Kamakura Corporation reported Monday that it has made major advances in credit modeling accuracy in Version 4.1 of its Kamakura Risk Information Services (KRIS) default probabilities and correlations, which were released to clients today.
In the Technical Guide accompanying Version 4.1, Kamakura reports the highest listed company default model accuracy figures ever publicly disclosed by a credit model vendor. The accuracy measures are required under the "Basel II" capital accords announced in 2004 by the Basel Committee on Banking Supervision. KRIS Version 4.1 includes two new models: the fourth generation Jarrow-Chava reduced form default probability model and the fourth generation Jarrow-Merton hybrid model, which combines the explanatory variables of the Jarrow-Chava model with Kamakura's Merton-based default probabilities. The KRIS third generation Jarrow-Chava model has been very popular with clients and is being retained in KRIS 4.1 for existing clients of the KRIS Service.
The KRIS default probability and correlation service is unique in its multiple models approach and its continuous improvement strategy. Kamakura releases annual versions of its default probability models in a manner parallel to the annual releases of new versions of its Kamakura Risk Manager integrated risk management software package.
"Kamakura is committed to the continuous innovation that is a natural outgrowth of accumulating more and more credit market data," said Warren Sherman, Kamakura President and Chief Operating Officer. "We find that with each year of additional data, more explanatory variables reveal themselves as key drivers of public company defaults. The annual release cycle for the KRIS default probabilities and correlations institutionalizes the kind of progress our clients say is essential. World-wide, major financial institutions and corporations have consistently told us that they don’t believe the argument that 30-year old default probability technology is still the state of the art. The results of our model tests and recent papers from researchers at Harvard and University of Michigan prove this view is correct."
Accuracy in credit modeling is measured using a standard measure called the ROC accuracy ratio, a measure dominant in the field of statistics for more than 50 years. The ROC accuracy ratio is 100% for a model that perfectly predicts default and 50% for a model with no predictive power at all. The ROC accuracy ratio for the fourth generation Jarrow-Chava model from Kamakura is 95.54% for a data set that spans 1.42 million observations on all public companies, both corporates and financial institutions, for the U.S. and Canada between January 1990 and October 2004. The ROC accuracy ratio for the fourth generation Jarrow-Merton hybrid model was reported in the KRIS Technical Guide as 95.45%.
"These ratios are far higher than the 90.50% accuracy reported in March, 2005 by a major rating agency on a sample over a similar time span but which was much easier to model—the largest firms in the universe. Regulators world wide have said to us, We are delighted that competition has come to the default probability business and that the basis for the competition is accuracy. For that reason, Kamakura continues to offer any national regulatory agency the right to audit Kamakura’s test results subject only to a signed confidentiality agreement," added Kamakura founder and chief executive officer Dr. Donald van Deventer.