Financial market predictions with deep learning

Abstract

Forecasting the financial market has proven to be a challenging task due to high volatility. However, with the growing involvement of computational methods in econometrics, models built with deep learning neural networks have been more accurate in capturing the dynamics of financial market data compared to the commonly used time series models such as the ARIMA and GARCH models. In this study, four deep learning models were applied to eight separate investments, namely stocks (AAPL, TSLA, ROKU, BAC), currency exchange rates (GBP/USD and USD/SEK) and exchange-traded funds (SQQQ and SPXS) to compare their forecasting abilities. The four deep learning models consists of three recurrent neural networks (RNN) which are the vanilla recurrent network (VRNN), long short-term memory (LSTM) and gated recurrent units (GRU), along with the convolutional neural networks (CNN). The models were tuned to be time efficient and evaluated with RMSE and MAPE. Results show that GRU was the overall best model, with exceptions to the LSTM performing better with the exchange traded funds.