@

초록

We evaluate alternative forecasting strategies for the Korean housing market by comparing a DSGE model with several machine-learning methods in an identical empirical setting. Ridge regression yields lower MAE, MAPE, and RMSE than the DSGE model, while being considerably easier to implement. Simpler penalized linear models (LASSO and Ridge) outperform more complex machine-learning methods, including MLP and XGBoost. This pattern is consistent with the tendency of high-capacity models to overfit limited training data, which lowers their accuracy on out-of-sample data. Variable-importance rankings from the machine-learning models align closely with the structural drivers emphasized in the DSGE framework, suggesting that these data-driven methods can retain economic interpretability.

키워드