Predicting Medium-Term Stock Index Direction Using Constituent Stocks and Machine Learning

Abstract

Predicting stock index movements is challenging due to market randomness. This paper addresses the problem of predicting medium-term stock index direction, an area with limited coverage in existing literature. We propose a novel model based on machine learning algorithms that employs relative indicators of constituent stocks within an index. The objective is to identify market entry points where the likelihood of achieving a significant return threshold is higher. To achieve this, supervised binary classifiers and other machine learning techniques have been applied, setting the class label '1' for significant index rises occurring within a defined medium term of 70 trading days and the class label '0' otherwise. Three indices were investigated: Nasdaq 100, where a significant rise was defined as a 10% increase, Dow Jones Industrial Average with an 8% increase, and the German Dax at 6%. Our investigation into different methods of utilizing constituent stocks revealed that focusing on the most weighted stocks yields the most promising results across various stock indices. The proposed model dynamically selects the most effective classifiers, SVM, KNN, Voting Classifier, and RF, tailored to varying market conditions. Employing a rolling forecast method, it utilizes the relative indicators on heavily weighted stocks and the index, demonstrating accuracy up to 0.97 and F1-scores for the '1' label up to 0.90. This enhances the ability to determine the optimal timing for market entry and, crucially, when the chances for high returns are limited. Additionally, we illuminate the conditions under which the model is most effective.