A Study on Relevant Features for Intraday S&P 500 Prediction Using a Hybrid Feature Selection Approach

Abstract

This paper investigates relevant features for the prediction of intraday S&P 500 returns. In contrast to most previous research, the problem is approached as a four class classification problem to account for the magnitude of the returns and not only the direction of price movements. A novel framework for feature selection using a hybrid approach is developed that combines correlation as a fast filter method, with the wrapper method differential evolution feature selection (DEFS) that deploys distance-based classifiers (k-nearest neighbor, fuzzy k-nearest neighbor, and multi-local power mean fuzzy k-nearest neighbor) as evaluation criterion. The experimental results show that feature selection successfully discarded features for this application to improve the test set accuracies or, at a minimum, lead to similar accuracies than using the entire feature subset. Moreover, all setups in this study ranked technical indicators such as 5-day simple moving average as the most relevant features in this application. In contrast, the features based on other stock indices, commodities, and simple price and volume information were a minority within the top 10 and top 50 features. The prediction accuracies for the positive return class considerably higher than the negative class predictions with over $70\%$ accuracy compared to $30\%$.