Deep Learning and the Cross-Section of Stock Returns: Neural Networks Combining Price and Fundamental Information

Abstract

Traditional empirical finance research uses hand-engineering and trial-and-error to look for the anomalies in the cross-section of stock returns. In this paper, we take advantage of deep learning and utilize both the price and fundamental information to separate stocks’ winners from losers. For the first model, we used a 2-layer Long Short-Term Memory (LSTM) neural network with past 80 days’ return information as inputs to predict the next day’s return and find a before-trading-cost monthly return of 29.58% with a t-statistic of 26.81. The return of the stocks shows a strong short-term reversal pattern. For the second model, we design a novel 2-layer LSTM and Multi-layer Perceptron (MLP) hybrid neural network and utilize the monthly return and annual accounting data to predict the returns in the next month. We achieve a monthly return of 2.37% with a t-statistic of 8.97 before trading cost from 1993 through 2017. We use TAQ intraday data to explicitly estimate the trading cost and find that profits of the daily trading strategy in the first model turn negative. However, the trading strategy utilizing both the price and fundamental information in the second model keeps significantly positive with a monthly return of 1.57% and t-statistic of 6.03 after the trading cost. We show that the 2-layer LSTM and MLP hybrid model performs better than the MLP-only and the hand-engineering momentum and short-term reversal double sort trading strategy.