2. Why does the portfolio’s structure matter? 2.1. Statistical issues linked with the portfolio risk heterogeneity The most straightforward concern is with statistical issues. Credit risk is one of the areas in economic and financial empirical research where the paucity of data is most problematic. Default events, relative to the total number of credit loans, remain rare. At a bank portfolio and sub-segment level, it is very unlikely that the biggest corporate risk class accounts for more than a few tens of thousands names. Even in the retail market where portfolios can be expected to be much larger, possibly in the millions, the calibration of high-level confidence default statistics would still require the availability of lengthy series and extensive data. Hence, it appears rather intuitive that the portfolio’s structure puts constraints on the availability of data and the accuracy of risk estimates. In this respect, Basel II requires that banks should improve the documentation of default events and the accumulation of statistical data in the future. However, besides data limitation, portfolios’ structure also matters for the statistical accuracy of internal rating estimates. Those risks evolve over time, and ratings’ migrations occur leading to the heterogeneity and instability of data pools underlying the internal rating segmentation. Moreover, rating migration means that following changes in the borrowers’ credit quality, internal ratings change or more precisely, borrowers change rating class. This can also be interpreted as shifts in the risk distribution of the bank’s portfolio reflecting macroeconomic fluctuations. One could expect that similar macroeconomic changes should be reflected similarly in the risk estimates of two rating systems benchmarked, and in this regard, this is not a concern, unless the perception of these changes by the bank and its rating system are different. Some important insights on this issue are provided by Heiftield, in RTF (2005), who suggests that the dynamics of a bank’s rating system lead to very different risk estimates. These are difficult to compare and even back test with regard to the actual defaults observed (e.g. in the case of rating systems that are sensitive to the business cycle). Thus, statistical consistency would require that comparable risk estimates are corrected for the portfolio’s heterogeneity.