FORECASTING STOCK MARKET DYNAMICS USING BIDIRECTIONAL LONG SHORT-TERM MEMORY

Abstract

This study forecasts stock market dynamics using machine learning techniques. Specifically, we use long short-term memory (LSTM) and bidirectional LSTM (Bi-LSTM) networks to predict the spot index return and implied volatility series in the Korean market. The Bi-LSTM model exhibits better out-of-sample forecasting performance than the LSTM and classic autoregressive models do, reflecting the fact that the Bi-LSTM model learns data patterns more accurately through a bidirectional process. The Bi-LSTM model with the longest time lag (i.e., 22 days) exhibits the best performance in predicting returns and volatility over the entire sample period. In contrast, during the global financial crisis and COVID-19 pandemic periods, when the stock market dynamics are unstable, Bi-LSTM models with shorter time lags (i.e., five or ten days) predict volatility more accurately.