Forecasting the EUR/PLN Exchange RateUsing Classical and Artificial Intelligence Methods:An Empirical Comparison of ARIMA, XGBoost, LSTMand Hybrid Models on NBP Data 2015-2026
Purpose: The primary objective of this paper is to conduct an empirical assessment of the forecasting performance of various exchange rate prediction approaches and to investigate whether contemporary machine learning techniques and hybrid modeling frameworks are capable of outperforming traditional econometric models. Design/Methodology/Approach: This paper presents an empirical comparison of four approaches to short-term daily EUR/PLN exchange rate forecasting: the classical ARIMA(1,1,1) model, the gradient boosting algorithm XGBoost, a single-layer LSTM recurrent neural network, and a hybrid ARIMA+XGBoost model. Experiments were conducted on daily data from the National Bank of Poland spanning 2 January 2015 to 27 February 2026 (2,815 observations), using a rolling one-step-ahead forecast procedure and a train/test split with the final year as an out-of-sample evaluation period (252 observations). Findings: Results indicate that ARIMA(1.1.1) achieves the lowest error (RMSE=0.0107, MAPE=0.17%), marginally outperforming XGBoost (MAPE=0.19%) and LSTM (MAPE=0.29%). The hybrid ARIMA+XGBoost model, however, exhibits an anomalous result (MAPE=2.35%), generating a near-constant forecast line at approximately 4.14 PLN throughout the entire test period a result thirteen times worse than the standalone ARIMA benchmark. This anomaly, visible as a flat line in the forecast visualization, constitutes an open methodological question addressed in the second part of this series. Practical Implications: The methodologies and empirical findings presented in this paper may provide practical value for organizations for which exchange rate forecasting plays a critical role. These include, in particular, corporations, financial institutions, and central banks, all of which continuously seek advanced approaches to currency risk management, financial instrument valuation, economic monitoring, and decision-making in international trade. Originality/Value: The paper contributes to the literature on exchange rate forecasting by assessing the performance of contemporary artificial intelligence methods and hybrid modeling approaches in a real-world financial market environment. Furthermore, it identifies novel methodological issues that warrant further investigation and provide directions for future research.