Training of LSTM Configurations for Alpha Generation through Trend Prediction in SPY

In this study, we aimed to improve the prediction of SPY trends using Long Short-Term Memory (LSTM) networks through advanced hyperparameter optimization techniques. Initially, manual tuning of key parameters such as window size, batch size, learning rate, and the number of neurons was performed to establish a baseline performance. The model was trained on financial data with various technical indicators after extensive preprocessing, including winsorization, wavelet transform for noise reduction, and normalization. To enhance the model's accuracy, Bayesian optimization and DEAP neuroevolution were employed. Bayesian optimization, leveraging probabilistic modeling, proved particularly effective in navigating the high-dimensional hyperparameter space, significantly improving the precision and recall of the model. DEAP neuroevolution, inspired by natural selection, provided another layer of optimization, though it was found to be less effective than Bayesian optimization in this context. The Bayesian optimized model was further analyzed for consistency. It was observed that training with a higher standard deviation constant and testing with a lower one significantly improved the precision of trend predictions. This strategy revealed hidden trends, enhancing the model's predictive power. The results showed precision improvements from 67% to 83% for negative trends and 62% to 85% for positive trends, despite a slight reduction in recall. The variability in monthly profits was estimated to range between 2.1% to 6.3%, which is reasonable given the model's high precision and the dynamic nature of financial markets. This study underscores the importance of regular retraining and adaptive strategies to maintain model performance over time.