Comparative Forecasting Performance of ML Regression Models: Evidence from Borsa Istanbul
Comparative Forecasting Performance of ML Regression Models: Evidence from Borsa Istanbul

Summary/Abstract: Purpose - This study examines stock price forecasting using regression-based machine learning models by leveraging the daily closing prices of 10 companies traded on Borsa Istanbul (Borsa İstanbul) over the 2016–2025 period. The research aims to inform model selection in financial time series forecasting by comparing predictive performance across forecast horizons defined as short-term (≤60 trading days ahead), medium-term (61–180), and long-term (≥180). Design/ methodology/approach - Five regression-based machine learning models were employed: Linear Regression, Bayesian Linear Regression, Decision Tree Regression, Neural Network Regression, and Poisson Regression. Data were split in chronological order (no shuffling). Models were trained using three training shares (80%, 90%, and 99%) and evaluated out-of-sample for each forecast horizon. Performance was assessed consistently across the manuscript using Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE; reported as “error rate (%)”), minimum–maximum percentage error, and the coefficient of determination (R²). Findings - Results show that no single model dominates across all horizons. Neural Network Regression performs best for long-term forecasts, while Decision Tree Regression and Poisson Regression perform comparatively better in the medium term. For short-term forecasts, Decision Tree Regression delivers the lowest error. Linear and Bayesian Linear Regression exhibit relatively higher errors under short- and medium-term volatility. Discussion - The findings indicate that forecasting performance depends on the investment horizon, the temporal structure of the series, and prevailing market conditions. By providing a comparative evaluation on long-span, recent Borsa Istanbul data, the study suggests that algorithm choice in investment strategies should be optimized jointly by forecast horizon and data characteristics.

• Details
• Contents