Forecasting movements in the stock market has long been one of the most desired yet frustrating tasks in data science. Stock markets are volatile, emotional, and driven by factors that no individual model could ever hope to capture entirely. In this study, we analyze and compare how well LSTM neural networks can predict stock movement trends and, more specifically, what impact the architecture itself may have. Using ten years of daily trade data (from 2014 through 2024) for three stocks – Apple Inc. (AAPL), Reliance Industries (RELIANCE.NS), and NIFTY 50 index, we have compared how well five different architectures – starting with a simple one-layer LSTM to a more sophisticated approach of a three-layer stacked LSTM with attention mechanisms – fared in their predictions. Our feature selection ranges from technical analysis tools, such as RSI, MACD, and Bollinger bands, to economic indicators, like VIX and spread between treasury yields. Of our models, the stacked LSTM with attention turned out to have the best performance, with a MAPE of 1.73% and accuracy in predicting price direction of 84.6%, significantly better than any other model tried (ARIMA, GRU, XGBoost, LSTMs). We have tested our models on the stock market crash caused by COVID-19 in early 2020 to see their performance during out-of-domain situations, i.e., when market conditions drastically change, as well as what those results imply about the use of these models in practice.
Verma et al. (Mon,) studied this question.