Machine Learning and Expected Returns

Abstract

Financial data is characterized by a low signal-to-noise ratio making it difficult to identify robust functional forms that map the characteristics of financial securities to expected returns (Lettau and Pelger, 2020). In this paper, we modify the standard prediction problem in empirical asset pricing by replacing realized returns with an estimator for expected return developed by Martin and Wagner (2019). We use a neural network to map expected returns to 164 stock characteristics and their interactions with eight macroeconomic time-series, resulting in 1476 predictors. Portfolios based on the predictions from the neural network generate risk-adjusted returns with respect to the Fama-French 6-factor model in the range of 1.4% (t-statistic of 3.04) to 1.2% (t-statistic of 2.65) before and after transaction costs; out-of-sample. The corresponding Sharpe ratios are 1.15 and 1.06. A similar analysis based on realized returns results in Sharpe ratios below the market portfolio.

Keywords: Machine Learning, Big Data, Equity Options, Return Prediction, Cross-Section of Returns, Transaction Costs

JEL Classification: G12, G11