Quantifying and minimizing the uncertainty in tail estimates
An estimate of capital requirements which is simply based on a single 'best guess' number e.g. "my 99.5% 1-year VaR is £1,000m" provides limited information as it says nothing about the statistical error around this estimate. What if the ESG is run under a different seed and the number changes to £1,100m or £900m? Furthermore, it is an estimate of a single point on the distribution (the 99.5th percentile). What about the 99.9th? Or the 99.5th CVaR? This note outlines a methodology (extreme value theory) for answering these questions.
