Delta of tomorrow: Economic variables affecting LSTM neural network model’s accuracy and returns in stock price prediction

Abstract

This paper studies economic variables affecting returns and accuracy of a market-neutral stock portfolio constructed by using Long Short-Term Memory (LSTM) model, which is a specific state-of-art sequence learning technique under Recurrent Neural Networks (RNN) (Sezer et al., 2020). By using out-of-sample daily returns of SP500 constituents from December 1992 to September 2022, I study how changes in different economic variables across different economic situations affect the accuracy and returns of the LSTM model-based portfolio.

Based on the study I find that: (i) Federal funds effective rate (FEDFUNDS) and the term spread (TS30) are the most significant predictors of LSTM portfolio returns and accuracies (ii) The predictive power of macroeconomic variables increase during higher market volatility resulting in higher R2 values and greater number of statistically significant variables. Despite the increased predictability, the returns tend to decrease during higher volatility seasons, indicating that the model is affected by the noise during such periods. This aligns with the efficient market hypothesis, indicating that the model rather captures noise than valuable information during high volatility periods (iii) the LSTM model’s decision-making for choosing stocks into the portfolio tends change within different economic seasons, possibly affecting the predictability of the model over time. This rationalizes the explanations why certain macroeconomic variables might be statistically significant only for limited periods of time.

Building on top of established studies on the performance of different neural network techniques, this study improves our understanding of how macroeconomic variables influence the returns and accuracy of LSTM predictions in different economic conditions, and how different economic seasons shape the LSTM model outcomes and thus, its predictability.