How XGBoost May Help in Multivariate EPS Forecasting for Companies Listed on the Warsaw Stock Exchange

Abstract

Purpose | The present analysis evaluates the forecasting capabilities of seven distinct methodologies: seasonal random walk (SRW), Laurent framework (L), Lev and Thiagarajan approach (LT), Residual Income methodology (RI), Pope and Wang framework (PW), Earnings Persistence approach (EP), and Hou, van Dijk, and Zhang methodology (HDZ) in predicting earnings per share. Research method | It uses the eXtreme Gradient Boosting (XGBoost) approach, which can handle non-stationary data. This approach is one of the most efficient to apply for tabular data. To examine accuracy of predictions various metrics are utilised and the statistical significance of difference between them is tested using the Wilcoxon and Brunner-Munzel tests. Results | Both the seasonal random walk framework and the Pope Wang methodology, when implemented through XGBoost, exhibited minimal prediction errors and generated superior representations of the Polish market dynamics relative to alternative approaches. Originality / value / implications / recommendations | Such comprehensive application of multiple models for earnings per share prediction within the Polish market utilising XGBoost methodology represents an unprecedented approach. The notable effectiveness of the relatively simple seasonal random walk framework potentially reflects the less complex characteristics of the Polish equity market. In contrast, the robust performance of the Pope Wang framework indicates the importance of specific financial indicators.