Deep Learning for Equity Time Series Prediction

Abstract

We examine the performance of Deep Learning methods applied to equity financial time series. Predicting equity time series is a crucial topic in Finance. To form equity portfolios and do asset allocation, we need to predict returns, compute their risk, and optimize market impact. One of the modeling benefits of Deep Learning architectures is the ability to model non-linear highly dimensional problems. The lack of transparency and a rigorous mathematical theory could be considered less positive sides. The fact that most progress in Deep Learning has been made by trial and error is also cumbersome. Equity financial time series is a challenging domain with some stylized facts: weak stationarity, fat tails in return distributions, small data sets compared to other areas of Artificial Intelligence (AI), slow decay of autocorrelation in returns, and volatility clustering, to name the most important ones. We perform a comparative study between Long ShortTerm Memory Networks (LSTM), Recurrent Neural Networks (RNN), Deep Feed-Forward neural networks (DNN), and Gated Recurrent Unit Networks (GRU). We perform two types of studies. The first focused on a univariate test, and the second a multivariate test. Our tests show that the LSTM performs the best compared to other Deep Learning and classical machine learning models. In terms of performance metrics, the LSTM is better than the baseline model. We also show that the predictions are better than chance. There is enough evidence thatRNN and LSTM can deal with stationary time series and learn the data generating process. Nevertheless, predicting equity non-stationary time series, with market developments like the one caused by the COVID-19 pandemic in 2020, is challenging.