Comparison of hyperparameter tuning optimization methods for LSTM and stock price prediction

Abstract

This paper investigates using Long Short-Term Memory (LSTM) networks for predicting stock prices, focusing on major stocks like AAPL, MSFT, TSLA, META, and GOOG from 2016 to 2021. We employ several technical analysis indicators, such as moving averages, the relative strength index, and others, as inputs to our LSTM model. The study involves data preprocessing, optimization, and tuning of ten different hyperparameters to enhance the performance of LSTM model. A comparative analysis of optimization techniques, including standard random search, Bayesian, Nevergrad optimization, covariance matrix adaptation optimization, and other bio-inspired algorithms, shows variations in performance across most datasets. Performance is measured using typical statistical metrics like mean squared error, R2 score, and others, with results showing varying prediction accuracies among different stocks. The study highlights the critical influence of data quality on LSTM performance and suggests further research into optimal hyperparameter tuning for enhancing AI-driven financial analytics.