Portfolio Margining Using PCA Latent Factors

Filtered historical simulation (FHS)—a simple method of calculating Value-at-Risk that reacts quickly to changes in market volatility—is a popular method for calculating margin at central counterparties. However, FHS does not address how correlation can vary through time. Typically, in margin systems, each risk factor is filtered individually so that the computational burden increases linearly as the number of risk factors grows. We propose an alternative method that filters historical returns using latent risk factors derived from principal component analysis. We compare this method's performance with "traditional" FHS for different simulated and constructed portfolios. The proposed method performs much better when there are large changes in correlation. It also performs well when that is not the case, although some care needs to be taken with certain concentrated portfolios. At the same time, the computational requirements can be reduced significantly. Backtesting comparisons are performed using data from 2020 when markets were stressed by the COVID-19 crisis.