Combining deep learning with econometric models: volatility forecasting using the KAN-GARCH-MIDAS framework

Machine learning and deep learning are increasingly applied in finance, yet few studies explore how they can enhance traditional econometric models. This study proposes an innovative KAN-GM model, integrating the Kolmogorov–Arnold network (KAN) with the GARCH-MIDAS model to extract nonlinear macroeconomic features for volatility forecasting. Empirical results show that KAN-GM outperforms traditional GARCH in MAE and MedAE, consistently ranks in the optimal model set via MCS tests, and demonstrates strong cross-market adaptability (stocks and forex). It also maintains robustness pre- and post-COVID-19. The model effectively combines deep learning and econometrics, improving financial risk prediction.