Evaluation of Alternative Approaches in Classification Algorithms for Prediction of Stock Market Index: Case of Crobex

Introduction

Stock market index is often used to describe the aggregate movements in a market as it represents a measurement of value of a section of the stock market and reflects the level of the free float value. Since its changes involve a stochastic component, it is very difficult to predict and forecasting stock market index value has always been one of the hottest topics in research. One of the ways of analyzing stock movement is technical analysis which consists of using past market information to predict the direction of prices.

The problem of automatic recognition of favorable days for intra-day trading can be modeled as a binary classification problem. Day trading consists in buying and selling financial instruments within the same trading day and favorable for intra-day trading means that the increase between the opening price and the closing price of the same day is large enough for obtaining a profit by buying at the opening price and selling at the closing price. Thereby “large enough” is defined by selecting a certain threshold. The underlying classification problem consists of attribution of one of two labels to daily records according to values of a selection of technical indicators. The class that is assigned to each daily record is 1 if the subsequent day is favorable for intra-day trading and 0 otherwise. The problem is to discover the relation between the selected indicators at a certain day and the market situation at the following day that would determine the class of day that is observed, according to technical analysis (Lo & Hasanhodzic, 2010). The objective, as in any classification problem, is to achieve the highest accuracy possible when comparing the prediction with the real class.

The goal of this chapter is to evaluate and compare different classification algorithms on automatic recognition of favorable days for intra-day trading. Classification problems has been intensively studied and there are many classification algorithms available today (Bruni & Bianchi, 2015; Hastie, Tibshirani, & Friedman, 2001). Different families of algorithms are tested: neural network, support vector machine, random forest, as well as k-nearest neighbors and naïve Bayes classifier as more typical classifiers. The idea is to test different families of algorithms but to give more attention to algorithms which are used more rarely than typical statistical methods which are still considered to be “mainstream” although is a lot of research in alternative approaches nowadays too. The algorithms will be applied to Croatian stock index CROBEX and it is the first attempt of such kind to the best of authors’ knowledge and should fill some gap in this area of research observed in literature (Šego & Škrinjarić, 2018). The results will give insight to effectiveness of technical analysis in predicting the index value, as well to observe the accuracy of different classification methods.