique improves forecast accuracy by

67%–100%.

Gradojevic, Nikola, and Jing Yang. 2006. “Non-Linear, Non-Parametric,

Non-Fundamental Exchange Rate Forecasting.” Journal of Forecasting 25 (4):

227–45. https://doi.org/10.1002/for.986.

e authors examine the exchange rate forecasting performance of ANNs.

Compared with naive random walk models and linear time series models,

ANNs perform well in forecasting the CAD/USD exchange rate based on

the root mean squared error criterion.

Groth, Sven S., and Jan Muntermann. 2011. “An Intraday Market Risk

Management Approach Based on Textual Analysis.” Decision Support Systems

50 (4): 680–91. https://doi.org/10.1016/j.dss.2010.08.019.

e authors study the potential gains from using newly obtained qualita-

tive text data to explain intraday market risk. Using Naive Bayes, K-nearest

neighbor, neural network, and SVMs for textual analysis, the authors

conclude that these techniques are helpful in discovering intraday market

exposure, with the latter model (SVM) being the best in terms of compu-

tational eciency and classication accuracy.

Articial Intelligence in Asset Management

Gu, Shihao, Bryan T. Kelly, and Dacheng Xiu. 2020. “Empirical Asset

Pricing via Machine Learning.” Review of Financial Studies 33 (5): 2223–73.

e authors show that ML techniques (trees and neural nets) may outper-

form traditional cross-section and time-series models in predicting asset

risk premia. e gain in predictive performance comes from the ability of

ML methods to allow nonlinear predictive interactions that may be infea-

sible with traditional statistical methods.

Gu, Shihao, Bryan T. Kelly, and Dacheng Xiu. 2019. “Autoencoder Asset

Pricing Models.” Yale ICF Working Paper No. 2019-04; Chicago Booth

Research Paper No. 19-24. https://ssrn.com/abstract=3335536.

e study’s authors propose an autoencoder neural network latent factor

modeling for asset pricing. Although other studies incorporate a linear-

ity assumption, the authors’ approach allows for nonlinear relationships

between factor exposures and asset characteristics. Incorporating the no-

arbitrage condition into the ML framework, the new latent factor asset

pricing model produces far smaller out-of-sample errors relative to other

leading asset pricing modeling techniques, such as Fama–French, principal

component analysis, and linear conditioning methods.

Hagenau, Michael, Michael Liebmann, and Dirk Neumann. 2013.

“Automated News Reading: Stock Price Prediction Based on Financial News

Using Context-Capturing Features.” Decision Support Systems 55 (3): 685–97.

https://doi.org/10.1016/j.dss.2013.02.006.

In this article, the authors describe the application of textual nancial

information to stock price prediction. Unlike other, similar approaches, the

one in this study uses a feature selection that allows market feedback. By

selecting the feedback-based relevant features, the approach reduces over-

tting problems and signicantly improves classication accuracy. As a

result, stock return prediction accuracy increased by as much as 76%.

Hamid, Shaikh A., and Zahid Iqbal. 2004. “Using Neural Networks for

Forecasting Volatility of S&P 500 Index Futures Prices.” Journal of Business

Research 57 (10): 1116–25. https://doi.org/10.1016/S0148-2963(03)00043-2.

e authors of this study examine the performance of the popular neural

networks technique in forecasting the volatility of S&P 500 Index futures

prices. e results show that volatility forecasts from neural networks out-

perform implied volatility forecasts obtained from the Barone–Adesi and

Whaley American futures options pricing model.

References

Han, Shuo, and Rung-Ching Chen. 2007. “Using SVM with Financial

Statement Analysis for Prediction of Stocks.” Communications of the IIMA 7 (4):

article 8.

e authors propose applying SVMs to nancial statement analysis for

stock market prediction. Using nancial indices improves forecast accuracy

and is more reliable than using technical indices.

Haykin, Simon. 2009. Neural Networks and Learning Machines, 3rd ed.

New York: Pearson.

In this comprehensive textbook, the author presents an up-to-date treat-

ment of neural networks. e book consists of six main parts that focus on

supervised learning, kernel methods based on radial-basis function net-

works, regularization methods, unsupervised learning, RL, and nonlinear

feedback systems.

Heaton, James B., Nick G. Polson, and Jan H. Witte. 2017. “Deep Learning

for Finance: Deep Portfolios.” Applied Stochastic Models in Business and

Industry 33 (1): 3–12. https://doi.org/10.1002/asmb.2209.

Traditional prediction and classication methods from nancial economics

might be impractical or dicult to apply to nancial problems, given the

large and complex nature of data. e authors describe DL hierarchical

models that can improve performance in nancial prediction problems and

classication.

Hendricks, Dieter, and Diane Wilcox. 2014. “A Reinforcement Learning

Extension to the Almgren–Chriss Framework for Optimal Trade Execution.”

In 2014 IEEE Conference on Computational Intelligence for Financial

Engineering & Economics (CIFEr), 457–64. London: IEEE.

e authors present an RL technique to optimize the liquidation volume tra-

jectory. e model is based on a standard Almgren–Chriss model, a popular

trade execution strategy that assumes risk aversion to trade execution. e

proposed technique is able to change the volume of trade using information

on market spread and volume dynamics. Empirical tests show promising

results. Specically, on average, the new technique can improve post-trade

implementation shortfall by up to 10.3% relative to the base model.

Hong, Taeho, and Ingoo Han. 2002. “Knowledge-Based Data Mining

of News Information on the Internet Using Cognitive Maps and Neural

Networks.” Expert Systems with Applications 23 (1): 1–8. https://doi.

org/10.1016/S0957-4174(02)00022-2.

Articial Intelligence in Asset Management

e authors of this study create the Knowledge-Based News Miner

(KBNMiner), which integrates cognitive maps with neural networks.

Unlike the time-series models used in interest rate forecasting, the tech-

nique uses prior news on interest rates in predicting interest rates. Relative

to neural network and random walk models, KBNMiner leads to improved

interest rate predictions.

Hu, Yong, Kang Liu, Xiangzhou Zhang, Lijun Su, E.W.T. Ngai, and Mei

Liu. 2015. “Application of Evolutionary Computation for Rule Discovery in

Stock Algorithmic Trading: A Literature Review.” Applied Soft Computing 36:

534–51. https://doi.org/10.1016/j.asoc.2015.07.008.

e authors explore the literature on the application of evolutionary com-

putation in stock algorithmic trading. ey observe that most of the trad-

ing techniques considered in the surveyed studies do well in the downtrend

but poorly in the uptrend, which is likely because of the problems associ-

ated with the selection of factors and the transaction costs.

Huang, Chien-Feng. 2012. “A Hybrid Stock Selection Model Using Genetic

Algorithms and Support Vector Regression.” Applied Soft Computing 12 (2):

807–18. https://doi.org/10.1016/j.asoc.2011.10.009.

Stock selection problems are particularly challenging in the area of invest-

ment research. is author examines a hybrid approach for stock selection

using SVR and genetic algorithms. In the rst stage, SVR chooses the top-

ranked stocks. In the second stage, genetic algorithms are used to optimize

the portfolio parameters and feature selection. e results indicate that this

hybrid approach signicantly outperforms the benchmark model.

Huang, Wei, Yoshiteru Nakamori, and Shou-Yang Wang. 2005. “Forecasting

Stock Market Movement Direction with Support Vector Machine.”

Computers & Operations Research 32 (10): 2513–22. https://doi.org/10.1016/j.

asoc.2011.10.009.

e authors examine the performance of SVMs, a popular learning algo-

rithm, in predicting the direction of the Nikkei 225 Index. Relative to

forecasts obtained via linear discriminant analysis, quadratic discriminant

analysis, and Elman backpropagation neural networks, forecasts reached

via SVM are more accurate in predicting the direction of the stock index.

e forecasting performance can be further improved by combining SVM

with other classication techniques.

Huang, Zan, Hsinchun Chen, Chia-Jung Hsu, Wun-Hwa Chen, and

Soushan Wu. 2004. “Credit Rating Analysis with Support Vector Machines

References

and Neural Networks: A Market Comparative Study.” Decision Support

Systems 37 (4): 543–58. https://doi.org/10.1016/S0167-9236(03)00086-1.

e authors of this study apply SVMs to corporate credit rating analysis.

Both the SVM and benchmark neural network techniques yield approxi-

mately 80% prediction accuracy for US and Taiwan market data. In addi-

tion, the authors implement input nancial variable contribution analysis

to help interpret the neural network results.

Hutchinson, James M., Andrew W. Lo, and Tomaso Poggio. 1994. “A

Nonparametric Approach to Pricing and Hedging Derivative Securities

via Learning Networks.” Journal of Finance 49 (3): 851–89. https://doi.

org/10.1111/j.1540-6261.1994.tb00081.x.

In this article, the authors present a nonparametric pricing approach for

estimating pricing and hedging derivative securities. Unlike a parametric

approach for estimating the arbitrage-based pricing formula, a nonpara-

metric approach does not require that the dynamics of asset prices be

known in advance. Using data on pricing and delta hedging of S&P 500

futures options, the authors nd that in the majority of cases, a network-

pricing formula outperforms the naive traditional Black–Scholes model.

In addition, compared with other popular models—such as ordinary least

squares, radial basis function networks, multilayer perceptron network,

and projection pursuit—the neural network pricing formula is computa-

tionally more ecient and accurate when the price dynamics of the under-

lying asset are unknown.

James, Gareth, Daniela Witten, Trevor Hastie, and Robert Tibshirani.

2017. An Introduction to Statistical Learning with Applications in R. New York:

Springer.

In this book, the authors provide a fundamental understanding of statisti-

cal learning, including LASSO, classication and regression trees, boost-

ing, and SVMs. Unlike other textbooks in this new area, this one treats

these topics in a less technical manner and focuses on their application

using statistical software R.

Kaashoek, Johan F., and Herman K. van Dijk. 2002. “Neural Network

Pruning Applied to Real Exchange Rate Analysis.” Journal of Forecasting

21 (8): 559–77. https://doi.org/10.1002/for.835.

e authors apply neural network pruning to exchange rate forecasting.

Because neural networks consist of many cells, the number of cells is pruned

or reduced using basic descriptive procedures, such as multiple correlation

Articial Intelligence in Asset Management

coecients, principal component analysis of residuals, and graphical analy-

sis. Compared with standard autoregressive integrated moving average mod-

els, the proposed method performs better in terms of its predictive accuracy

and its ability to capture the long-term dynamics of exchange rates.

Katona, Zsolt, Marcus Painter, Panos N. Patatoukas, and Jean Zeng. 2018.

“On the Capital Market Consequences of Alternative Data: Evidence from

Outer Space.” 9th Miami Behavioral Finance Conference 2018. https://ssrn.

com/abstract=3222741.

e authors examine the use of satellite images of parking lot trac to

predict the earnings of major US retailers before public disclosure.

Ke, Zheng Tracy, Bryan T. Kelly, and Dacheng Xiu. 2019. “Predicting

Returns with Text Data.” https://ssrn.com/abstract=3389884.

e authors propose a text mining methodology that extracts sentiment

information from textual sources. is approach is simple, requires mini-

mal computing power, and can be adapted to the dataset being used. A

simple trading strategy that buys assets with positive recent news senti-

ment and sells assets with negative sentiment using this method generates

higher out-of-sample abnormal returns than similar strategies based on

sentiment scores from commercial vendors such as RavenPack.

Kearney, Colm, and Sha Liu. 2014. “Textual Sentiment in Finance: A Survey

of Methods and Models.” International Review of Financial Analysis 33:

171–85. https://doi.org/10.1016/j.irfa.2014.02.006.

is article’s authors provide a comprehensive review of the text-based sen-

timent literature and qualitative information sources used in prior research,

as well as the most frequently used textual data analysis methods and their

empirical applications. Further, they identify possible opportunities and

directions for future research using text-based qualitative data.

Kearns, Michael, and Yuriy Nevmyvaka. 2013. “Machine Learning for Market

Microstructure and High Frequency Trading.” In High Frequency Trading:

New Realities for Traders, Markets, and Regulators, edited by Maureen O’Hara,

Marcos Lopez de Prado, and David Easley, 91–124. London: Risk Books.

e authors explore the application of ML approaches to high-frequency

trading and microstructure data. ey consider a few case studies used in

solving dierent trading problems. Specically, they review the applica-

tion of ML methods in optimal trade execution problems and in predicting

equity limit order book price movements.

References

Kercheval, Alec N., and Yuan Zhang. 2015. “Modelling High-Frequency

Limit Order Book Dynamics with Support Vector Machines.” Quantitative

Finance 15 (8): 1315–29. https://doi.org/10.1080/14697688.2015.1032546.

e authors develop an SVM-based learning model to examine the

dynamics of information contained in a limit order book. A limit order

book records high-frequency trading activity grouped by asks and bids,

including time and type of transaction, order price, and volume. e

authors show the usefulness of features selected by the proposed method

for forecasting short-term price dynamics.

Kim, Steven H., and Hyun Ju Noh. 1997. “Predictability of Interest Rates

Using Data Mining Tools: A Comparative Analysis of Korea and the US.”

Expert Systems with Applications 13 (2): 85–95. https://doi.org/10.1016/

S0957-4174(97)00010-9.

Forecasting techniques used in predicting interest rates have generally

failed to improve over the simple random walk model. In this article, the

authors consider neural networks, case-based reasoning, and the combina-

tion of the two to forecast interest rates in the United States and Korea.

Case-based reasoning uses accumulated past experiences in making deci-

sions. Interestingly, these techniques are superior to the random walk

model for the US market but cannot outperform the random walk model

for the Korean market.

Kirilenko, Andrei, Albert S. Kyle, Mehrdad Samadi, and Tugkan Tuzun.

2017. “e Flash Crash: High-Frequency Trading in an Electronic Market.”

Journal of Finance 72 (3): 967–98. https://doi.org/10.1111/jo.12498.

e authors investigate intraday market intermediation in an electronic

market around the time of large and temporary selling pressure that

occurred on 6 May 2010. Known as the “ash crash,” decline in prices and

increase in trading volume was triggered by a trader who initiated a large

sell program to sell E-mini S&P 500 stock index futures. is kind of

large selling pressure may potentially lead to a market crash. Using empiri-

cal data from 6 May 2010 and the three days leading up to that date, the

study nds evidence that most intraday intermediaries did not change their

trading behavior despite large and temporary selling pressure.

Kirilenko, Andrei A., and Andrew W. Lo. 2013. “Moore’s Law versus

Murphy’s Law: Algorithmic Trading and Its Discontents.” Journal of Economic

Perspectives 27 (2): 51–72. https://doi.org/10.1257/jep.27.2.51.

Articial Intelligence in Asset Management

In this article, the authors survey algorithmic trading and its history, its

major drivers, challenges, and the resulting unintended consequences for

the nancial system. ey argue that although technological advances have

reduced nancial transaction costs drastically, current nancial regulations

need to be aligned with the requirements of the digital age.

Kofman, Paul, and Ian G. Sharpe. 2003. “Using Multiple Imputation in

the Analysis of Incomplete Observations in Finance.” Journal of Financial

Econometrics 1 (2): 216–49. https://doi.org/10.1093/jjnec/nbg013.

e authors examine the application of multiple imputation methods.

When applied to two nancial datasets involving severe data incomplete-

ness, the imputation methods outperform the ad hoc approaches com-

monly used in the nance literature.

Kolm, Petter N., and Gordon Ritter. Forthcoming. “Modern Perspectives on

Reinforcement Learning in Finance.” Journal of Machine Learning in Finance

1 (1). https://ssrn.com/abstract=3449401.

e authors provide an overview of RL applications in nance. RL allows

for solving dynamic optimization problems such as the pricing and hedg-

ing of contingent claims, investment and portfolio allocation, buying and

selling a portfolio of securities subject to transaction costs, market mak-

ing, asset–liability management, and optimization of tax consequences in

a model-free way. e authors also highlight some of the challenges with

using RL.

Kolm, Petter N., Reha Tütüncü, and Frank J. Fabozzi. 2014. “60 Years of

Portfolio Optimization: Practical Challenges and Current Trends.” European

Journal of Operational Research 234 (2): 356–71. https://doi.org/10.1016/j.

ejor.2013.10.060.

e authors review approaches that aim at addressing some of the practical

challenges of portfolio optimization, including accounting for transaction

costs, portfolio management constraints, and sensitivity to estimates of

expected returns and covariances. ey also illustrate several developments

in portfolio optimization, including diversication methods, risk-parity

portfolios, the mixing of dierent sources of alpha, and practical multi-

period portfolio optimization.

Kumar, P. Ravi, and Vadlamani Ravi. 2007. “Bankruptcy Prediction in Banks

and Firms via Statistical and Intelligent Techniques – A Review.” European

Journal of Operational Research 180 (1): 1–28. https://doi.org/10.1016/j.

ejor.2006.08.043.

References

In this comprehensive review, the authors consider the application of sta-

tistical and intelligent techniques for bankruptcy prediction between 1985

and 2005. Historically, researchers have used many dierent techniques to

forecast bank and rm bankruptcy, including traditional statistical meth-

ods, neural networks, case-based reasoning, decision trees, evolutionary

approaches, rough set–based techniques, fuzzy logic, SVMs, and various

hybrid models.

Lam, Monica. 2004. “Neural Network Techniques for Financial Performance

Prediction: Integrating Fundamental and Technical Analysis.” Decision Support

Systems 37 (4): 567–81. https://doi.org/10.1016/S0167-9236(03)00088-5.

e author examines the eect of integrating technical and fundamental

analysis to forecast the rate of return on common shareholders’ equity.

Empirical results show that a neural network incorporating nancial state-

ment and macroeconomic variables performs signicantly better than

the market return. It fails, however, to outperform the maximum bench-

mark—the average return of the top one-third of companies with the

highest returns.

Leshik, Edward, and Jane Cralle. 2011. An Introduction to Algorithmic Trading:

Basic to Advanced Strategies. Chichester, UK: John Wiley and Sons.

e authors of this book provide a background on algorithmic trading and

describe current algorithmic trading strategies. e book contains two

broadly dened parts. e rst part is an introduction to trading algo-

rithms, current popular algorithms, and how to use and optimize these

trading algorithms. e second part provides readers with tools and exam-

ples to help them learn the trading strategies the authors have developed.

Leung, Henry, and ai Ton. 2015. “e Impact of Internet Stock Message

Boards on Cross-Sectional Returns of Small-Capitalization Stocks.” Journal of

Banking & Finance 55: 37–55. https://doi.org/10.1016/j.jbankn.2015.01.009.

In this article, the authors study the stock return eects of stock messages

posted on the Australian online stock message board HotCopper of the

Australian Securities Exchange. e empirical evidence suggests that both

the number of messages and their sentiment have a positive impact on the

current return of underperforming small stocks and their trading volume.

Large stocks, however, do not appear to be inuenced by online message

sentiment.

Li, Xiaodong, Xiaodi Huang, Xiaotie Deng, and Shanfeng Zhu. 2014.

“Enhancing Quantitative Intra-Day Stock Return Prediction by Integrating

Articial Intelligence in Asset Management

Both Market News and Stock Prices Information.” Neurocomputing 142:

228–38. https://doi.org/10.1016/j.neucom.2014.04.043.

Previous studies on stock prices and market news consider either market

news or past stock prices a predictor of future stock returns. In this article,

the authors integrate these two sources of information using a kernel learn-

ing technique. e empirical evidence suggests that the proposed method

outperforms three baseline models (market news, past stock prices, and

both market news and past stock prices).

Liao, Shu-Hsien, and Shan-Yuan Chou. 2013. “Data Mining Investigation

of Co-Movements on the Taiwan and China Stock Markets for Future

Investment Portfolio.” Expert Systems with Applications 40 (5): 1542–54.

https://doi.org/10.1016/j.eswa.2012.08.075.

Taiwan and China signed an Economic Cooperation Framework

Agreement in 2010. In this article, the authors investigate the co-move-

ments in the two countries’ stock markets after this agreement. To do so,

they identify 30 categories of stock indices and analyze their behavior using

patterns, rules, and cluster analysis. e empirical results show that for the

Taiwan stock market, electronics, nancial and insurance, and semicon-

ductor stock indices strongly move together with the TAIEX Index. For

the Hong Kong stock market, real estate, telecommunications, and nan-

cial services stock indices move with the HSI Index, and for the Shenzhen

stock market, manufacturing, machinery, and electronics stock indices

move with the SZSE Index. e authors also discuss the co-movement

across these three stock markets.

Lin, Chin-Shien, Haider A. Khan, Ruei-Yuan Chang, and Ying-Chieh

Wang. 2008. “A New Approach to Modeling Early Warning Systems for

Currency Crises: Can a Machine-Learning Fuzzy Expert System Predict

Currency Crises Eectively?” Journal of International Money and Finance

27 (7): 1098–121. https://doi.org/10.1016/j.jimonn.2008.05.006.

In this study, the authors develop a hybrid causal model to predict currency

crises. is hybrid model, constructed by integrating neural networks with

a fuzzy logic model, can predict currency crises better than other popular

methods, such as neural networks and logistic regression.

Lopez de Prado, Marcos. 2016. “Building Diversied Portfolios that

Outperform Out of Sample.” Journal of Portfolio Management 42 (4): 59–69.

https://doi.org/10.3905/jpm.2016.42.4.059.

References

Markowitz’s critical line algorithm (CLA) is based on quadratic optimiza-

tion. e CLA procedure is unstable mainly because the algorithm requires

the inversion of a covariance matrix. e authors introduce a hierarchical

risk parity (HRP) method, an alternative algorithm that applies graph the-

ory and ML, which uses the information in the covariance matrix without

requiring a matrix inversion. By replacing the covariance matrix with a tree

structure, the HRP method allows for solution of the portfolio optimiza-

tion problem even when the covariance matrix is singular. Monte Carlo

experiments show that HRP produces lower out-of-sample variance than

does the CLA procedure, whereas CLA delivers minimum-variance port-

folios only in sample.

Loterman, Gert, Iain Brown, David Martens, Christophe Mues, and Bart

Baesens. 2012. “Benchmarking Regression Algorithms for Loss Given

Default Modeling.” International Journal of Forecasting 28 (1): 161–70. https://

doi.org/10.1016/j.ijforecast.2011.01.006.

e existing literature on credit risk models focuses mainly on the prob-

ability of default, while loss given default (LGD) remains an inadequately

studied area. e authors of this study consider numerous regression meth-

ods to model the LGD problem. Using data on losses from several major

international banks, the considered models can explain 4% to 43% of the

variation in LGD. Out of 24 models in the study, SVMs and neural net-

works perform signicantly better than traditional statistical techniques.

Lowe, David. 1994. “Novel Exploitation of Neural Network Methods in

Financial Markets.” In Proceedings of 1994 IEEE International Conference on

Neural Networks, vol. 6, 3623–28. Orlando, FL: IEEE.

e author introduces feedforward networks for portfolio management,

demonstrating that neural network methods have the advantage of being

able to approximate nonlinearities in the data and to optimize the portfolio

under constraints.

Majhi, Ritanjali, Ganapati Panda, and Gadhadhar Sahoo. 2009. “Ecient

Prediction of Exchange Rates with Low Complexity Articial Neural

Network Models.” Expert Systems with Applications 36 (1): 181–89. https://doi.

org/10.1016/j.eswa.2007.09.005.

In this article, the authors propose two ANN models—functional link

ANN (FLANN) and cascaded functional link ANN (CFLANN)—to

forecast currency exchange rates. ey nd that compared with Widrow’s

popular least-mean-square algorithm, both the CFLANN and FLANN

Articial Intelligence in Asset Management

models are superior in exchange rate forecasting, with the CFLANN

model performing the best among the three models considered.

Manahov, Viktor, Robert Hudson, and Bartosz Gebka. 2014. “Does High

Frequency Trading Aect Technical Analysis and Market Eciency? And If

So, How?” Journal of International Financial Markets, Institutions and Money

28: 131–57. https://doi.org/10.1016/j.intn.2013.11.002.

Approximately 40% of foreign exchange traders use technical analy-

sis for their trading rules, and others may implicitly rely on the ecient

market hypothesis. In this article, the authors examine the relationship

between technical analysis and high-frequency trading in foreign exchange

markets. ey develop a special adaptive form of strongly typed genetic

programming to forecast most frequently traded currency pairs using high-

frequency data. e results show that strongly typed genetic programming

signicantly outperforms traditional econometric forecasting models.

Importantly, excess returns are found to be both statistically and economi-

cally signicant, even when transaction costs are considered.

Manela, Asaf, and Alan Moreira. 2017. “News Implied Volatility and

Disaster Concerns.” Journal of Financial Economics 123 (1): 137–62. https://

doi.org/10.1016/j.jneco.2016.01.032.

In this article, the authors propose a text-based measure of uncertainty—

a news implied volatility (NVIX) measure—to study the relationship

between uncertainty and expected returns. Using front-page articles from

the Wall Street Journal starting in 1890, the constructed NVIX measure

captures the time variation in risk premia. Specically, a high NVIX mea-

sure indicates high future returns in normal times and rises just before

rare disaster events (e.g., wars, concerns about government policy, natural

disasters).

Manning, Christopher D., and Hinrich Schütze. 1999. Foundations of

Statistical Natural Language Processing. Cambridge, MA: MIT Press.

Statistical NLP applies probabilistic and information theory as well as

linear algebra to characterize linguistic observations. is graduate-level

textbook consists of four broad parts. Part 1 presents the essential math-

ematical and linguistic concepts. Part 2 focuses on word-based statistics

and inference. Part 3 is devoted to studying grammar-based statistical

techniques. Finally, Part 4 covers the applications of statistical techniques

to NLP.

References

Markowitz, Harry. 1952. “Portfolio Selection.” Journal of Finance 7 (1): 77–91.

In this seminal theory article, Markowitz describes the portfolio selection

problem using expected returns and their variances. Also known as the

mean–variance model, it chooses an ecient portfolio with the maximum

expected return for a given level of risk.

McCarthy, John, Marvin L. Minsky, Nathaniel Rochester, and Claude E.

Shannon. 2006. “A Proposal for the Dartmouth Summer Research Project on

Articial Intelligence, August 31, 1955.” AI Magazine 27 (4): 12–14.

is article involves a republication of the seminal 1955 Dartmouth

Summer Research Proposal on AI by four mathematicians. e authors

of this research proposal are believed to have been the rst to use the term

“articial intelligence.” e authors suggested conducting a 10-man study

on AI during the summer of 1956.

Michaud, Richard O., and Robert O. Michaud. 2008. Ecient Asset

Management: A Practical Guide to Stock Portfolio Optimization and Asset

Allocation, 2nd ed. Oxford, UK: Oxford University Press.

e authors discuss mean–variance portfolio optimization and its limita-

tions, then review alternative portfolio management approaches from a

statistical point of view. e topics covered include Markowitz eciency,

classic mean–variance optimization, traditional criticisms and alternatives,

unbounded mean–variance portfolio eciency, mean–variance eciency

with linear constraints, the resampled eciency frontier and its properties,

portfolio rebalancing and monitoring, input estimation, Bayes estimation

and caveats, and avoiding optimization errors.

Mitkov, Ruslan, ed. 2014. e Oxford Handbook of Computational Linguistics,

2nd ed. New York: Oxford University Press. https://doi.org/10.1093/oxfor

dhb/9780199573691.001.0001.

is book is devoted to describing major concepts, methods, and appli-

cations in computational linguistics. It provides an overview of the eld,

describing a broad range of current techniques used in NLP and providing

a comprehensive survey of current applications of NLP.

Molnar, Christoph. 2020. Interpretable Machine Learning: A Guide for

Making Black Box Models Explainable. https://christophm.github.io/

interpretable-ml-book/.

e author presents a free and comprehensive book on various ways of

facilitating the interpretation and understanding of ML predictions.

Articial Intelligence in Asset Management

Murphy, Kevin P. 2012. Machine Learning: A Probabilistic Perspective.

Cambridge, MA: MIT Press.

e author of this advanced textbook on ML focuses on probability theory

and distributions. e rst part of the book is devoted to ML concepts and

methods, probability theory and distributions, and Bayesian and frequen-

tist statistics. e second part presents linear and logistic regression and

generalized linear, mixture, and latent linear models. e third part dis-

cusses kernels, adaptive models (classication and regression trees, boost-

ing, neural networks, ensemble learning), Markov and hidden Markov

models, state space models, graphical models, variational inference, Monte

Carlo inference, and DL.

Nevmyvaka, Yuriy, Yi Feng, and Michael Kearns. 2006. “Reinforcement

Learning for Optimized Trade Execution.” In Proceedings of the 23rd

International Conference on Machine Learning, 673–80.

Optimized trade execution is an important problem in the eld of nance.

In this study, the authors apply RL to optimal trade execution using

NASDAQ market data. When market state variables are chosen carefully,

RL can improve trade optimization relative to other baseline execution

strategies.

Nuij, Wijnand, Viorel Milea, Frederik Hogenboom, Flavius Frasincar, and

Uzay Kaymak. 2014. “An Automated Framework for Incorporating News

into Stock Trading Strategies.” IEEE Transactions on Knowledge and Data

Engineering 26 (4): 823–35. https://doi.org/10.1109/TKDE.2013.133.

e authors introduce a framework that automatically incorporates news

into stock trading strategies. Using genetic programming to nd optimal

trading strategies, the authors achieve results that indicate that optimal

trading strategies include technical trading rules and, in many cases, news

variables as additional input.

Nuti, Giuseppe, Mahnoosh Mirghaemi, Philip Treleaven, and Chaiyakorn

Yingsaeree. 2011. “Algorithmic Trading.” Computer 44 (11): 61–69. https://

doi.org/10.1109/MC.2011.31.

Trading in the nancial sector uses automated systems that are fast and

complex. In this article, the authors provide an overview of trading algo-

rithms and how such systems work. In particular, the authors explain the

trading objective, trading process, electronic trading execution, and trad-

ing analysis, and they provide some examples.

References

Oh, Kyong Jo, and Ingoo Han. 2000. “Using Change-Point Detection

to Support Articial Neural Networks for Interest Rates Forecasting.”

Expert Systems with Applications 19 (2): 105–15. https://doi.org/10.1016/

S0957-4174(00)00025-7.

Interest rates change according to the monetary policy of governments,

and the authors of this study propose to identify intervals between these

change points and use this information in predicting interest rates. ey

use a backpropagation neural network (BPN) to detect the change-point

groups of interest rates and apply the BPN technique again to interest rate

forecasting. e proposed BPN technique with change-point detection

outperforms the simple BPN technique at a statistically signicant level.

Papaioannou, Georgios V., and Daniel Giamouridis. Forthcoming. Enhancing

Alpha Signals from Trade Ideas Data Using Supervised Learning, in Machine

Learning and Asset Management. Springer.

In this chapter, the researchers use trade investment ideas along with

supervised ML. Trade ideas are market experts’ recommendations that

institutional investors often use. Investment trade ideas are classied into

two classes (success or failure) using supervised ML methods, specically

random forests and gradient boosting trees. In addition to stock character-

istics, the authors use characteristics of the contributor to the investment

trade idea. e overall results demonstrate a performance improvement of

more than 1% for long ideas and of more than 2% for short ideas.

Park, Saerom, Jaewook Lee, and Youngdoo Son. 2016. “Predicting Market

Impact Costs Using Nonparametric Machine Learning Models.” PLoS One

11 (2): 1–13. https://doi.org/10.1371/journal.pone.0150243.

Transaction costs aect the prots of investment strategies. e authors

seek to more accurately predict market impact cost, which is the result of

the dierence between the initial stock price and the actual price after the

transaction. Using data on the US stock market, the authors apply non-

parametric ML techniques (neural networks, Bayesian neural network,

Gaussian process, SVR) to predict market impact cost. e empirical

results suggest that nonparametric ML models generally outperform their

parametric counterparts.

Patel, Keyur, and Marshall Lincoln. 2019. It’s Not Magic: Weighing the

Risks of AI in Financial Services. London: Centre for the Study of Financial

Innovation. http://www.cs.org/s/Magic_10-19_v12_Proof.pdf.

Articial Intelligence in Asset Management

e authors oer a detailed review of the potential benets and risks of

applying AI in the nancial services industry. e report is divided into

three sections. Section 1 introduces AI in the nancial services industry.

Section 2 discusses AI’s potential benets in nancial services, such as

improvements in security, compliance, and risk management. e report

devotes much attention to Section 3, which presents the potential risks of

applying AI and ML in nancial services. It identies 12 key risks, includ-

ing those related to ethical challenges.

Peña, Tonatiuh, Seran Martinez, and Bolanle Abudu. 2011. “Bankruptcy

Prediction: A Comparison of Some Statistical and Machine Learning

Techniques.” In Computational Methods in Economic Dynamics, Dynamic

Modeling and Econometrics in Economics and Finance, Vol. 13, edited by

Herbert Dawid and Willi Semmler, 109–31. London: Springer. https://doi.

org/10.1007/978-3-642-16943-4_6.

e authors examine the accuracy of statistical and ML methods in pre-

dicting bank failures. ey introduce Gaussian processes for classication

and evaluate the processes’ performance relative to other statistical and

ML techniques (logistic regression, discriminant analysis, least-squares

SVMs). e study nds that forecasts generated from dierent instances

of Gaussian process classiers can compete with the results from popular

techniques.

Rasekhschae, Keywan Christian, and Robert C. Jones. 2019. “Machine

Learning for Stock Selection.” Financial Analysts Journal 75 (3): 70–88.

https://doi.org/10.1080/0015198X.2019.1596678.

ML methods are gaining in popularity among nancial practitioners

because they can better capture dynamic relationships between predictors

and expected returns. Given the noisy historical nancial data, however,

the risk of overtting poses a real challenge. e authors discuss two main

ways of overcoming the overtting problem when using ML for predicting

the cross-section of stock returns. Combining dierent forecasts reduces

noise. e authors recommend forecast combination along dierent

dimensions, such as from dierent forecasting techniques, based on dier-

ent training sets, and for dierent horizons. Similarly, feature engineering

can help mitigate the overtting problem by increasing the signal-to-noise

ratio.

Rapach, David E., Jack K. Strauss, Jun Tu, and Guofu Zhou. 2019. “Industry

Return Predictability: A Machine Learning Approach.” Journal of Financial

Data Science 1 (3): 9–28.

References

e authors apply ML to predict industry returns. Specically, they use

LASSO regression to t sparse models that include lagged industry returns

for 30 industries. e LASSO-selected variables are then estimated using

an ordinary least-squares model to lessen the eect of downward bias in

estimated coecients from the LASSO model. In-sample and out-of-

sample predictions of industry returns provide evidence for the relevance of

information in lagged industry returns.

Rapach, David E., Jack K. Strauss, and Guofu Zhou. 2013. “International

Stock Return Predictability: What Is the Role of the United States?” Journal

of Finance 68 (4): 1633–62. https://doi.org/10.1111/jo.12041.

Stock return predictability has received signicant attention in the litera-

ture. In this study, the authors introduce a new powerful predictor of stock

returns in industrialized countries. ey nd that lagged US market returns

can dramatically improve stock return predictability in other industrial-

ized countries, whereas lagged non-US returns are not good predictors of

stock returns in the United States. e contribution of lagged US returns to

stock return predictability in non-US industrialized countries is explained

through a news-diusion model, in which shocks to US stock returns are

reected in equity prices in other industrialized countries with a lag.

Renault, omas. 2017. “Intraday Online Investor Sentiment and Return

Patterns in the U.S. Stock Market.” Journal of Banking & Finance 84: 25–40.

https://doi.org/10.1016/j.jbankn.2017.07.002.

Using investor opinions and ideas about stock market returns posted on

the Stocktwits blog, the authors construct investor sentiment data to study

its relationship with US stock returns. ey provide evidence that inves-

tor sentiment is an important variable for forecasting intraday stock index

returns.

Ribeiro, Bernardete, Catarina Silva, Ning Chen, Armando Vieira, and João

Carvalho das Neves. 2012. “Enhanced Default Risk Models with SVM+.”

Expert Systems with Applications 39 (11): 10140–52. https://doi.org/10.1016/j.

eswa.2012.02.142.

Recent advances in bankruptcy prediction consider adding additional

information, such as marketing reports, competitors landscape, economic

environment, customers screening, and industry trends. is additional

information can be incorporated into an SVM. Using data on French com-

panies, the authors demonstrate that their adaptation produces a better

bankruptcy prediction than does a baseline SVM.

Articial Intelligence in Asset Management

Ristolainen, Kim. 2018. “Predicting Banking Crises with Articial Neural

Networks: e Role of Nonlinearity and Heterogeneity.” Scandinavian

Journal of Economics 120 (1): 31–62. https://doi.org/10.1111/sjoe.12216.

Early warning systems help predict coming banking crises. Rather than

using traditional linear models, such as logistic regression, the author in

this study builds early warning systems using an ANN model. For regional

as well as international data, the proposed ANN model outperforms logis-

tic regression in predicting all banking crises two years in advance, given

the information about earlier crises.

Russell, Stuart, and Peter Norvig. 2010. Articial Intelligence: A Modern

Approach, 3rd ed. Upper Saddle River, NJ: Pearson.

e authors of this textbook provide an in-depth review of AI and cover

introductory concepts as well as recent advances in the eld. Topics include

intelligent agents, problem-solving agents, search algorithms, logic agents,

rst-order logic and inference, planning, uncertainty in knowledge and

reasoning, learning (e.g., learning from examples and learning probabilistic

models), NLP and communication, and robotics.

Sabharwal, Chaman L. 2018. “e Rise of Machine Learning and Robo-

Advisors in Banking.” Journal of Banking Technology 2: 28–43. https://www.

idrbt.ac.in/assets/publications/Journals/Volume_02/No_02/Chapter_02.pdf.

e author discusses the current use of ML and its future role in the nan-

cial sector. Robo-advisors used by the largest banks in the United States are

examples that the nancial sector has embraced ML for banking services.

Still, ML has yet to achieve its biggest impact in the nance industry.

Schumaker, Robert P., and Hsinchun Chen. 2006. “Textual Analysis of Stock

Market Prediction Using Financial News Articles.” AMCIS 2006 Proceedings,

185: 1431–40. https://pdfs.semanticscholar.org/db74/80f28a68b95ed35701b8

4a282d6ebd8eb366.pdf.

e authors consider the impact of nancial news on stock prices.

Specically, three textual document representations—bag of words, noun

phrases, and named entities—obtained from news articles are considered.

Using SVMs, the authors analyze the impact of news articles on stock

prices 20 minutes after a news article is published. e study yields two

interesting results. First, compared with linear models, the SVM nds that

nancial news has a statistically signicant impact on stock prices. Second,

various textual analysis approaches yield dierent stock return prediction

References

performance. Compared with the popular bag of words, noun phrase tex-

tual representation results in better prediction performance.

Sevim, Cuneyt, Asil Oztekin, Ozkan Bali, Serkan Gumus, and Erkam

Guresen. 2014. “Developing an Early Warning System to Predict Currency

Crises.” European Journal of Operational Research 237 (3): 1095–104. https://

doi.org/10.1016/j.ejor.2014.02.047.

To predict currency crises in the Turkish economy, the authors use a nan-

cial pressure index—which measures drastic deviations in the exchange

rate and drastic decreases in foreign exchange reserves—as the dependent

variable and 32 macroeconomic variables as independent variables. e

three models considered in this article—ANNs, logistic regression, and

decision trees—are able to predict the 1994 and 2001 crises 12 months in

advance and with 95% accuracy.

Simon, Dan. 2013. Evolutionary Optimization Algorithms. Hoboken, NJ: John

Wiley & Sons, Inc.

is applied textbook, divided into ve parts, is devoted to studying evolu-

tionary algorithms for optimization. Part 1 discusses types of optimization

problems and algorithms. Part 2 reviews natural genetics and their history

and describes the use of articial genetic algorithms for solving optimiza-

tion problems. In Part 3, the discussion centers on related techniques, such

as ant colony optimization, particle swarm optimization, and dierential

evolution. Part 4 is devoted to special types of optimization problems

(discrete, constrained, and multi-objective optimization problems) and

problems associated with reducing the computational costs of evolutionary

algorithms. Finally, Part 5 provides a practical guide on how to address

problems (checking for bugs and problems in the code and software) and

how to measure the performance of an algorithm against standard bench-

mark optimization problems.

Skolpadungket, Prisadarng, Keshav Dahal, and Napat Harnpornchai.

2016. “Handling Model Risk in Portfolio Selection Using Multi-Objective

Genetic Algorithm.” In Articial Intelligence in Financial Markets: New

Developments in Quantitative Trading and Investment, edited by Christian

Dunice, Peter Middleton, Andreas Karathanasopolous, and Konstantinos

eolatos, 285–310. London: Palgrave Macmillan. https://doi.org/10.1057/

978-1-137-48880-0_10.

e classical Markowitz (mean–variance) portfolio optimization model

assumes that asset returns are normally distributed. In reality, means and

Articial Intelligence in Asset Management

volatilities of asset returns tend to vary, which requires forecasting these vari-

ables to construct an optimal portfolio. e authors present a solution to the

portfolio optimization problem using a multi-objective genetic algorithm to

account for the inaccuracy inherent in forecasting models. is model risk can

be reduced when an approximation of the Sharpe ratio error of the portfolio

of assets is added as an additional objective to the portfolio optimization task.

Sprenger, Timm O., Philipp G. Sandner, Andranik Tumasjan, and Isabell

M. Welpe. 2014. “News or Noise? Using Twitter to Identify and Understand

Company-Specic News Flow.” Journal of Business Finance & Accounting 41

(7–8): 791–830. https://doi.org/10.1111/jbfa.12086.

Using 400,000 S&P 500 stock-related messages from Twitter, the authors

compare company returns just before positive and negative news. Good

news tends to have a larger information leakage and a greater impact on

stock returns than bad news.

Tam, Kar Yan. 1991. “Neural Network Models and the Prediction

of Bank Bankruptcy.” Omega 19 (5): 429–45. https://doi.org/10.1016/

0305-0483(91)90060-7.

Bank bankruptcy rst increased signicantly in the 1980s. e author uses

a neural network technique to predict bank failure, showing that the pro-

posed method performs better than traditional statistical methods in terms

of robustness, forecast accuracy, adaptability, and explanatory capability.

Tan, Pang-Ning, Michael Steinbach, Anuj Karpatne, and Vipin Kumar.

2018. “Data Mining Cluster Analysis: Basic Concepts and Algorithms.” In

Introduction to Data Mining, 2nd ed., 525–603. New York: Pearson.

e authors provide an overview of cluster analysis and illustrate its appli-

cation in dierent elds. In cluster analysis, data are partitioned into mul-

tiple groups or clusters that share some traits common within their group.

e authors present dierent clustering techniques (e.g., K-means, hier-

archical clustering) and discuss the strengths and weaknesses of various

clustering methods.

Tan, Zhiyong, Chai Quek, and Philip Y.K. Cheng. 2011. “Stock Trading with

Cycles: A Financial Application of ANFIS and Reinforcement Learning.”

Expert Systems with Applications 38 (5): 4741–55. https://doi.org/10.1016/j.

eswa.2010.09.001.

e authors develop a new non-arbitrage algorithmic trading algorithm

based on an adaptive network fuzzy inference system (ANFIS) and RL

techniques. e proposed method predicts the changes in the long-term

References

movement of prices and is able to outperform trading algorithms such

as DENFIS and RSPOP. Experimental trading outcomes using ve US

stocks indicate that on average, total returns using the new framework are

higher by approximately 50 percentage points.

Teräsvirta, Timo, Dick van Dijk, and Marcelo C. Medeiros. 2005. “Linear Models,

Smooth Transition Autoregressions, and Neural Networks for Forecasting

Macroeconomic Time Series: A Re-Examination.” International Journal of

Forecasting 21 (4): 755–74. https://doi.org/10.1016/j.ijforecast.2005.04.010.

e authors consider three techniques—linear autoregressive, smooth tran-

sition autoregressive (STAR), and neural network time-series models—to

forecast macroeconomic variables. Using a dynamic model specication,

the authors produce results indicating that the dynamic STAR model per-

forms better than the linear autoregressive and several xed STAR models

in terms of forecast accuracy. Neural network models can produce more

accurate forecasts when the forecast horizon is long and when the Bayesian

regularization is applied to the model.

Tibshirani, Robert. 1996. “Regression Shrinkage and Selection via the

Lasso.” Journal of the Royal Statistical Society. Series B. Methodological 58 (1):

267–88. https://doi.org/10.1111/j.2517-6161.1996.tb02080.x.

e author introduces the LASSO method in linear regression models. By

adding a penalty term to the mean squared error minimization problem,

LASSO shrinks some coecients to zero and produces models that are

interpretable. Compared with other variable selection techniques, such as

subset selection and ridge regression, LASSO is better suited for small- to

moderate-sized numbers of moderate-sized eects.

Tsai, Chih-Fong, Yuah-Chiao Lin, David C. Yen, and Yan-Min Chen. 2011.

“Predicting Stock Returns by Classier Ensembles.” Applied Soft Computing

11 (2): 2452–59. https://doi.org/10.1016/j.asoc.2010.10.001.

In this article, the authors compare the prediction performance of ensemble

models of classication for stock returns. ey nd that relative to single

classiers, classier ensembles perform well in terms of return on invest-

ment and prediction accuracy.

Tsai, Chih-Fong, and Jhen-Wei Wu. 2008. “Using Neural Network

Ensembles for Bankruptcy Prediction and Credit Scoring.” Expert Systems

with Applications 34 (4): 2639–49. https://doi.org/10.1016/j.eswa.2007.05.019.

e authors compare the forecast accuracy of multiple neural network clas-

siers with that of a single best neural network classier for bankruptcy

Articial Intelligence in Asset Management

prediction and credit scoring problems. e empirical results show that

multiple classiers tend to perform worse than a single best neural net-

work classier. If type I or type II errors are considered, however, neither

method outperforms the other in terms of prediction accuracy.

Vapnik, Vladimir N. 2000. e Nature of Statistical Learning eory, 2nd ed.

New York: Springer. https://doi.org/10.1007/978-1-4757-3264-1.

e author reviews statistical learning for small data samples that do not

rely on a priori information. e book, which is appropriate to be used as

a graduate-level textbook on learning theory, is divided into three parts:

the general theory of learning, support vector estimation, and statistical

foundations of learning theory.

Varetto, Franco. 1998. “Genetic Algorithms Applications in the Analysis of

Insolvency Risk.” Journal of Banking & Finance 22 (10–11): 1421–39. https://

doi.org/10.1016/S0378-4266(98)00059-4.

e author studies insolvency risk and compares results obtained from tra-

ditional linear discriminant analysis and genetic algorithms. Using data on

Italian companies from 1982 to 1995, the author nds that linear discrimi-

nant analysis better predicts insolvency. e genetic algorithms produce

results faster and with less contribution from the nancial analyst, how-

ever, and can therefore serve as an eective tool in bankruptcy risk analysis.

Verikas, Antanas, Zivile Kalsyte, Marija Bacauskiene, and Adas Gelzinis.

2010. “Hybrid and Ensemble-Based Soft Computing Techniques in

Bankruptcy Prediction: A Survey.” Soft Computing 14: 995–1010. https://doi.

org/10.1007/s00500-009-0490-5.

e authors review literature on hybrid and ensemble-based soft computing

techniques used in bankruptcy prediction studies. Because the literature

on bankruptcy prediction rarely reports condence intervals of prediction

results, and studies use vastly dierent data, comparisons of the obtained

results are not feasible. Instead, the authors focus on specic techniques

and their ensembles used in bankruptcy prediction.

Vui, Chang Sim, Gan Kim Soon, Chin Kim On, Rayner Alfred, and Patricia

Anthony. 2013. “A Review of Stock Market Prediction with Articial

Neural Network.” In 2013 IEEE International Conference on Control System,

Computing and Engineering (ICCSCE), 477–82.

ANN can be a useful technique for stock market prediction, given the non-

linear and volatile nature of stock market dynamics. e authors provide a

References

brief review of the literature on the application of various ANN approaches

in predicting stock market returns.

Xing, Frank Z., Erik Cambria, and Roy E. Welsch. 2018. “Natural Language

Based Financial Forecasting: A Survey.” Articial Intelligence Review 50:

49–73. https://doi.org/10.1007/s10462-017-9588-9.

In recent years, the number of papers using textual sentiment data for nancial

forecasting has been increasing. e authors review the natural language based

nancial forecasting literature, discussing the history of NLP techniques, cur-

rent text processing techniques, and algorithms for predictive models.

Xue, Jingming, Qiang Liu, Miaomiao Li, Xinwang Liu, Yongkai Ye, Siqi

Wang, and Jianping Yin. 2018. “Incremental Multiple Kernel Extreme

Learning Machine and Its Application in Robo-Advisors.” Soft Computing

22: 3507–17. https://doi.org/10.1007/s00500-018-3031-2.

Robo-advisors currently used in the nance industry provide investors

with nancial advice previously provided by nance sector employees. e

authors suggest that the ML algorithm robo-advisors use may be less suit-

able when information is heterogeneous. ey introduce an incremental

multiple kernel extreme learning machine model, which can initialize and

simultaneously update the training dataset and combine information from

dierent data sources. e proposed method is able to eciently solve clas-

sication problems and can thus be used as an algorithm for robo-advisors.

Yao, Jingtao, Yili Li, and Chew Lim Tan. 2000. “Option Price Forecasting

Using Neural Networks.” Omega 28 (4): 455–66. https://doi.org/10.1016/

S0305-0483(99)00066-3.

e authors of this study use a neural network option pricing model and

show that the proposed approach achieves dierent performance results

depending on how data are partitioned into groups based on moneyness.

Using Japanese Nikkei 225 Futures data, the authors nd that a neural

network pricing model outperforms the Black–Scholes model when mar-

kets are volatile, whereas the latter performs better when its theoretical

assumption of constant volatility holds. e traditional Black–Scholes

model performs better for at-the-money options. For in-the-money and

out-of-the-money options, a neural network pricing model may be more

appropriate when the preferred strategy is high risk and high return.

Yu, Lean, Shouyang Wang, and Kin Keung Lai. 2008. “Neural Network-

Based Mean–Variance-Skewness Model for Portfolio Selection.” Computers &

Operations Research 35 (1): 34–46. https://doi.org/10.1016/j.cor.2006.02.012.

Articial Intelligence in Asset Management

Researchers have been studying extensions of Markowitz’s mean–variance

model of portfolio optimization. Recent work in this area recommends

incorporating higher-order moments, such as skewness, especially when

asset returns are not normally distributed. e authors of this study present

a neural network–based mean–variance-skewness model of portfolio opti-

mization. ey integrate this model with investors’ risk preferences, dier-

ent forecasts, and trading strategies and show that the proposed algorithm

is computationally fast and ecient in solving the triple trade-os in the

mean–variance-skewness portfolio optimization problem.

Yu, Lean, Shouyang Wang, Kin Keung Lai, and Fenghua Wen. 2010.

“A Multiscale Neural Network Learning Paradigm for Financial Crisis

Forecasting.” Neurocomputing 73 (4–6): 716–25. https://doi.org/10.1016/j.

neucom.2008.11.035.

e authors propose a novel approach to predicting exchange rates. ey

develop a multiscale neural network learning algorithm for the exchange

rates, which are decomposed into multiple independent intrinsic mode

components using the Hilbert EMD (empirical mode decomposition)

algorithm. e ndings show that the new EMD-based multiscale neural

network learning approach performs well in forecasting bank crises and is

superior to other classication methods.

Zetzsche, Dirk A., Douglas W. Arner, Ross P. Buckley, and Brian W.

Tang. 2020. “Articial Intelligence in Finance: Putting the Human in the

Loop.” CFTE Academic Paper Series: Centre for Finance, Technology and

Entrepreneurship, no. 1; University of Hong Kong Faculty of Law Research

Paper No. 2020/006. https://ssrn.com/abstract=3531711.

e authors develop a regulatory roadmap for the use of AI in nance,

focusing in particular on human responsibility and highlighting the neces-

sity of human involvement. After describing various cases of AI’s use in

nance, the authors discuss a range of potential issues, then consider the

regulatory challenges and tools available. e key issues identied are

increased information asymmetries, data dependencies, and system inter-

dependencies leading to unexpected consequences.

Zhang, Gioqinang, and Michael Y. Hu. 1998. “Neural Network Forecasting

of the British Pound/US Dollar Exchange Rate.” Omega 26 (4): 495–506.

https://doi.org/10.1016/S0305-0483(98)00003-6.

Numerous studies successfully apply neural networks to exchange rate

forecasting. In this study, the authors look at the impact of the number

References

of parameters to be estimated on exchange rate forecasting because neural

networks require the estimation of many parameters. Using the GBP/USD

exchange rate forecasting problem as an empirical exercise, the authors

nd that the number of input nodes and hidden nodes aects the forecast-

ing performance of neural networks.

Zhang, Guoqiang, Michael Y. Hu, B. Eddy Patuwo, and Daniel C. Indro.

1999. “Articial Neural Networks in Bankruptcy Prediction: General

Framework and Cross-Validation Analysis.” European Journal of Operational

Research 116 (1): 16–32. https://doi.org/10.1016/S0377-2217(98)00051-4.

e authors provide a comprehensive review of neural network methods

used in bankruptcy prediction studies and nd that these methods are

superior to logistic regression models in bankruptcy forecasting and clas-

sication. e authors explain that the superior performance of neural net-

works results from their link to Bayesian posterior probabilities.

Zheng, Ban, Eric Moulines, and Frederic Abergel. 2013. “Price Jump

Prediction in a Limit Order Book.” Journal of Mathematical Finance 3 (2):

242–55. https://doi.org/10.4236/jmf.2013.32024.

e authors study the empirical relationship between bid–ask limit order

liquidity balance and trade direction and the usefulness of information

contained in limit order books. e results show that bid–ask liquidity bal-

ance helps to predict the direction of the next trade. Further, the authors

examine the intertrade price jump using logistic regression, where the most

informative features in limit order books are selected by the LASSO vari-

able selection technique. Empirical results using data on French stocks

show that trade sign, market order sign, and liquidity on the best bid–ask

are all important factors for price jump prediction.

Zimmermann, Hans Georg, Ralph Neuneier, and Ralph Grothmann. 2002.

“Active Portfolio-Management Based on Error Correction Neural Networks.”

In Advances in Neural Information Processing Systems 14, edited by omas G.

Dietterich, Suzanna Becker, and Zoubin Ghahramani, 1465–72. Cambridge,

MA: MIT Press.

In this study, the authors combine the Black–Litterman portfolio optimi-

zation model with neural networks to forecast excess returns. e fore-

casts of the expected return are based on error correction neural networks

that use the previous model’s error. Using data from 21 nancial markets

in G–7 countries, the proposed portfolio optimization model is shown to

outperform a benchmark portfolio.

Ameritech

Anonymous

Robert D. Arnott

Theodore R. Aronson, CFA

Asahi Mutual Life Insurance Company

Batterymarch Financial

Management

Boston Company

Boston Partners Asset Management,

L.P.

Gary P. Brinson, CFA

Brinson Partners, Inc.

Capital Group International, Inc.

Concord Capital Management

Dai-Ichi Life Insurance Company

Daiwa Securities

Mr. and Mrs. Jeffrey Diermeier

Gifford Fong Associates

Investment Counsel Association

of America, Inc.

Jacobs Levy Equity Management

John A. Gunn, CFA

John B. Neff, CFA

Jon L. Hagler Foundation

Long-Term Credit Bank of Japan, Ltd.

Lynch, Jones & Ryan, LLC

Meiji Mutual Life Insurance

Company

Miller Anderson & Sherrerd, LLP

Nikko Securities Co., Ltd.

Nippon Life Insurance Company of

Japan

Nomura Securities Co., Ltd.

Payden & Rygel

Provident National Bank

Frank K. Reilly, CFA

Salomon Brothers

Sassoon Holdings Pte. Ltd.

Scudder Stevens & Clark

Security Analysts Association

of Japan

Shaw Data Securities, Inc.

Sit Investment Associates, Inc.

Standish, Ayer & Wood, Inc.

State Farm Insurance Company

Sumitomo Life America, Inc.

T. Rowe Price Associates, Inc.

Templeton Investment Counsel Inc.

Frank Trainer, CFA

Travelers Insurance Co.

USF&G Companies

Yamaichi Securities Co., Ltd.

Named Endowments

The CFA Institute Research Foundation acknowledges with sincere gratitude the

generous contributions of the Named Endowment participants listed below.

Gifts of at least US$100,000 qualify donors for membership in the Named Endowment

category, which recognizes in perpetuity the commitment toward unbiased,

practitioner-oriented, relevant research that these ﬁrms and individuals have

expressed through their generous support of the CFA Institute Research Foundation.

For more on upcoming Research Foundation

publications and webcasts, please visit

www.cfainstitute.org/en/research/foundation.

Senior Research Fellows

Financial Services Analyst Association

Available online at www.cfainstitute.org

BARTRAM, BRANKE, AND MOTAHARILITERATURE REVIEW / ARTIFICIAL INTELLIGENCE IN ASSET MANAGEMENT

9 781952 927027

ISBN 978-1-952927-02-7