Stock turning points classification using a novel discrete learning-based methodology

Abstract

Achieving accurate turning point (TP) forecasting techniques can provide investors with tools to make profitable trading decisions, by offering the opportunity to buy = low and sell = high. Forecasting stock time series TP is known to be one of the most important but also extremely challenging issues. The stock TP forecasting problem is a classification problem where the decision to buy or sell a stock has to be made at each point in time. The first step to achieve this goal is to identify and label the TPs present in the history of the stock time series through a detection technique. Once the labels are achieved, the corresponding classifier is trained to assign labels to new, invisible observations. According to the literature, improving the accuracy of the classifier leads to an increase in the forecasted TPs’ accuracy. Various intelligent classification techniques with different properties have been developed, Over the past decades, to achieve more accurate classification results. However, the parameter estimation of classifiers is done by optimizing a continuous distance-based cost function; and since the goal of the classification problem is discrete, applying a continuous cost function is unreasonable, or at least quite inefficient. To address this, we introduce a novel discrete learning-based methodology for estimating the unknown parameters of the classifier and thereby forecasting stock TPs. In contrast to the existing classification techniques, the proposed methodology is tailored to the discrete goal function of the classification. The present study chooses the multilayer perceptron (MLP), which is one of the most widely-used intelligent classification techniques, to implement the proposed methodology. Although it can be generally shown that the forecasting performance of the proposed discrete learning-based multilayer perceptron model (DIMLP) will not be worse than that of its continuous learning-based counterparts, this paper compares it to other classification techniques to determine the superiority of the proposed DIMLP model. Experimental results using randomly selected datasets from the Shanghai Stock Exchange show that the proposed DIMLP model is superior to its counterparts including classical MLP, Support Vector Machine, Random Forest, Decision Tree, K-Nearest Neighbor, Logistic Regression, Radial Basis Function Network, Generalized Logistic Regression, Probabilistic Neural Network, and AdaBoost. Consequently, further support is provided for the hypothesis that using a discrete learning-based function as a cost function for classification purposes is appropriate and effective. © 2023, The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature.