When Does Machine Learning Outperform Statistical Models? Trade Forecasting with Data Irregularity in OECD and Emerging Economies

초록

This study examines how the structural characteristics and irregularity of trade data affect the predictive performance of empirical modeling tools. Using trade panel data from 80 countries, we distinguish between OECD and emerging economies to assess whether the relative effectiveness of forecasting methods varies across data environments. It compares the out-of-sample forecasting performance of traditional statistical models and machine learning approaches in predicting export growth. Pooled OLS is employed as a benchmark model, while Random Forest-based machine learning and Transformer-based deep learning models serve as alternative tools. Predictive accuracy is evaluated using RMSE, MAE, MAPE, and out-of-sample R2. The results show that machine learning models consistently outperform pooled regression models in trade growth forecasting. Importantly, this performance advantage is substantially stronger in emerging economies, where trade data exhibit higher volatility and structural irregularity, than in OECD countries. Robustness checks using alternative panel regression specifications confirm that the predictive superiority of machine learning models is amplified in more irregular data environments. Unlike existing studies that treat model superiority as universal, this study demonstrates that the predictive advantage of machine learning is conditional on data structure. By explicitly linking forecasting performance to trade data irregularity, it offers new methodological guidance for selecting appropriate analytical tools in trade research.

키워드