The Expected Returns on Machine-Learning Strategies

Abstract

We estimate the expected returns of machine learning-based anomaly trading strategies and quantify the impact of three factors often overlooked in the previous literature: transaction costs, post-publication decay, and the post-decimalization era of high liquidity. Despite a cumulative performance reduction averaging about 57% when accounting for these three factors, sophisticated machine learning strategies remain profitable, particularly those employing Long Short-Term Memory (LSTM) models. We estimate that our most effective strategy, the one based on an LSTM model with one hidden layer, has an expected gross (net) Sharpe Ratio of 0.94 (0.84). We rationalize these findings in a simple theoretical framework where technological diffusion gradually erodes trading profits while superior signal processing capabilities allow the extraction of alpha from increasingly complex information.

Keywords: Stock market anomalies, machine learning models, return prediction, transaction costs, asset pricing models.

JEL Classification: G11, G12, G14, C45, C58