Estimating Value at Risk for Central Counterparties: A Generative AI Approach

Central counterparties (CCPs) play an important role in the stability of financial markets by helping to mitigate systemic risk. It is important that CCPs have robust risk management tools in order to protect themselves from failing due to a defaulting member. In this paper a new model using a Bidirectional Generative Adversarial Network (BiGAN), a type of generative AI model, is proposed to estimate Value at Risk (VaR), a common metric used to compute initial margin for CCPs. Fifteen years of closing prices from the S&P 500 index was used as the dataset. The model was backtested for a period of four years. The results were evaluated using the number of breaches, Kupiec test, excess margin as well as two procyclicality measures: the standard deviation and peak to trough ratio of margin. The results were compared against three widely used models: filtered historical simulation method, historical VaR and parametric VaR. The results from this study showed that VaR computed using the BiGAN model produced 19 breaches on average for the four-year test period. While the experimental results show the proposed model is comparable with other models in terms of accuracy, its standard deviation for margin calls is lower which results in more short-term stability and a lower excess margin compared to the traditional models. The results of this research encourage further research on using BiGAN to estimate VAR.