Forecasting and optimization stock predictions: Varying asset profile, time window, and hyperparameter factors

Abstract

Machine learning has made significant progress in various fields, including financial markets. Numerous studies have applied different machine learning algorithms to predict stock market behavior, but these studies often face challenges in terms of data acquisition and preparation, algorithm design, hyperparameter optimization, and feature selection, as well as the inherent volatility of stocks. In this work, our aim is to review the literature for comprehensive studies that address these challenges and enhance the state-of-the-art by introducing novel factors, such as multi-time windows, training batch size, stopping criteria, training data ratio, and financial technical indicators. We observe statistical significance when varying the training period window, with a p-value lower than 0.0001. However, genetic-based hyperparameter optimization brings about a significant 40% improvement compared to random-grid search. Concerning the inclusion of technical indicators, we see little improvement in terms of prediction accuracy, but there is some improvement in directional prediction accuracy across several stocks. Overall, the results show high variation with respect to the time window chosen for conducting a study. Additionally, we discover that the characteristics of the stock and the time period, including the length of the time period and the specific start and end dates, significantly impact prediction accuracy.