You know what I think about when I hear about the epic failure of all these fancy financial models that were designed to calculate risk? I think about the Atlantic Cod. These fish used to be everywhere. (Once upon a time, they were considered the cash crop of the ocean. Spanish fishing vessels would trek across the Atlantic just to fish the abundant cod off the coast of Canada.) Now the Newfoundland cod fishery is gone, yet another victim of overfishing.
The story of cod is usually told as the tragedy of trawlers. A trawler is boat designed to drag a massive net behind it. These nets are weighted, so that they cling to the bottom of the ocean floor. They sweep up everything for miles and miles. Most of the haul is trash - trawlers leave a trail of dead, unwanted fish - but they can also capture thousands of cod in a single haul. The use of radar made these trawlers even more efficient; now they knew exactly where to drop their nets. The result was a boom in caught cod: by the late 1960's, fishermen were hauling in more than 800,000 tons of cod every year.
But trawlers aren't entirely to blame. Their catch was still within the legal limits. (Cheating, of course, remained a big problem. Many fishing boats caught way too many fish, just as fraudulent lending helped implode the subprime market.) In fact, the Canadian government had been concerned about the cod population for decades. In the 1970's, the government instituted strict regulations that limited the total catch to just 16 percent of the total cod population. The tricky part, of course, was coming up with the population estimates in the first place. It's hard to know how many fish to catch if you don't know how many fish there are. But fishery scientists were confident that their sophisticated models were accurate. They had randomly selected areas of the ocean to sample and then, through the use of a complicated algorithm, arrived at their total estimate of the cod population. They predicted that the new regulations would allow the cod stock to steadily increase. Fish and the fishing industry would both thrive.
The models were all wrong. The cod population never grew. By the late 1980's, even the trawlers couldn't find cod. It was now clear that the scientists had made some grievous errors. The fishermen hadn't been catching 16 percent of the cod population; they had been catching 60 percent of the cod population. The models were off by a factor of four. "For the cod fishery," write Orrin Pilkey and Linda Pilkey-Jarvis, in their excellent book Useless Arithmetic: Why Environmental Scientists Can't Predict the Future, "as for most of earth's surface systems, whether biological or geological, the complex interaction of huge numbers of parameters make mathematical modeling on a scale of predictive accuracy that would be useful to fishers a virtual impossibility."
People love models, especially when they're big, complex and quantitative. Models make us feel safe. They take the uncertainty of the future and break it down into neat, bite-sized equations. But here's the problem with models, which is really a problem with the human mind. We become so focused on the predictions of the model - be it the cod population, or the risk of mortgage derivatives - that we stop questioning the basic assumptions of the model. (Instead, the confirmation bias seeps in and we devote way too much mental energy to proving the model true.) It's not just about black swans or random outliers. After all, there was no black swan event that triggered this most recent financial mess. There was simply an exquisite model, churning out extremely profitable predictions, that happened to be based on a false premise. Hopefully, the markets will recover quicker than the Atlantic cod.
Comments (163)
I often dislike models, but for different reasons than what you listed above. Much of my thesis work was devoted to modeling release of drug from polymer coatings -- different application, same idea.
>>People love models, especially when they're big, complex and quantitative. Models make us feel safe.
>>(Instead, the confirmation bias seeps in and we devote way too much mental energy to proving the model true.)
As you are probably aware, it's not possible to prove a model to be true or false, although evidence can be accumulated that either supports or refutes aspects of the model. Models, like any scientific theory, need testing. The real problem comes in when researchers don't ask the right scientific question when they test the model. Is the researcher looking for a model that accurately reflects an underlying mechanism (the cod model probably did) or is the researcher interested in an accurate prediction of the output (the cod model didn't)? Often we have the opposite problem: a model can be constructed such that it produces incredibly accurate predictions, but the underlying mechanism has never been demonstrated, and so parameter extraction from those sorts of models is a shaky science.
I think one of the biggest problems in current-day modeling is a failure to recognize that models must be tested. "Testing" does not mean "check whether my model predicts my output" or "verify the exact value of my parameters". "Testing" means that experimental conditions must be varied and the output must be measured (which output you measure depends on whether you're looking for an accurate mechanism or an accurate prediction). If the cod-modelers wanted to convince me of anything, they would have had to try fishing at a certain percentage for 5 years, check the population growth/decline, change the fishing percentage and check again.
Of course nobody would ever do that. But that's my point. Even though I'm in the business of modeling, and I think modeling is an extremely important scientific tool, we could all be a little more critical of the methods by which people generate models.
Let me digress with a story. I recently attended an epidemiology talk where the speaker demonstrated that her data could be fit well by four models. Three models where the same model with slightly different assumptions about the start and end points of the data. The fourth model was empirical (using a SPLINE for goodness' sake). All four models were used to estimate one parameter. The speaker -- who is a leader in the field, with many awards and publications to their name -- went so far as to conclude in writing, with a bolded, fun-colored powerpoint note, that "The use of four models confirms that this phenomenon is real." My jaw almost dropped.
I could fit any data you give me with an unlimited number of fancy models. For many of the models that I generate, I could give logical explanations of how they were constructed. I could ascribe their operators to real phenomenon. I could probably rationalize the parameter estimates. All of these activities can be performed for models that do not reflect the physics of the system or produce accurate predictions. It is to the detriment of good science that the purpose and testability of models is so widely misunderstood.
(By the way, it's ironic -- though only to me -- that you posted this today. I'm working on a manuscript that details the problems inherent in particular sorts of modeling. I haven't worked on it in more than a month, and this morning, I opened it up for what will be its final edit.)
Posted by: Rachael | October 28, 2008 10:53 AM