The Comparison of Feature Selection Methods and Feature Combinations Based on the Stock Prediction

Abstract

Predicting short-term stock prices is a significant and challenging research area due to market volatility. Machine learning (ML) uses algorithms to learn patterns from data, improving prediction accuracy over time. Feature selection (FS) methods enhance model accuracy and efficiency. Evaluating and selecting the best FS methods and feature combinations is essential for improving prediction performance. This paper evaluates three feature selection (FS) methods by scoring technical indicators and using three models with 30 different indicator combinations to predict outcomes. Error rates are used to measure accuracy. The analysis reveals that incorporating all features yields the lowest average error rate in this paper, and Williams R is attached to the greatest importance in this paper. Among the FS methods, Mutual Information (MI) and Random Forest (RF) outperform the correlation coefficient. Future work can focus on two main areas: considering more indicators and combinations, and exploring additional feature selection (FS) methods to identify the best one.