Moody's Introduces Default Rate Forecast


By Martin S. Fridson, chief high yield strategist, Merrill Lynch

[Moody's Investors Service began issuing a monthly, year-ahead forecast of its trailing 12-month default rate. It was described in the rating agency's August Special Comment entitled "Predicting Default Rates: A Model for Moody's Issuer-Based Default Rate" by Sean Keenan, Jorge Sobehart and David Hamilton. The following is a response by Martin Fridson.]

Moody's entry into the default rate forecasting arena is a courageous move, given the criticism that is almost certain to follow.

For starters, market participants will scoff at Moody's forecasting model for failing to incorporate supposedly important factors. In addition, every time the agency forecasts a rise in the default rate (which, if the model is valid, must occur at least once in every cycle), we can expect high yield hypesters to ridicule the agency's failure to recognize the alleged irrelevance of all historical precedent. For the most part, the intellectual level of these attacks will rank with the pro forma denunciations of downgrades by companies that subsequently file for bankruptcy.

First, on the matter of the model's structure, readers should note that the Moody's model is empirically based.

Keenan, Sobehart and Hamilton have not simply expressed an opinion about how the world ought to work. Rather, they have selected explanatory variables that display a significant statistical correlation with the default rate. In the process of culling the significant variables, they rejected many others that might logically seem to influence the default rate, but had no measurable impact. These included more detailed breakdowns of the ratings mix and a variety of macroeconomic variables.

While this empirical approach provides a valuable benefit, namely, rigor, it also has the disadvantage of opening the Moody's analysts to attacks by nonrigorous propagandists who hope to impress unsophisticated onlookers. In a competition of this sort, those who follow a rigorous method are at a hopeless disadvantage to those who follow no method at all. In private discussions, the detractors can assert, with relative impunity, that Keenan, Sobehart and Hamilton failed to consider a supposedly vital determinant of default rates that they in fact tested and rejected on eminently defensible grounds.

One particularly odious version of the no-method approach is the spurious one-factor model. In default rate discussions, this usually takes the form of describing the default rate as purely a function of the level of economic activity. The Moody's study makes a useful contribution by refuting in a new way a canard previously discredited by Helwege and Kleiman (1996), among others.

Keenan, Sobehart and Hamilton show that while industrial production is useful, on the whole, as an explanatory variable, its correspondence with the default rate was quite low around 1973 and in the early 1990s (the Moody's researchers add that as the universe of high yield issuers becomes increasingly global, a measure of U.S. economic activity is becoming an increasingly imperfect proxy).

In light of such periods of false signals, a well-designed multivariate model is inherently superior to a univariate model. Even so, we fully expect self-proclaimed high yield experts to continue "explaining" to the financial press that default rates are sure rise (or fall) in the coming period in light of the prospects for economic growth.

As for the predictable condemnations of the bearish forecasts that Moody's is bound to publish from time to time, readers should consider that there are two sound ways to assess the validity of a forecast.

First, one can wait for the outcome and determine, after the fact, whether the prediction was accurate. Second, one can assess the logic underlying the forecast before the outcome is known.

In some instances, paradoxically, the latter method is sometimes more sound than the former. Consider, for example, a forecast of a tennis match's outcome, based on a "who-has-beaten-whom" method that has been statistically validated.

Suppose that of the ten players whom both Jones and Smith have been matched against during the past year, nine lost to Jones and eight beat Smith. At even odds, one would surely bet on Jones to beat Smith. If Jones contracts a severe stomach virus the day before the match and comes onto the court in a weakened condition, a victory by Smith does not invalidate a forecast of a Jones triumph, made one week earlier. Ex post, the forecast that Smith would win proves to be a better forecast, but whatever method was used to derive it is not one we would recommend. Even though it proved wrong, one can argue that the prediction of a victory by Jones was the superior forecast.