Data

So far, the model has been designed and implemented, and is running smoothly and producing results. However, remember GIGO: the results can only be useful if the model risk behind the input data, ie, model parameters and position-level data, is being addressed properly. In the model we have been using, three sets of data will have to be considered in total:

• the PDs assigned to the positions;

• the EADs and LGD levels assigned to the positions; and

• the correlation parameter ρ.

In general, the very generation of these data is of a quantitative and statistical nature. It is therefore quite natural to think about the implied model risk, and to formulate criteria for validation, also using quantitative and statistical methods. However, this is not the full story. Credit portfolio models will always suffer from the problem of possible lack of representativeness. The exposures in the current portfolio have not (yet) defaulted (at least most of them) by definition. On the other hand, to estimate parameters statistically, one needs either data from actual defaults or some mechanism (in other words, another model) connecting different kinds of information (credit spread time