Quantifying credit portfolio sensitivity to asset correlations with interpretable generative neural networks

We propose a novel approach for quantifying the sensitivity of credit portfolio value-at-risk to asset correlations with the use of synthetic financial correlation matrixes generated with deep learning models. In previous work, generative adversarial networks (GANs) were employed to demonstrate the generation of plausible correlation matrixes that capture the essential characteristics observed in empirical correlation matrixes estimated on asset returns. Instead of GANs, we employ variational autoencoders (VAEs) to achieve a more interpretable latent space representation and to obtain a generator of plausible correlation matrixes by sampling the VAE’s latent space. Through our analysis, we reveal that the VAE’s latent space can be a useful tool to capture the crucial factors impacting portfolio diversification, particularly in relation to the sensitivity of credit portfolios to changes in asset correlations. A VAE trained on the historical time series of correlation matrixes is used to generate synthetic correlation matrixes that satisfy a set of expected financial properties. Our analysis provides clear indications that the capacity for realistic data augmentation provided by VAEs, combined with the ability to obtain model interpretability, can prove useful for risk management, enhancing the resilience and accuracy of models when backtesting, as past data may exhibit biases and might not contain the essential high-stress events required for evaluating diverse risk scenarios.