Can The Classical Economic Model Improve The Performance Of Deep Learning? A GDP Forecasting Example

Model ensemble is considered as a powerful tool to deal with the overfitting to train data when Deep Learning (DL) models is applied to small size sample. With the application to GDP forecasting, we find significant overfitting to the validation set which also limit the power of model ensemble. We propose the Filtering Ensemble Method (FEM) which use the Classical Economic Model (CEM) as a benchmark to filter overfitted DL models. Results show that the FEM successfully improves the performance of DL models, and the Two-step Prediction Method (TSPM) further enhances their accuracy. Besides, regression equations confirm the possibility of overfitting of DL models on validation sets and the effectiveness of CEMs in filtering overfitted DL models. The study highlights the importance of combining DL models with CEMs in macroeconomic forecasting and suggests that incorporating economic knowledge is critical for the successful application of DL models in economics.