Thursday, August 16, 2007

How Computer Models In the Financial Markets Missed the Mark

Some of the more recent victims of the subprime mortgage mess are hedge funds called "quant funds." They use computer models to trade, and apparently their models never anticipated that the markets would do what they are now doing. Prominent among these quant funds are three Goldman Sachs sponsored funds. The people who get jobs at Goldman are often in the top 0.1% of their B-school classes, so this is a little disturbing, to say the least.

The details of how the quant fund computer models work are generally kept confidential, for commercial reasons. So we don't know what happened at any particular quant fund. But some overarching problems have emerged. First, the models didn't account for the kind of risks that we have today. Models are based on statistical analyses, which by definition use past data. If something hasn't been seen in the past, the analysis won't pick it up. Today's market conditions are unprecedented. Never before have we had such enormous derivatives markets, with exposures reaching all over the globe. Never before have we had a lending process that favored the creation of huge amounts of risky loans. But that's exactly what the mortgage backed securities market did by encouraging mortgage lenders to make no doc, adjustable rate loans to people who hadn't a prayer of repaying them. We discussed this problem in our blog at

Furthermore, never before have we had a situation where large numbers of illiquid CDOs were being offered for sale simultaneously. While the people that created these computer models try to anticipate potential market messes (through a process called "stress testing"), they can't imagine everything that might happen.

Second, the models didn't account for the kind of volatility we have today. A fundamental premise of the computer models is that there are mathematical relationships between various asset classes that can be discovered by the power of the computer and exploited for investment purposes. If this isn't true, one can forget about computer modeling and go back to picking stocks by throwing darts at pages of the business section of the newspaper or reading 10-Ks. Anyone who believes in the reliability of computer models has to assume, explicitly or implicity, that there is a limit to volatility. If the potential for volatility is infinite, then the model cannot ultimately be said to be reliable.

The problem is that the potential for volatility is infinite--or at least beyond the imaginings of the folks that created Wall Street's computer models. Volatility today is coming at us from many different directions, but two of the most important causes are: (a) the enormous amount of risk created by the plethora of really dumb subprime and other adjustable rate mortgage loans made in the last few years, (b) the inability to sell the CDOs in which those mortgages are packaged, which has forced hedge funds to sell other assets. The models apparently didn't have a way to deal with a world where incentives existed to create massive amounts of unusually risky assets. They also didn't anticipate that hedge fund investors would respond to the unanticipated crisis with an old-fashioned run on the bank, making large numbers of withdrawal requests. These requests would force asset sales that would drive down the prices of stocks, corporate bonds, commercial paper, municipal bonds and who knows what else.

Third, the models were part of a herd. Many of the quant funds use similar models. So they took similar risks. When those risks proved to be imprudent, they all got caught in the same vortex of asset liquidations. Herd animals can be easily stampeded, and now that crocodiles have appeared at the river crossing, the wildebeest are scrambling for any foothold on dry land.

Much of the problem is a result of the index fund. The index fund was one of the most revolutionary concepts in personal finance in the last century. It gave the individual investor a shot at earning the market average without needing a lot of capital or knowing much about finance. You could spend all your spare time gardening or playing canasta, and still earn respectable returns. In order to justify their increasingly infamous 2% and 20% fees, the hedge funds had to find a way to beat the index fund. Using the massive amounts of RAM and memory capacity of modern computers, they could scour enormous financial databases for any anomaly, inefficiency or squirm in asset values that could be exploited to obtain above average returns. Then, they'd employ linear and nonlinear regression analyses, and other statistical methodologies, to show that when salmon begin to clear the lowest set of fish ladders on the Columbia River, gold prices per ounce would be at least 100 times silver prices per ounce.

The ubiquity of the computer can blind us to its weaknesses. It does marvelous things, like operate 4-liter automobile engines to generate twice the horsepower the same sized engine could provide 40 years ago, fly commercial airliners (Airbus jets are almost entirely flown by computer), send space probes to distant planets, and provide a worldwide communications mechanism through which you are accessing this blog. But the computer also creates an illusion of certainty and authority. It's so capable that we begin to believe it can't be wrong. In that respect, the computer makes us stupid. People who rely a great deal on computers often lose the willingness to think for themselves.

There's an old saying among computer programmers: garbage in, garbage out. It recognizes that a computer is just a machine and will only do what it's told. If the computer is told to do something incomplete or incorrect, it will do something incomplete or incorrect. If a computerized trading model doesn't recognize all the potential anomalies that can occur in the financial markets, the computer may well eventually direct trading in the wrong direction.

Some people predicted the subprime mortgage mess before it happened. To have picked up on the budding crisis, a computer would have had to take in information about human nature, such as the reasons for increased market demand for higher risk CDOs (the desire of institutional investors for higher yielding debt securities than Treasuries that nevertheless were supposedly safe), and the resulting appetite among investment bankers and mortgage brokers for loans whose only realistic expectation for repayment was a continually rising real estate market. But these things can't be reduced to numbers. They have to be intuited from knowledge of incentives and desires.

A computer has no intuition. While its massive analytical power can crunch through tons of historical trading data, it has no ability to read the unprecedented tea leaves in the current market and intuit a new and different future. The backwards looking nature of computer modeling won't anticipate unexpected conditions like we have today. And that's the ultimate problem with computer models. Life inevitably is unprecedented. No one day is quite like any other. Sure, computer models can generate profits a lot of the time, just as one Big Mac tastes pretty much like any other Big Mac. But you can't eat Big Macs for every meal.

Markets are a process of human interaction. Like people, they will do unprecedented things. There's no computer model for mortgage and financial market professionals greedily following short term incentives while recklessly disregarding the interests of borrowers and investors. There's no computer model for how to handle the investor panic that now fuels the withdrawal requests that are forcing large-scale liquidations of any hedge fund asset that gets a bid. Even as we see the benefits of computers, we should also recognize their limitations.

We've been here before, with the Long Term Capital Management bailout in 1998. As now, unprecedented conditions in the financial markets led to the collapse of a hedge fund that was in the process of going down and taking the financial system with it. Have we learned from that experience? Or do computers continue to mesmerize us?

The Boeing 777 has computerized flight controls. But it allows the pilot to override the computer and fly the plane manually in emergency situations. If you're going to invest in a hedge fund that uses a computer model to trade, make sure it has a manual override, and a pilot with the experience, judgment, wisdom and guts to take control when the markets become turbulent.

Crime news: here's a new fraud--glass eating.

No comments: