The Basel Traffic Light Test

Regulators need a simple, mechanical way to check whether a bank's internal Value-at-Risk (VaR) model is trustworthy enough to base capital requirements on. The Basel traffic light test does this by counting "exceptions" — days over the past 250 trading days where the bank's actual loss exceeded its predicted 99% VaR — and comparing that count against thresholds calibrated to what a well-behaved model should produce by chance alone.

A 99% VaR model should, on average, be breached about 2.5 times a year (1% of 250 days) if it's correctly calibrated. The traffic light framework sets the green zone at 0–4 exceptions (consistent with a well-specified model, differences from the expected 2.5 attributed to normal statistical noise), yellow at 5–9 (raises concern, may still be statistical noise but is unlikely under a correct model), and red at 10 or more (statistically implausible under a correct model, treated as strong evidence the model understates risk). Each zone carries a consequence: green means no action, yellow adds a graduated capital multiplier penalty that increases with the exception count, and red triggers a much larger penalty along with regulatory scrutiny of the model itself.

The zones are derived from a binomial probability calculation: under a correctly calibrated model, the number of exceptions in 250 trials with a true 1% breach probability follows a binomial distribution, and the yellow/red cutoffs are set at points where the cumulative probability of seeing that many or more exceptions purely by chance becomes small enough to be suspicious.