Global Computer-Fueled Economic Crash

[Dave]

Axioms, downturns, and a global (computer?) crash

One of the oldest and best acronyms in computing is GIGO—"garbage in, garbage out"—and it sums up a truth about the world of automated number crunching that is often forgotten by both programmers and end users. The GIGO principle means that no matter how detailed and accurate your computational model is, the results are always dependent on the quality of your input. A well-crafted model that is supplied with a credible, realistic set of inputs can greatly improve the quality of your forecasting and decision-making. But when you start feeding that same model biased data and outright lies, the results can be catastrophic. Just ask Wall Street.

When bankers do models

Right now, there are no shortage of attempts under way to blame our current economic woes on this or that party or factor or institution—and, hey, why not? There's plenty of blame to go around. But when the history of this epoch-making period in capitalism is written, there will definitely be a few chapters devoted to computers and the people who ran them.

Two very brief first drafts of that history have already come out, one of them a New York Times article from earlier this month. In the piece, journalist Stephen Lohr looks at the role of financial engineering in the crisis and concludes that the computer-driven risk assessment models that were used to value many of the complex securities at the root of the current crisis weren't necessarily badly designed. Rather, the money men took these elaborate tools—all of them designed by a class of math and physics PhDs known as "quants"—and loaded them with bad data built on faulty assumptions.

Perhaps the downturn's ur-assumption—the mistaken notion at the root of so much of the world's present pain—was that the value of US real estate would continue to rise indefinitely. That single, very faulty assumption was used as input for model after model, and the resulting output was sliced, diced, baked, and served up to investors from Los Angeles to the Ukraine.

In this respect, you can think of the present crisis as the fruit of Wall Street's (willful?) forgetfulness of two important truisms: the aforementioned GIGO principle and the famous economic dictum, attributed variously to Lord Acton and Herbert Stein, that "things that can't go on forever, don't."

When models do bankers
The second writer to recently tackle the role of computer models in the downturn is Barry Ritholtz, proprietor of The Big Picture blog and one of the handful of macroeconomists who's now getting plenty of well-deserved attention for having publicly and loudly called the downturn well in advance. Ritholtz's new article in Scientific American explains that there were more faulty assumptions at work than just the idea that the real estate boom was permanent. (As another economist pointed out recently, whenever everyone thinks that a particular asset class's value will continue to rise indefinitely, that's called a "bubble.")

Ritholtz writes, the following about the way in which the quants built their models:

"As Benoit Mandelbrot, the fractal pioneer who is a longtime critic of mainstream financial theory, wrote in Scientific American in 1999, established modeling techniques presume falsely that radically large market shifts are unlikely and that all price changes are statistically independent; today's fluctuations have nothing to do with tomorrow's—and one bank's portfolio is unrelated to the next's. Here is where reality and rocket science diverge."

The bit about the models' presumptions about the unlikelihood of radically large market shifts echoes the thinking of Nasim Taleb, another prominent "permabear" who has greatly enhanced his fortune and his reputation because of his accuracy and profitability in forecasting the recent fall.

A global hard reset?
Taleb also has another major problem with computers that has nothing to do with their uses and abuses in banking. From Taleb's perspective, computers have made the whole of the modern economy too complex and too efficient. From inventory management systems that ensure that retail outlets hold the optimal amount of inventory (no less and no more) for a given day and location, to the massive options pricing machines that time trades with millisecond precision, the entirety of the computer-driven global economy is like one massive model that was assembled—most of it over the course of the past decade—on the governing assumption that the future would look pretty much like the past. And when that widely shared assumption breaks down, then the system ceases to behave in a predictable way, because it has been too finely tuned to operate under a set of parameters that no longer pertain. (Or so the argument goes.)

In computer science terms, you could say that both Taleb and Mandelbrot, in a recent and very scary interview with Charlie Rose, have essentially argued that the current global system is in an "undefined state." This means that there's no way to predict what its output will be, which is why attempts by governments to meddle massively with the inputs will definitely have some kind of impact, but nobody can say what it really is. Government intervention becomes the equivalent of "percussive maintenance," i.e., beating on the side of the machine on the chance that you'll magically unbreak it.

If Ritholtz, Taleb, Mandelbrot, and the rest of the computer modeling and financial engineering naysayers are correct about the big picture, then we really are arguably in the midst a bona fide computer crash. Not an individual computer crash, of course, but a computer crash in the sense of Sun Microsystems' erstwhile marketing slogan, "the network is the computer." That is, we have all of these machines in different sectors of the economy, and we've networked all of them together either directly (via an actual network) or indirectly (by using the collective "output" of machines in one sector as input for the machines in another sector), and like any other computer system the whole thing hums along nicely... up until the point when it doesn't.

It may be that we should fundamentally rethink our pervasive reliance on these machines that we've come to fetishize, especially if we wind up with a hard reset. It's not that they aren't very effective tools, but rather that they're so devastatingly effective, especially they're networked together and Metcalfe's Law kicks in to multiply not only our intelligence, but our collective human frailty.

I found this article quite enlightening, particularly how the total networking of many systems caused the entire thing to crash, when one part of the system is fed bad data.

[Darth Wong]

It's basically the electronic version of the good old herd mentality. You have a processing algorithm which accounts for everyday circumstances but not extraordinary ones, and you have a feedback mechanism whereby a lot of computers doing the same thing will trigger other computers to follow. And of course, you have no intelligent leadership: all features of the herd mentality.

[Count Chocula]

If I remember correctly, about 50% of NASDAQ trades are program trades, i.e. automated trades with no people in the middle. These programs are very sophisticated, and many are designed to adapt to changing market conditions by monitoring trends (typically intra-day) and adjusting buy and sell parameters accordingly. The kicker is that there's an even chance that these algorithms are responding to other program trades, rather than reasoned human decisions.

This often results in "piling-on" activity, where a few blocks of trading establish an up or down trend in a particular sector or individual stock, then the programs come on-line and exaggerate the up or down motion.

All of these trading programs are built on algorithms devised by people who design the systems to fit into certain parameters, based on the programmers' expectations. When something outside the box comes up, wacky things - like limit up or limit down days - happen.

If you are a large institutional investor, like Merrill Lynch or BofA, I imagine there are people on staff who know this, and use it to swing trades their way courtesy of the program trading.

[aerius]

Not only that, most of the idiots who run these programs don't even bother running a "stress test" scenario to see what would happen if things don't work out as well as expected. What would the defaults be like, who would take a loss, how much, they never planned out any of that stuff. Now they're all getting forcibly liquidated and bankrupted for their lack of planning.

[aerius]

If you are a large institutional investor, like Merrill Lynch or BofA, I imagine there are people on staff who know this, and use it to swing trades their way courtesy of the program trading.

Yup, it's called the "2 o'clock pump". The market will trend upward around 2pm for no reason at all on most days. If you're a big trader you can make a fortune on this just by sitting in front of your computer from noon to 3pm every day.

[Count Chocula]

I'm afraid to ask this, but I will:

Do institutional "investors" ever figure out default risks or potential losses on derivative positions? It's been a few years, but the functions (LOL) used to calculate derivative odds all assumed liquid markets - I never saw any routines that worked where there were breaks in continuity.

If not, how the hell do people with 4 years or more of higher education fail to take failure risk into account? Is it ignorance, or willful blindness?

EDIT: I'm talking about the guys in the suits, not the programmers.

[Starglider]

Not only that, most of the idiots who run these programs don't even bother running a "stress test" scenario to see what would happen if things don't work out as well as expected. What would the defaults be like, who would take a loss, how much, they never planned out any of that stuff. Now they're all getting forcibly liquidated and bankrupted for their lack of planning.

Time-to-market pressure in financial software is insane (or at least, it used to be). The primary concern was the fear of falling behind in the constant race to squeeze out more margin and find and exploit weaknesses in the competitor's models. Testing is mostly about making sure there are no 'exploits' other banks could use. Exhaustive testing or verification of software like that takes months if not years, by which time it would be obsolete. A decent reusable specialised test harness would cut that a lot, maybe by an order of magnitude, but it would be a major investment to build and maintain. Up until recently no one could justify expending scarce and expensive expert programmer time on something that didn't seem to be needed.

You know maybe it is time we went back to the financial sector with our product, its automated verification and test suite generation capabilities may suddenly be in great demand. Thanks for the idea.

[Illuminatus Primus]

Nassim Taleb sums up what I think about quantitative risk management tools.

[J]

The absurdity of structured finance models was brilliantly summed up in a recent Dilbert strip.

[Uraniun235]

Couldn't this happen even without computers though? You could send a bunch of guys out onto the floor each day with instructions that are basically "if x happens then trade y, else do nothing", and hand out instructions according to whatever model you're subscribing to at the time. Is this really an issue with relying on computers or is it just an issue of relying on unrealistic theories?

[Sea Skimmer]

Couldn't this happen even without computers though? You could send a bunch of guys out onto the floor each day with instructions that are basically "if x happens then trade y, else do nothing", and hand out instructions according to whatever model you're subscribing to at the time. Is this really an issue with relying on computers or is it just an issue of relying on unrealistic theories?

Yeah it could, but when you start involving computes and huge numbers of simulations, it becomes harder for the CEOs and other not super computer savvy people running the companies to reject the advice they just paid millions of dollars for.