The sign test is a non-parametric statistical test used to evaluate whether the median of the differences between two paired samples (in this case, the losses predicted by the two VaR models) is significantly different from zero. In the context of comparing VaR models, it checks whether one model consistently produces higher or lower losses compared to the other, without assuming a specific distribution for the differences.

A is incorrect: The sign test evaluates whether the median difference is significantly different from zero, not specifically the frequency of underestimation.

B is incorrect: While comparing exceptions is a legitimate aspect of VaR model validation, the sign test specifically evaluates the median loss differential between two paired observations. It does not focus on exception frequencies or rates.

D is incorrect: Exception rates are typically analyzed using backtesting or specific statistical tests designed for frequency-based evaluations. The sign test, in contrast, examines paired loss differentials and does not directly address exception rates.

Things to Remember

• The sign test is a non-parametric test, meaning it does not rely on strong distributional assumptions about the loss differentials.
• It is relatively simple to implement and provides a practical way to formalize model comparisons.