QuarcHome
Model catalogueBack to the model index

Expected Shortfall backtesting

Evaluates the accuracy of tail forecasts with test requirements matched to the forecast outputs.

How it works

An Expected Shortfall forecast describes the average loss in a specified tail. Backtesting can use joint forecasts of VaR and ES, or another test with explicit requirements on the forecasted distribution. The test and the available forecast outputs must match. Tail magnitude and frequency both matter.

Worked example

Suppose a daily 95% VaR forecast is 5 and the ES forecast is 8. In two samples the same number of days exceed 5, but the first sample averages 7 on those days and the second averages 15. Frequency alone cannot distinguish the tail severity. A formal joint test also accounts for the alignment of the realised losses, quantile forecasts and conditional scenarios.

Validation and limitations

Record the confidence level, whether forecasts are conditional and the profit-and-loss definition used in validation. Sparse tails and clustered losses limit power. Inspect contributions from large observations and avoid interpreting a failure to reject as proof of accuracy.

Research sources