Backtesting Risk Models

Backtesting closes the loop between forecast and outcome: each day, compare the model with realized profit and loss and count exceptions, days the forecast was exceeded. A 99 percent VaR should be breached about one day in a hundred; frequent breaches mean understatement. Kupiec’s proportion-of-failures test treats the count as a binomial draw, and Christoffersen’s independence test asks whether exceptions cluster, a sign of missed regimes. Basel codified this over a 250-business-day window, sorting exceptions into green, amber, and red zones; the red zone escalates capital and forces a return to simpler models. Done honestly, backtesting separates validation from theater, and it feeds model governance, the final node.