3. Portfolio Management
© 2020 CFA Institute Research Foundation. All rights reserved. 11
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
(Rasekhschae, Christian, and Jones 2019; Borghi and De Rossi, forthcom-
ing). Recent ndings indicate that AI signals generate signicant prots in
both short and long positions (0.78% abnormal returns per month for a long-
only, value-weighted portfolio) and that these prots remain statistically and
economically signicant 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 dicult, because their prices and payos are
not well dened 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