LSTM Networks in Financial Decision-Making: Stock Price Prediction and Portfolio Optimization

Abstract

This research seeks to assess how well Long Short-Term Memory (LSTM) networks predict stock prices, as well as how LSTM might be applied to portfolio optimization. The research proposes a diversified portfolio containing 20 stocks of the Casablanca Stock Exchange, and CVaR (Conditional Value At Risk) was used for portfolio optimization, relying on historical data and LSTM predicted data for the same period. The predictive performance of the LSTM model was quite robust when looking at Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R-squared (R2). However, when considering portfolio optimization with predicted LSTM compared to historical data, the expected return and CVaR are substantially different, as are the optimal portfolio asset weights. The findings suggest that it is essential to be careful and deliberate in using LSTM for stock price forecasting, especially when analyzing portfolio optimization and asset weight.