Dear Mr. Buffett_ What an Investor Learns 1,269 Miles From Wall Street - Janet M. Tavakoli [14]
If you play with coins or dice, you can learn a lot about the outcomes by flipping and throwing them thousands of times and recording the results. A Monte Carlo simulation uses a computer to throw a whole lot of random inputs into a model. It is like shaking a newly made chair to see how stable it is. Financial firms use correlation models to look at what happens when corporations default.The model tries to determine if other companies will behave similarly when one company strengthens or weakens. The models are highly unstable. They are like a chair that collapses beneath you as soon as you sit on it. Small changes to model inputs result in huge changes to the results.
If you play with coins or dice, you know exactly what your inputs are and you can model all potential outcomes. You can examine the coins (heads or tails per coin), and you can model all of the possible outcomes.You can examine dice (one to six dots on each face of each cube), and again, a mathematical model can describe all potential outcomes. We do not have to guess at the inputs for dice and cards; they are known in advance and the relationship between the inputs does not change, even though we may use a Monte Carlo model to randomize the inputs (the flips and tosses).
The inputs to credit models are a bit of a guess, since we rely on data approximations to come up with the inputs in the first place. Furthermore, the relationships between the inputs can change. Most of the data describing how one corporation behaves in relationship to another is based on market prices such as stock prices or the prices of credit default swaps based on corporate debt. Moreover, there is very little of this already-suspect data to work with. The results are guesses about relative price or yield spread movements, which result in a guess about the correlations. When a credit upset occurs in a financial sector, correlations that were previously fractional numbers tend to converge to one. Everything seems to fall apart at once. A model will calculate the wrong answer to nine decimal places, but it cannot tell you it is the wrong answer.
The biggest problem with the models is that even if they temporarily get the correct answer, they do not tell you what you need to know. Wall Street estimates asset correlations instead of the necessary default correlations. Furthermore, the overwhelming flaw in the methodology is that if you want to make up a default correlation between two companies, you must make the false assumption that default probability does not vary, but of course it does. Even if the models measured the default probability of individual companies—and they do not—if a company defaults, you still have to guess the recovery rate, the amount left over, if any, after all obligations are paid. You cannot solve for two independent items of information from a single piece of information such as a letter grade or a price. You cannot get both the probability that a company will default and the amount of money you will have left if it does default.
Warren warns Wall Street it is about to get into a fight it cannot win, and Wall Street comes anyway. The models are incapable of generating the information Warren has in his head. Warren says he doesn’t use a model, but he does. Warren himself is the model. He has spent so much time reading annual reports that he has a good idea of whether or not a company will default, and if it does, how much will be leftover after everyone else is paid off. Moreover, he knows