QuarcHome
Model catalogueBack to the model index

CAViaR

Models a conditional return quantile directly using quantile regression.

How it works

CAViaR models a conditional return quantile directly. Its recursion can use lagged quantile forecasts and features of recent returns. Coefficients minimise quantile loss, which weights misses differently according to the target tail probability. A model can therefore estimate a tail threshold without first fitting a full return distribution.

Worked example

For a lower 1% return quantile of −2%, a realised return of −3% counts as one exceedance. In 250 correctly calibrated days the expected count is 2.5. Also record the size of the miss and whether violations occur in clusters. One sample count has considerable uncertainty.

Validation and limitations

Check exceedance frequency, independence and stability through volatile periods. State whether the output is expressed as a lower return quantile or as positive loss VaR. A single quantile supplies no tail mean. Compare conditional coverage.

Research sources