Performance testing of margin models using time series similarity

Max Wong and Patrick Ge Pei

Abstract

Risk management has traditionally used backtesting for quality assurance of its margin models, as required by regulators. Backtesting focuses on statistical unbiasedness in terms of frequency coverage and temporal independence of margin breaches. However, a model that has been perfectly backtested may be ineffective in times of crisis. In practice, a risk manager is more concerned about three main features of margin models that are not measured by backtesting: the magnitude of the margin shortfall, the rate at which margin increases (which will cause procyclical risk) and the cost-efficiency of margin usage. We suggest a “regulatory target” model in which all three requirements are fulfilled according to the utility function of the regulator. Practical margin models used by financial institutions can then be benchmarked against this model using empirical similarity.