AI-Powered Reduced-Form Model for Default Rate Forecasting

This study aims to combine deep and recurrent neural networks with a reduced-form portfolio model to predict future default rates across economic sectors. The industry-specific forecasts for Italian default rates produced with the proposed approach demonstrate its effectiveness, achieving significant levels of explained variance. The results obtained show that enhancing a reduced-form model by integrating it with neural networks is possible and practical for multivariate forecasting of future default frequencies. In our analysis, we utilize the recently proposed RecessionRisk+, a reduced-form latent-factor model developed for default and recession risk management applications as an improvement of the well-known CreditRisk+ model. The model has been empirically verified to exhibit some predictive power concerning future default rates. However, the theoretical framework underlying the model does not provide the elements necessary to define a proper estimator for forecasting the target default rates, leaving space for the application of a neural network framework to retrieve the latent information useful for default rate forecasting purposes. Among the neural network models tested in combination with RecessionRisk+, the best results are obtained with shallow LSTM networks.