Efficient Market Risk Assessment with Multi-Fidelity Learning: A Case Study on Bermudan Swaptions

This paper investigates the use of statistical learning models to accelerate financial risk evaluation under stringent regulatory frameworks. We first apply Gaussian Process Regression (GPR) to construct accurate surrogate models for efficiently estimating key risk measures such as Value-at-Risk (VaR) and Expected Shortfall (ES). We then propose an extension based on multi-fidelity Gaussian Process Regression (mGPR), combining highfidelity pricing data (from the LGM-1F model) with fast low-fidelity approximations. The approach is validated on portfolios of Bermudan swaptions, which are computationally intensive to value due to their early-exercise features. Numerical experiments show that GPR alone provides substantial speed-ups while maintaining the required precision, and that mGPR achieves further gains, reducing computational costs by up to a factor of four. These results offer a scalable and accurate solution for real-time risk calculations under FRTB requirements. The methodology also opens promising directions for applying multiscale learning techniques to other quantitative finance challenges.