Some Empirical Results on the Merton Model

Eric Falkenstein

Andrew Boral

12/18/00

Moody's Risk Management Services

Published in April 2001 Risk Professional

            For quantitative risk professionals, the Merton model of default provides a great segue into a previously qualitative and subjective area.   Using the Merton framework one can simultaneous address both individual default risk, and then, using the same model, work this information into a portfolio algorithm.  Elegant and efficient!  To put this into perspective, remember that loan underwriting and junk bond analysis is still a highly narrative field, where 'the story' is often more important than any raw numbers. 

This article examines the Merton model, commenting on its power for predicting default, and its sensitivity to refinements. Moody's uniquely large dataset of defaulting companies gives us a glimpse of relationships that can only be conjectured. Unlike other empirical reports which document anecdotal, qualitative, or pricing results, this is a test of the Merton model at distinguishing between future defaulters vs. nondefaulters.  Our findings are as follows: the model is a powerful measure of default risk, refinements add only marginally to its power, and outside-the-box augmentations make it significantly better.

            What is the Merton Model?

            What makes the Merton model so attractive to quants is that it is so totally familiar to them in another context.  The 'Merton Model' in credit circles refers to a model that treats equity as an option on the firm's assets.  Simply consider equity a call option, the total liabilities the strike price, and the value of the firm's assets the value of the underlying asset.  This makes sense theoretically because equity is a residual: equity holders get the upside if things go well (e.g. Microsoft), but nothing if things go poorly (Pets.com).

            Given the option framework, one can work backwards from observable information and calculate a default metric.  Specifically, take the market value of equity (E), the volatility of assets (sA), the face value of liabilities (L), and the interest rate (r), and the market value of assets (MVA). Using an options approach, the market value of equity is the result of the Black-Scholes formula:

 

          

 

where

 

The problem is that we have two unknowns: MVA and sA, so we need another equation.  Luckily we have one, in that the relation between the volatility of equity and the volatility of assets is proportional to the value of assets over equity.  Specifically, the volatility of equity can be found using the function

 

Given equations (1.1) and (1.2) we can solve for MVA and sA using standard iterative techniques that come embedded within most software packages.  The final step is to then calculate a default metric from this information.  Given the market value of assets and a the value of liabilities, and assuming that a default occurs if MVA<L, then the number of standard deviations to default is expressed as:

      

This final formula is a practitioner refinement to the original Merton model, replacing the lognormal assumption with a normal approximation.  There are other refinements made to interest rate, the time horizon, the point at which default really occurs, that are not within the Merton model explicitly, but well within the spirit.  Space constrains our ability to explain these here, but they are relatively straightforward adjustments that most on-the-ground practitioners are well aware of. 

Testing Preliminaries

            Before jumping in to testing results, we need to define our testing metric.  Given that most people do not have a sufficient number of defaults to ever do these types of tests, it is understandable that this is not an entirely familiar performance metric.  The tests below use Moody's proprietary U.S. dataset of 14,500 nonfinancial firms from 1980-2000, and 1,450 defaulting companies.

Figure 1

Power and Accuracy Ratios

 

The power curve itself can be defined as follows.  It maps the fraction of all companies with the worst score (horizontal axis) onto the fraction of defaulting companies within that group (vertical axis).  Ideally, if the sample contained 10% defaulters, then a perfect model would exclude all those defaulters at 10% of the sample excluded: the lowest ranked companies would include all the defaulting companies. Purely uninformative and highly informative models are illustrated above.

In reality defaulters will not be perfectly discriminated, creating a curvilinear function: at 10% of the sample excluded 30% of the defaulters would be excluded, at 20% of the sample 40% of the defaulters would be excluded, etc.  This creates a line that is bowed out towards the upper left (Northwest) of the chart: the greater the bow, the greater the area underneath, the better the model. 

By taking the ratio of the area under the curve but above the 45 degree line, and comparing to the area under the curve of a 'perfect' model, we can turn the power curve into a number, something we at Moody's call Accuracy Ratios. They are somewhat like R2's of regressions, and can only be used to compare models on identical datasets, but they provide a clear metric of power, with 0 being totally uninformative, and 1 perfectly informative. The higher the Accuracy Ratio, the better.  In our tests we use a 2 year horizon, because the Merton model does relatively better at this horizon and we wish to give it every benefit.

Testing Results

            First we compare the Merton model against another well-worn measure of corporate health.  Specifically we will use what we have found to be the best single metric of profitability, the ratio of Net Income/Assets.  At the 1 year horizon, we see that Merton model does dominate NI/A as a measure of risk.  As the standard errors on these measures are less than 0.01, the difference reported is statistically significant.  However, it is not as if the Merton framework, by itself, takes us into a new dimension of model accuracy. 

Table 1.  The Merton Model versus Naive Measures

Accuracy Ratios

US companies

1980-99, 1450 defaulting companies

Merton

0.748

Net Income / Assets

0.732

            Can we make it better?  Actually, we already have, as we noted we used a functional form that uses the difference between the value of assets and the value of liabilities in the strike price.  These and other adjustments are outlined in popular books on the implementation of the Merton model (see Caoutette, Altman, and Narayanan (1998), or Crouhy, Galai, and Mark (2001)).  Thus when we talk about the Merton model, most people refer to version 2.0, not the 1974 version. 

            The first check to see how sensitive the model is to refinements is to see how much lift is generated by refinements.  Consider the following 'sophisticated alternative'.  Instead of the assumption that default can only occur at maturity, we will instead use a barrier option approach, so that a default occurs if the value of the firm falls below the distress point  at any time prior to maturity.  The results, in Table 2 below, suggest that the Merton model does make a rather strong improvement over the naive method for integrating market information.  Yet while the barrier option refinement is in the right direction, it is sufficiently modest to be considered not worth the trouble. 

Table 2  Choosing a Functional Form

Accuracy Ratios

US companies

1980-99, 1450 defaulting companies

Merton

0.746

Barrier

0.748

            Thus the Merton model is better than other simple metrics of firm health, such as NI/A, not materially affected by refinements of barriers.  But what about an outside-the-box refinement?  Some would say this is like painting a mustache on the Mona Lisa, but this isn't art, its engineering.  Outside-the-box refinements have been going on since Ptolemy's mapped the heavens. 

 A Hybrid Approach

            A specific alternative we want to test is our best version of the Merton model, versus a hybrid approach that uses the Merton and NI/A.  Theoretically, all the needed information needed for default prediction is in the equity price and the liability structure.  A compelling argument for this view is that if equity information and the capital structure is insufficient, that would imply you could outperform an equity benchmark like the S&P500.  Many studies have demonstrated that professionals rarely outperform such benchmarks over long periods.  The rebuke 'if you're so smart why aren't you rich' really rings true for people who say equity markets are clearly and often irrational.  Yet this analogy is not so tight.  If the model is perfectly specified, there is a one-to-one mapping between market information and default probabilities (or at least, a rank-ordering of default propensities).  If the model is imperfectly specified, however, other information may help predict default rates even though it can not help predict future equity prices. 

            To add the information together, Merton and NI/A, we need to normalize these inputs so that their their impacts are equally weighted.  To do this, we converted both variables to percentiles and added them together.  A high Merton percentile means a company is many standard deviations from its default point; similarly, a high NI/A ratio is a high percentile, and also implies good credit.  Thus we simply added the two percentiles together for this simple hybrid model. 

            The result is in Table3.  Our best Merton model, adjusting for barriers, a perpetual options, interest rates, and several other items unmentioned, scored a 0.75.  Adding NI/A upped the score to 0.799.  Quite an improvement, much more than any refinements to the Merton model 2.0. 

 Table 3.   Adding Excluded Variables

Accuracy Ratios

US companies

1980-99, 1450 defaulting companies

'Best' Merton Model

0.748

 

Simple Merton and  Net Income/Assets

 

0.799

Conclusion

Criticizing a model as being 'insufficient' is usually a straw man argument, yet there does exist an energetic group who actually hold this extreme view for the Merton model: that it's not only timely and powerful, but sufficient.  While the information above is not a complete analysis, it hits on our strongest argument.  It is purely an empirical issue as to whether the Merton model, with all the refinements to interest rates, liability structure, default points, and other inside-the-box adjustments, says everything there is to say about default probabilities.  Our data suggest it doesn't.

While insufficient, the Merton model is a valuable part of any credit analysts toolkit. At Moody's we find it useful information for our traditional analysts, and also for our own hybrid quantitative default model that uses equity information and financial ratios.

References

 Crouhy, Michel, Dan Galai, and Robert Mark.  Risk Management.  McGraw-Hill,  2001.

Caoutette, John, Edward Altman, and Paul Narayanan.  Managing Credit Risk.  John Wiley & Sons, 1998.

Merton Model Merton Model Merton Model Merton Model Merton Model

Credit Risk Default Risk Credit Risk Default Risk Credit Risk Default Risk