Thompson Sampling vs Ols: Analyzing Tariff Impact on Export Decisions

This paper, based on the Melitz model, examines how effective trade costs influenced European Union (EU) steel exports to the U.S. between 2008 and 2024. It employs both Ordinary Least Squares (OLS) regression and the Thompson Sampling (TS) algorithm to forecast export volumes, using cumulative regret and its growth rate as key evaluation metrics. Empirical results show that the OLS model achieves a lower initial cumulative regret, with an average of 5.3% deviation from actual exports in the early years. However, over time, the TS algorithm surpasses OLS in adaptability, with the growth rate of cumulative regret declining from 6.7% to 2.1% after 2020, indicating improved long-term prediction accuracy under uncertain conditions. The dynamic updating mechanism of TS allows it to adjust to policy shifts and economic volatility more effectively than static models. These findings suggest that reinforcement learning-based models like TS are better suited for forecasting in rapidly changing trade environments. The study offers valuable insights for firms and policymakers seeking robust tools to anticipate trade flows and develop adaptive export strategies.