Predicting Style Factor Returns and Group/Sector Returns Using Long and Short-Term Memory (“LSTM”) Deep Learning Neural Networks

Abstract

of applications including time series analysis, document classification, speech,
and voice recognition. In this study we employ LSTM for predicting out-ofsample
style factor returns and group/sector stock returns derived from the
countries, industries, and style explanatory variables of a cross-section factor
model. The data considered for the analysis is from September 2013 to June 2023
(approx. 10 years) of 4 style factor returns and 9 stock market groups/sectors for
the South African stock market. One of the challenges of using factor models to
forecast returns is the assumption that the prior consecutive observations are
independent of each other as a result they do not account for the previous
observations. Deep learning models like the LSTM are more accurate in predicting
these sources of expected returns with their time-series behavior they can
accurately predict markets where the effects of multiple market variables have
interdependence. The results show that LSTM model is a powerful tool that can
be used to predict returns which can help investors and portfolio managers who
make investment decisions by grouping stocks into style, countries, or sectors.