Cross-Country Transfer Learning for FDI Forecasting in Small Macroeconomic Datasets
This study examines whether transfer learning can improve machine learning performance under data-scarce forecasting conditions. The empirical application focuses on forecasting foreign direct investment (FDI) inflows using CatBoostRegressor. The source domain is represented by a broad international panel of countries, while the target domain consists of Central Asian economies.
CatBoost is evaluated in two settings: without transfer learning, where the model is trained only on target-region data, and with transfer learning, where the model is first pretrained on the source domain and then fine-tuned on the target domain. The results are compared with simple baseline models, including a lagged-target naive model and a regional mean benchmark.
The findings show that transfer learning substantially improves predictive accuracy. Compared with the target-only CatBoost model, the transfer learning model reduces RMSE from 4.0421 to 3.0426 and MAE from 3.3504 to 2.3369. also improves from − 5.8205 to − 0.0761. These results suggest that transfer learning can help stabilize machine learning forecasts when target-domain observations are limited.