A Long Short-Term Memory Network-Based Approach for Predicting the Trends in the S&P 500 Index

The price level of a considered security in the stock market is largely determined by perceived consumer demand. Variations in price level in the stock market are essentially a manifestation of public psychology aiming to either cut losses or maximize profits. The assumption that market movements are controlled by a guiding trend is proved by establishing that the time-series variations of daily S&P 500 indices are a non-stationary process using the Dickey–Fuller test. This paper presents a long short-term memory (LSTM) network adaptation of an autoregressive predictive model for S&P 500 indices. Later, an LSTM network-based model is devised to learn the underlying patterns that are responsible for driving the trends in market dynamics. The proposed model accepts as input at each time-step a vector of prior S&P indices over the past two weeks and estimates the current market index. The network weights assume the correct settings by training the model using daily S&P indices over the past ten years. The proposed model decisively captures subtleties in market trends, which are in turn verified using the Granger causality approach. The result demonstrates that the proposed model achieved an accuracy of 82.6%, whereas random forest results in 79.12%. The results obtained are truly the latest in this line of predictive models and reaffirm that LSTM networks are powerful learning models capable of deciphering intricate underlying patterns. © 2023, The Institution of Engineers (India).