Are Natural Language Processing methods applicable to EPS forecasting in Poland?

Abstract

Accurate earnings forecasts are crucial for successful investment outcomes, especially in emerging markets like Poland, where analyst coverage is limited. This study investigated the applicability of natural language processing (NLP) techniques, specifically FastText and FinBERT word embeddings, combined with a gradient-boosting decision tree (XGBoost) machine learning algorithm, to forecast earnings per share (EPS) for companies listed on the Warsaw Stock Exchange from 2010 to 2019. The performance of these models was compared with a seasonal random walk (SRW) model. The SRW model consistently demonstrated the lowest error rates, as measured by the mean arctangent absolute percentage error, and outperformed the NLP-based models across different periods and error metrics. The superior performance of the simple SRW model can be attributed to the overparameterization and overfitting tendencies of the complex NLP models, as well as the relatively straightforward dynamics of the Polish stock market. The findings suggest that the application of sophisticated NLP techniques for EPS forecasting in Poland may not be justified, and that the SRW model provides a more accurate representation of the market’s behavior.