A1950
Title: Machine learning estimation of Armington elasticities and their implications for trade policy simulation
Authors: Jong-Hwan Ko - Pukyong National University (Korea, South) [presenting]
Il Do Ha - Pukyong National University (Korea, South)
Abstract: Computable General Equilibrium (CGE) models are widely used in trade policy analysis, but key behavioral parameters such as Armington elasticities are usually calibrated from a single benchmark Social Accounting Matrix rather than estimated from data. This approach produces fixed values without uncertainty and may distort policy results when parameters are mis-specified. A machine learning framework is proposed to estimate sector- and region-specific Armington elasticities and examine how using these estimates changes trade policy simulation outcomes. Using panel data from GTAP versions 6-12 for 1997-2023, together with bilateral trade data from UN Comtrade, XGBoost, Gaussian Process Regression, and Bayesian Ridge Regression are applied to estimate heterogeneous elasticities across 16 regions and 22 sectors. The results show substantial cross-sector and cross-region variation that is masked by conventional calibration practices. These estimated elasticities are incorporated into a 16-region, 22-sector global CGE model to simulate U.S. tariff increases against major trading partners in 2025-2026. Compared with standard calibration, the ML-based estimates produce meaningfully different policy outcomes, including larger welfare losses in import-competing sectors and weaker export gains in capital-intensive sectors. The findings suggest that data-driven elasticity estimation can improve the realism and policy relevance of CGE-based trade analysis.
