Is overlapping data a problem in VaR backtesting?

Question

AcadiFi · Accepted Answer

Overlapping lookback windows are normal in VaR forecasting. A model recalibrated each day will naturally use many of the same historical observations as yesterday's model. That is not automatically a backtesting flaw. The bigger concern arises when the realized loss horizon overlaps, because overlapping multi-day P/L observations are dependent and can distort statistical tests.

Suppose Meridian Market Risk estimates one-day 99 percent VaR using a 500-day rolling window. Day t and day t+1 use highly overlapping calibration samples, but each forecast is compared with the next day's nonoverlapping P/L. That structure is common. If the desk instead backtests ten-day VaR using daily rolling ten-day losses, adjacent realized losses share nine days and cannot be treated as independent exceptions.

The FRM lesson is to identify what overlaps. Calibration data overlap is expected. Evaluation-period overlap is more delicate. Backtesting tests the model as it would have been used at the time, based only on information then available. The statistical interpretation of exceptions depends on whether the realized outcomes are independent enough for the chosen test. On the exam, also state the horizon, confidence level, data source, modeling assumption, and validation step, because those choices usually explain why two reasonable VaR answers differ materially. Practice more in our FRM question bank.

Is overlapping data a problem in VaR backtesting?

Master Part II with our FRM Course

Related Questions

Practice Questions