Using LSTM-Predicted Stock Prices and Risk-Adjusted Performance Metrics to Optimize Portfolios in the European Market

Abstract

[EN] Long short-term memory (LSTM) neural networks allow to capture long-range dependencies and non-linearities in sequential data and can learn complex patterns and relationships in the data improving the accuracy of future stock price predictions. Since classical portfolio optimization is highly sensitive to the estimated parameters used to construct an optimal portfolio, the purpose of our research is to leverage LSTM abilities to predict the parameters accurately and create portfolios that generate superior results.We predict the prices of the 50 components of the EURO STOXX 50® Index using LSTM and create prediction-based optimal portfolios for ten different investment time horizons. We define the risk as a combination of the standard deviation and the performance of the evaluation metrics obtained testing our model, allowing us to use a measure for the risk based on the level of confidence the model has in the prediction.Our portfolios consistently beat the market over the analyzed investment scenarios from 2021 until the first half of 2022 and are robust for both growing and bear markets. The proposed model achieves an average MAE of 0.01634, an average MSE of 0.00047, and an average accuracy of 95.8% in predicting the direction of the stock movements over the ten proposed periods.Our research contributes to the field of finance by providing an innovative framework for portfolio optimization that leverages the power of LSTM-based stock price prediction and risk-adjusted performance metrics.