Backtesting value at risk: how Kupiec and Christoffersen tests verify a risk model
A value-at-risk figure is a probabilistic claim about future losses. The 95% one-day VaR says losses will exceed the figure on roughly 5% of days. Like any probabilistic claim, it can be tested against realised data—and a model that fails the test is not a model worth using. VaR backtesting is the formal methodology for that verification.
What VaR backtesting is
VaR backtesting compares the predicted exception frequency from a VaR model against the realised exception frequency in out-of-sample data. An exception occurs whenever the realised loss exceeds the predicted VaR threshold. A 95% one-day VaR over 250 trading days should produce roughly 12.5 exceptions; the question backtesting addresses is whether the actual exception count is statistically consistent with that expectation.
The two foundational tests are Kupiec's POF (proportion of failures) test (1995) and Christoffersen's conditional coverage test (1998). Kupiec tests only the unconditional exception frequency; Christoffersen extends the analysis to test whether exceptions are independent across time, capturing models that produce roughly the right average exception count but cluster their exceptions in stress periods.
How it works
The Kupiec POF test computes a likelihood ratio comparing the observed exception count against the model's expected count. Under the null hypothesis (model is correctly calibrated), the test statistic follows a chi-squared distribution with one degree of freedom. Rejecting the null—typically at the 5% significance level—means the realised exception frequency is statistically inconsistent with the model's claim.
Christoffersen's extension adds an independence test. A model can pass Kupiec (right average exception count) while clustering exceptions on consecutive days during a stress regime—a sign that the model fails to update fast enough during volatility shifts. The conditional coverage test combines Kupiec and the independence test into a joint statistic.
What the evidence shows
Empirical applications of Kupiec and Christoffersen tests to bank trading-book VaR models reveal that simple historical-simulation VaR often passes Kupiec but fails the independence component during stress periods (Berkowitz, Christoffersen & Pelletier, 2011). More-sophisticated models (filtered historical simulation, GARCH-based) tend to perform better on both criteria.
The Basel framework requires regulated banks to backtest their internal VaR models and adjusts capital requirements based on the number of exceptions in the prior 250 trading days. The framework's traffic-light system—green, yellow, red zones—is a direct application of the Kupiec methodology to regulatory practice.
Limitations and trade-offs
VaR backtesting requires substantial out-of-sample data to deliver statistical power. With only 250 daily observations, the test cannot reliably distinguish between a 95% VaR with 5% true exception rate and one with a 4% or 6% true exception rate. Practical backtesting therefore uses longer histories or higher-frequency data where available.
The methodology also tests only what it tests: VaR at a specified confidence level. A model that is well-calibrated at 95% can be poorly calibrated at 99%—and tail-risk applications often care more about the higher confidence level than the headline 95%. Tail-VaR backtesting (testing exceptional losses given an exception) extends the methodology and is increasingly standard in institutional applications.
VaR backtesting in pfolio
Formal VaR backtesting (Kupiec, Christoffersen) is not currently performed in pfolio Insights. The platform reports VaR and expected shortfall at standard confidence levels using the historical return distribution; investors who want to test exception frequency can do so externally using the underlying return series.
Related articles
Disclaimer
Get started now
