The Formula That Guessed How Houses Fall Together

February, two thousand nine. Wired magazine, front cover, a headline that still makes quants flinch. Recipe for Disaster. The Formula That Killed Wall Street.

The formula. Singular. One equation took down the global economy?

That's the claim the journalist Felix Salmon made. And the man at the center of it wasn't a trader or a CEO. He was a quantitative analyst. David X. Li. In two thousand, he published a paper introducing the Gaussian copula.

Which means what, in human?

Salmon put it beautifully on Marketplace. Li was trying to look at lots of different bonds and work out whether they were all moving in the same direction or not. That's it. Does your neighbor's mortgage going bad tell you anything about yours?

And obviously it does. If the factory in town closes, everybody's late on the fifteenth.

Hold that thought. Because inside Li's formula sits one number that answers exactly that question. Nobody on Earth could measure it. So they guessed. I'll tell you what that guess cost... but you need the magic trick first.

Magic trick.

If mortgages default independently, you can take a pool of thousands of them and slice it into layers. The bottom layer eats the first losses. The top layer only gets hurt if almost everything fails at once.

So the top layer looks bulletproof.

Triple A. Same rating as a government bond. Built out of subprime loans. That's alchemy. And the only thing standing between alchemy and arithmetic is that one input... correlation.

So where did they GET it? You can't just... vibe a number.

You sort of can. Real default data is rare. Defaults are unusual events, and the American housing market had barely ever fallen nationally. So Li took a shortcut. He used prices from the credit default swap market instead of real-world default history.

Hang on. He used market prices to estimate the risk... and then people used that risk to set market prices?

Yes. And it's worse than circular. Researchers who pulled the model apart afterwards pointed out that in a pool of a hundred and twenty-five names, you have something like seven thousand seven hundred and fifty separate pairwise relationships.

And the model hands you...

One number. Seven thousand seven hundred and fifty questions. One answer.

That's not a model, that's a wish.

And the market saw cracks early. There's a thing called the correlation smile, where different slices of the same deal imply different correlations, which is mathematically absurd. Sometimes it couldn't be calibrated at all. The papers say it worked, more or less... until early two thousand eight.

Which is roughly when the ground opened up.

Houses in Nevada, Florida, Ohio, California started defaulting together. Not independently. Together. And the low-correlation assumption wasn't slightly wrong. It was wrong in the exact direction that destroys the safest layer first.

Define destroys.

The Financial Crisis Inquiry Commission found that over ninety percent of the triple A mortgage securities issued in two thousand six and two thousand seven were downgraded to junk by two thousand eight.

Ninety percent. Of the ones rated safest on Earth.

And there were two other dials nobody talks about. How many loans default. And recovery, how many cents on the dollar you get back afterwards. The models assumed foreclosure sales would recover a healthy chunk. When the whole country is selling at once...

Recoveries collapse too. So all three assumptions failed in the same direction on the same morning.

That's model risk in one sentence.

Fine. So we hang David Li and go home.

No. And this is the part I want you to hear. Salmon never argued the math was evil. He argued Wall Street missed the assumptions underneath it. Other academics went further, and said Li was unfairly blamed for the formula that killed Wall Street. He wrote a tool. Other people chose to bet the house on it.

The formula didn't have a bonus target.

The formula didn't have a bonus target. And Salmon won the American Statistical Association's award for statistical reporting for that story, so the lesson was never anti-math. It was: know which of your numbers you guessed.

Okay, but that's two thousand eight. Is anyone still building these things?

Constantly. Different letters now. Corporate loans pooled instead of mortgages, sliced the same way. Triple A at the top, a thin equity layer at the bottom eating the first losses. Same three dials. Default rate. Recovery rate. Correlation.

And correlation is still a guess.

Unknowable in advance. By definition. You find out what your correlation assumption was worth on the day everything moves together... and that's the one day you don't get to re-run the model.

So next time someone tells me something is low risk...

Ask them what has to stay independent for that to be true. Your job. Your house price. Your pension. Your bank. It all feels separate... right up until one shock proves it never was.

That's the bit that makes it personal.

Every triple A rating ends with a clause nobody says out loud. This is safe... assuming things don't all break at once. The formula never lied to anyone. It just answered exactly the question it was asked.