13. Validation of model inputs/outputs

All qualitative standards defined for the control of consistency, accuracy and reliability of data sources of VaR also apply to Stressed VaR.

Underlyings for which institutions do not have a history of data complete enough to cover the reference period, should be shocked by approximation, using closely related underlyings (same market, similar structure and characteristics). Following the same process that has been approved for the institutions’ internal models, in order to ensure the quality of historical data used for the reference period, institutions should document the methodology followed for identifying and for proxying missing data. Institutions should also perform tests of the potential impact of the use of these proxies.

With a view to preserving arbitrage inequalities, institutions may need to apply data cleaning for Stressed VaR. Where this is the case, the removal of outliers from

16 historical data series should be appropriately justified and documented, as it should not end up decreasing the magnitude of extreme events.

As Stressed VaR entails, by definition, the application of highly stressed scenarios to current market parameters, which may lead to incoherent market conditions (e.g. negative forward rates) more frequently than within a VaR computation, institutions should monitor the calibration failures that may materialise. Institutions using full revaluation when estimating their Stressed VaR may be more frequently confronted with those calibration failures than institutions not using full revaluation, not because failures will not happen, but because their methodology will not enable them to spot these calibration failures when they occur.

D. Use test