Abstract
Abstract
This study investigates the influence of input window size and ensemble size on the forecasting performance of Decision Tree-based ensemble methods in one-step-ahead financial return prediction. Nineteen financial instruments spanning stocks, indices, commodities, and cryptocurrencies are evaluated under a chronological train-validation-test framework. Daily closing prices are transformed into log returns, and models are trained across rolling windows of 5, 10, 20, and 60 observations. The methods compared include Naive forecasting, Linear Regression, a standalone Decision Tree, Bagging-DT, Boosting-DT, and Stacking-DT, with ensemble configurations selected based on validation MAE. Linear Regression achieves the lowest average magnitude error and the best overall rank, while Bagging-DT and Boosting-DT substantially outperform both Naive forecasting and the standalone Decision Tree. The validation-selected ensemble reduces MAE across all 19 instruments, yielding an average out-of-sample improvement of 31.89 % relative to the single tree. A test-best diagnostic comparison suggests a higher improvement ceiling of 35.10 %, though this figure serves solely as an upper-bound reference. A 20-day window delivers the strongest average error performance, while a 60-day window and 100-estimator ensembles are most frequently selected at the individual instrument level. Larger ensembles incur additional computational cost without producing consistent accuracy gains. Friedman and Wilcoxon tests confirm significant performance differences among models, and Diebold-Mariano tests indicate that ensemble methods consistently outperform both Naive forecasting and the standalone Decision Tree. Overall, Decision Tree ensembles offer robust improvements over persistence-based and single-tree forecasting, though their advantage relative to Linear Regression remains instrument-dependent.