Christoffersen Interval Forecast Test

The Christoffersen Interval Forecast test extends VaR validation beyond simple exception counting. Named for Peter Christoffersen’s 1998 research, it checks not only whether exceptions occur at the right frequency (like the Kupiec test) but also whether they occur independently or cluster. Clustered exceptions reveal a VaR model that breaks under stress—a critical insight the Kupiec test misses.

Why clustering matters more than count alone

The Kupiec Proportion of Failures test asks: did we see the right number of exceptions? The Christoffersen test asks a harder question: were they scattered randomly, or did they pile up on consecutive days?

Consider two VaR models over 250 trading days, both showing 4 exceptions (both pass Kupiec). Model A had exceptions on days 40, 115, 180, and 240—roughly evenly spaced. Model B had exceptions on days 248, 249, 250, and 251—consecutive days. Both observed the same count, yet Model B’s clustering screams that something broke. Possibly a market regime shift: volatility spiked, correlations collapsed, or a risk factor moved in a way the model didn’t anticipate. Model A’s scattered exceptions suggest random variation, which is consistent with model soundness. Model B’s clustering suggests the model became obsolete mid-backtest and should be scrapped.

The Kupiec test treats both as equally credible. The Christoffersen test catches the difference.

The two-part test structure

Christoffersen’s test decomposes into two components:

1. Unconditional Coverage (UC) – the Kupiec test repackaged as a likelihood ratio. This part checks whether the proportion of exceptions matches the claimed confidence level. It is necessary but not sufficient.

2. Independence (Ind) – tests whether the probability of an exception on day t depends on whether day t−1 had an exception. If independence holds, exception probability is always p (e.g., 0.01 for 99% VaR). If violated, then after an exception, the next day’s exception probability is higher—clustering.

The overall test statistic combines these: LR = LR(UC) + LR(Ind), where LR stands for likelihood ratio. Under the null hypothesis (model is correct), this statistic follows a chi-squared distribution with 2 degrees of freedom. A value exceeding the critical threshold rejects the model.

The independence test in detail

Suppose you observe 4 exceptions over 250 days. Some patterns are:

• Exception on day 50, no exception on day 51 (transition: exception → no exception)
• Exception on day 180, exception on day 181 (transition: exception → exception) — a cluster
• No exception on day 100, no exception on day 101 (transition: no exception → no exception)

The independence test counts transitions between these states. If the model is sound, the probability of moving from “exception” to “exception” should equal the base exception probability p. If it’s much higher, exceptions are sticky—a hallmark of regime breakdown.

Formally, let:

• n₁₁ = count of day pairs (exception, exception)
• n₀₁ = count of day pairs (no exception, exception)
• n₁₀ = count of day pairs (exception, no exception)
• n₀₀ = count of day pairs (no exception, no exception)

The probability of an exception given yesterday was an exception is p₁ = n₁₁ / (n₁₁ + n₁₀). If independence holds, p₁ should equal p (the base rate). If p₁ » p, exceptions cluster, and the independence test flags it.

Why clustering reveals hidden model failures

During normal markets (the backtest window), a VaR model’s assumptions—normality of returns, stable correlations, flat volatility—might hold reasonably well. Exceptions are rare and scattered. But in a market crash, these assumptions shatter. Volatility spikes, correlations jump to 1 (all assets fall together), and price distributions develop fat tails. The model’s VaR forecast becomes laughably optimistic.

The 2008 financial crisis is the canonical example. Banks’ VaR models, backtested on calm markets (2003–2007), showed acceptable exception rates. But when the crisis hit, exceptions clustered: multiple days of losses far exceeding the forecast. The Kupiec test (applied retrospectively to the crisis period) still passed because the count was predictable. But the clustering—many consecutive days of extreme loss—revealed a model that had been structurally broken when it mattered most.

Christoffersen’s test, applied to pre-crisis backtests, would likely have flagged clustering if those backtests had included even a brief prior stress episode (e.g., the 1998 Russian default or 2000 tech crash). By checking independence, the test rewards models that remain calibrated across diverse market regimes.

Trade-off: power vs. specificity

The Christoffersen test is more powerful than Kupiec (more likely to detect a bad model), but at the cost of statistical power to detect specific alternatives. The independence component can be noisy if exceptions are few. For a 99% VaR over a normal year, you might see only 1–3 exceptions. With such small counts, the clustering test has limited discrimination.

This is why regulators often apply both tests:

• Kupiec for a quick check on frequency (low barrier to entry, easy to interpret).
• Christoffersen for deeper validation if the Kupiec test is passed (catches models that pass frequency but fail under stress).

Banks might also add stress testing and expected shortfall checks to the mix, tripling down on model validation.

Practical implementation and regulatory guidance

The Basel Committee recommends applying both Kupiec and Christoffersen tests in rolling windows (e.g., every quarter). Many banks compute a composite score:

• Pass both → green zone (model is sound)
• Pass Kupiec, fail Christoffersen → yellow zone (model’s frequency is right, but clustering is present; investigate)
• Fail Kupiec, pass Christoffersen → yellow zone (unlikely but possible; very noisy regime)
• Fail both → red zone (reject the model)

The yellow zones trigger investigations: has the portfolio composition changed? Have volatility regimes shifted? Is the risk factor model stale? Yellow-zone findings often lead to model refinements—adding volatility clustering (GARCH), regime-switching models, or heavier tails (Student-t instead of normal distribution).

The Dodd-Frank Act and Basel III frameworks mention both tests, though U.S. regulators have moved toward the traffic-light system (green/yellow/red) rather than insisting on one specific test. This flexibility encourages banks to use the best tools for their portfolios.

Limitations and extensions

The Christoffersen test assumes exceptions follow a simple two-state Markov chain (yesterday’s state influences today, but not further back). This is a simplification. Real market stress can show longer-term clustering: after a volatility spike, elevated volatility might persist for weeks. More sophisticated tests (e.g., duration-based tests, which measure time between exceptions) can capture longer persistence.

The test also assumes the VaR model is unchanged during the backtest. If the bank rebalances its portfolio or adjusts the model midway through, the test’s assumptions break. Rigorous backtesting requires a fixed model over the full window, otherwise the test conflates model changes with model failure.

Finally, the Christoffersen test is backward-looking. A model that passes both Kupiec and Christoffersen on historical data is not guaranteed to work on new, unseen markets. This is why banks combine backtesting with forward-looking validation: stress testing against hypothetical scenarios, expected shortfall calculations (which measure tail severity), and quarterly revalidation.

See also