What are structural breaks in time series data and how do they affect risk models?

A structural break occurs when the underlying data-generating process changes at some point in time — meaning the model parameters (mean, variance, correlations) shift permanently rather than temporarily. This is distinct from cyclical variation or outliers.

Why Structural Breaks Matter for Risk

Most risk models assume parameter stability: the VaR model calibrated on 2 years of data assumes the volatility and correlation structure of that period will persist. A structural break invalidates this assumption, potentially causing:

• VaR underestimation if the break increases volatility
• Hedge failure if correlations shift
• Model backtesting failure if parameters drift

Detection: The Chow Test

The Chow test checks whether a linear model's coefficients are the same across two sub-periods. The procedure:

1. Estimate the model on the full sample (restricted) — obtain RSS_full
2. Estimate separately on each sub-period — obtain RSS_1 and RSS_2
3. Compute the F-statistic:

> F = [(RSS_full - RSS_1 - RSS_2) / k] / [(RSS_1 + RSS_2) / (n - 2k)]

Where k = number of parameters and n = total observations. A large F-statistic (above the critical value) rejects the null of no structural break.

Example: Ironwood Risk Analytics

Ironwood Risk Analytics monitors a volatility model for EUR/USD. Their GARCH(1,1) model was calibrated through 2024. In early 2025, the ECB announced a major policy shift, and realized volatility jumped from an annualized 7% to 12%.

Ironwood applies a Chow test at the policy announcement date:

• RSS_full (2023-2025): 0.0842
• RSS_pre (2023 to announcement): 0.0312
• RSS_post (announcement to 2025): 0.0285
• k = 3 parameters, n = 500 observations

F = [(0.0842 - 0.0312 - 0.0285) / 3] / [(0.0312 + 0.0285) / 494]

F = [0.00817] / [0.0001209] = 67.6

The critical F(3, 494) at 1% is approximately 3.82. The test overwhelmingly rejects parameter stability — confirming a structural break.

Practical Responses to Structural Breaks

FRM exam tip: Know the Chow test mechanics and be able to compute the F-statistic. Also understand that CUSUM (cumulative sum) tests are an alternative that does not require knowing the break date in advance. Questions often describe a scenario where backtesting failures spike after an event and ask what went wrong — the answer is usually an undetected structural break.

Join our FRM community for more time series analysis discussion.