The chart reports the ROC accuracy ratio for the three new models and compares them to the KRIS version 4.0 Jarrow-Chava model and to the version 4.0 implementation of the Merton model of risky debt. The ROC accuracy ratio is based on the comparison of 3 billion pairs of one defaulting observation and one non-defaulting observation. The ROC accuracy ratio is the percentage of the time that the defaulting observation is correctly ranked as more risky than the non-defaulting observation. The chart below shows the ROC accuracy ratio for all models for the accuracy of predicting default in month N, conditional on surviving through month N-1.
The graph shows that the version 5 models have a 1month accuracy ratio of 95.10. The version 5 models are dramatically more accurate than legacy ratings and Kamakuraâs version 4.1 models at all maturities, particularly at the longer maturities. In fact, the 73.13 accuracy for predicting default in month 120 conditional on surviving 119 months is higher than the Merton modelâs accuracy, 72.49, in predicting default in month 15 conditional on surviving 14 months. The reduced form Jarrow-Chava model, therefore, provides 105 months of âearlier warningâ than the Merton models at a comparable level of accuracy. The dramatic improvements in long run accuracy stem from significant suggestions from the diverse KRIS client base, including insights previously announced from First Rand Bank in Johannesburg.
âDuring the credit crisis, it became starkly apparent to many market participants that legacy Merton default technology was inconsistent with the many failures of financial institutions around the world,â reported Dr. Donald R. van Deventer, Kamakura Corporation founder and CEO. âIn a crisis, cash is king, and the reduced form framework can take advantage of this insight that is ignored in the Merton framework. By extending the maturities of the KRIS default term structures to 10 years, we can now show early warning of the next âbusiness cycle in the makingâ to KRIS clients. One of the reasons for the very high model accuracy is the distinction that FNMA and FHLMC failed and then were rescued for some but not all liability holders. The major rating agencies, contrary to credit default swap market âdefaultâ definitions, ignored these failures of AAA-rated firms in their self-assessments released in February. KRIS 5.0 is based on the recognition that these failures, and those of many other firms, are too important to ignore.â
The KRIS default probabilities are available for 29,400 public firms in 33 countries. The KRIS service also includes implied ratings and credit default swap spreads for all 29,400 firms. KRIS default probabilities are available via the web site www.kris-online.com. KRIS default probabilities are also made available via the Reuters 3000Xtra service for 2,000 firms and via file transfer protocol for KRIS âpower users.â The KRIS default probabilities are integrated seamlessly in the KRIS-cpm credit portfolio manager simulation engine embedded in KRIS and with Kamakuraâs enterprise wide risk management system, Kamakura Risk Manager. In KRIS-cpm, forward looking default probabilities are driven by econometric links to macro-economic factors. In Kamakura Risk Manager, these linkages can be supplied by either the user or from the KRIS data base. These relationships define the degree of correlation between the default probabilities and events of default for all pairs of companies.