Feature Learning Behavior in Trade Forecasting: Evidence from Tree-Based Machine Learning Models

Received: 17 January 2026 | Revised: 9 February 2026 and 19 February 2026 | Accepted: 22 February 2026 | Online: 1 April 2026
Corresponding author: S. Balasubramanian

Abstract

This study investigates how tree-based machine learning models learn from historical trade data by examining the importance of lagged export and import values in the context of India's monthly aggregate trade series from Fiscal Year (FY) 1990–91 to FY 2024–25. Rather than prioritizing predictive accuracy, the analysis focuses on feature-based learning behavior using lagged inputs constructed within a supervised learning framework. A time-aware training–testing split is employed to preserve chronological integrity. Feature importance analysis is treated as the primary analytical outcome to evaluate how models allocate learning weights across short-term and medium-term temporal inputs. The results show that recent export lags consistently dominate model learning, whereas selected medium-term import lags also contribute meaningfully. Robustness checks using alternative lag specifications confirm the stability of short-term dominance. These findings demonstrate that tree-based trade forecasting models learn in a structured and selective manner rather than distributing attention uniformly across historical observations. By emphasizing interpretability over pure accuracy comparisons, the study provides deeper insight into temporal learning dynamics in machine learning-based trade forecasting.