Abstract

Global forest-product trade faces severe disruptions from external shocks, necessitating robust forecasting and risk assessment tools. To capture the complex, nonlinear dynamics of the forest bioeconomy, this study constructs a comprehensive panel covering 229 countries and five product categories from 1995 to 2023 using FAOSTAT and World Bank data. We propose Forest-TAN, a novel multi-horizon temporal attention network that integrates static country–product feature embeddings with causal temporal convolution and multi-head attention. This architecture allows for the joint forecasting of exports at t + 1, t + 3, and t + 5 horizons. Comprehensive evaluations against statistical and advanced deep learning baselines (including ARIMA, XGBoost, and Transformer models) demonstrate that Forest-TAN significantly mitigates long-memory decay and cross-horizon error accumulation, achieving the lowest forecasting errors (RMSE) across all horizons. Furthermore, based on these forward-looking forecasts, we construct the Forest Bioeconomy Trade Resilience Index (FBTRI). Explainability analysis using SHAP and FBTRI assessments reveal that high-income countries and high-value-added processed products possess substantial systemic advantages in shock absorption and trade resilience. Ultimately, this research provides an interpretable, quantitative framework for global trade risk early warning, resource constraint evaluation, and dynamic scenario simulation.

IPC Classification

Keywords