Slashdot Mirror
The Formula That Killed Wall Street
We recently discussed the perspective that the harrowing of Wall Street was caused by over-reliance on computer models that produced a single number to characterize risk. Wired has a piece profiling David X. Li, the quant behind the formula that enabled the creation of such simple risk models. "For five years, Li's formula, known as a Gaussian copula function, looked like an unambiguously positive breakthrough, a piece of financial technology that allowed hugely complex risks to be modeled with more ease and accuracy than ever before. With his brilliant spark of mathematical legerdemain, Li made it possible for traders to sell vast quantities of new securities, expanding financial markets to unimaginable levels. His method was adopted by everybody from bond investors and Wall Street banks to ratings agencies and regulators. ... [T]he real danger was created not because any given trader adopted it but because every trader did. In financial markets, everybody doing the same thing is the classic recipe for a bubble and inevitable bust."
In financial markets, everybody doing the same thing is the classic recipe for a bubble and inevitable bust.
Citation? Booms and busts are caused by, respectively, expansion and contraction of the money supply (usually in the form of bank credit), often accompanied by manipulated interest rates. The formulas used by lots of investing firms could cause clusters of errors, but the extent of types of companies (and governments) affected points to a more Austrian-style, systemic boom/bust rather than a single-(important-)sector miscalculation.
That Gaussian curves are a poor model for unlikely events has been known for quite some time. This is best explained by Nassim Taleb in the following books:
His main thesis is that the markets are essentially random and are basically impossible to predict in any meaningful way. Further there are unlikely unknown unknowns can cannot be predicted until the they occur, usually with disastrous consequences.
---- It won't be as bad as you fear or as good as you hope, but it will take twice as long as you plan.
Is global warming the new replacement for Godwin's Law?
Do you even lift?
These aren't the 'roids you're looking for.
it does not matter what model you use. Apparently they all created virtual worlds in big numbers (total value of derivatives and such is few times more than summed up gross domestic product of all countries on our planet) - this had to crash independently of the model - problem being that they used the same one. in other words: if all sheeple use the same model of reality then to make profit you need to use different one. Or to say it yet differently: if all sheeple do the same they create the bubble. nature of bubbles is that they burst when they reach physical limits of the stuff of which they are made. In our case it was human gullibility.
>>There is nothing wrong with using a model. Models are good.
Not in economics, they're not. The book Black Swan, which should be read by anyone interested in this topic, says that the hideous lie is that people claim that "they're better than nothing", when, in fact, they're worse than not having any model at all.
The LTC crash was caused by the founders (Nobel Laureates in Economics) having a model to quantify risk. IIRC, they used some sort of guassian model, taking the standard deviation of price movement as "risk". (http://en.wikipedia.org/wiki/Black-Scholes#Black.E2.80.93Scholes_model) This of course looked good until, quite suddenly, it wasn't and there was an event that their model predicted shouldn't have happened within the lifetime of the universe (that's the problem with using gaussians instead of cauchy curves or other fat-tailed distributions) and the company crashed and burned, and did a lot of collateral damage as well.
From the wikipedia article on LTC (http://en.wikipedia.org/wiki/Long-Term_Capital_Management): Merrill Lynch observed in its annual reports that mathematical risk models, "may provide a greater sense of security than warranted; therefore, reliance on these models should be limited."
yes, i completely agree with you. the focus on quarterly earnings is representative of "short-termism" everywhere, which is usually detrimental to long term value preservation.
i guess what i should have said is that greed is not going anywhere. harness it when you can and don't be surprised when it causes people to do things that harm others.
In China, they're using this slack time to upgrade the infrastructure, closing down old inefficient factories and building new ones with government CASH. Who's winning this round?
Not the millions of migrant chinese workers who have lost their jobs, which will probably also cause civil unrest. Also, the Chinese holding trillions of dollars in U.S. treasuries will also be slightly annoyed when the U.S. government inflates away their debts.
Finally, the vast majority of China's stimulus package was already announced before this major recession. You have the order backwards.
There's nothing advanced or innovative about a gaussian copula. It's a very simple mathematical trick, it doesn't say anything about finance in itself. It's a programming trick to go from a uniform distribution on a cube (easy to generate, run rnd() for each coordinate) to a multivariate gaussian with a specific covariance matrix. The way to do it is cholesky decomposition. This is OLD stuff.
Li's paper is a clever way to measure default correlation using correlation matrixes from asset returns. It's quite clever, and yes it's a pretty good model (more on that later)
This is not journalism, this is a bit of shit where the author decided having an "evil formula" would be cool. Look there's an "equal" sign, how can they be so sure... pffffffffffffffff.
I said it was a good model, yet it's been proven wrong hasn't it? Well, first of all, what has been shown to be wrong is the guesstimate of correlation that was input into the model. G.I.G.O
Plus, if you price a fixed income product and it produces higher than market return, you will borrow short term funds to invest them in it. In a free market that quickly drains the pool of saving and raises short term interest rate. Sure you end up losing money but no catastrophe. In a federal reserve system, well the short term rate stays what the fed says it should be and everyone piles on the arbitrage, creating sky high leveraged position.
Yeah the formula can be misleading, but for a true catastrophe, you need a federal reserve.
\u262D = \u5350
>>>they all created virtual worlds in big numbers - [the real world] had to crash independently of the model
Maybe we should invent a game for these bankers. World of Real Estate - where the goal is to get as many poor people into as many houses as possible, without investors learning the real housing value is only half the retail value. That way they can watch their virtual bubble go "boom" without affecting the rest of us in the real world.
"I disapprove of what you say, but I will defend to the death your right to say it." - historian Evelyn Beatrice Hall
Exactly. EVERY model that only sees rising house prices during it's data collection phase WILL assume that house prices will keep rising, and therefore tell bankers that dodgy mortgages are ok.
After all, as long as house prices keep rising, there is NO risk whatsoever in dodgy mortgages. Either you get the stated intrest (buyer pays mortgage) or you get the price rise of the house since the buyer bought it with your money (in the case of default) ... the risk of losing money in the deal is EXACTLY the chance that house prices drop. And house prices never dropped (significantly) in over 50 years ... obviously any statistical algorithm would have told you the risk was minimal.
Even more important that the limitations of a model are the assumtions taken in developing the model and/or feeding the data into the model, these should always be made clear to whomever the user of the model is, and it is then up to the user to decide if those assumtions are reasonable for their use of it.
The problem with this is most people's "just give me what I need to get the job done today" attitude. I've taught statistics in community college for a number of years, and I grapple with this a lot. Difficult enough to get people to perform the calculations for z-interval/test. Almost impossible to get them to consider the meta-analysis on whether the test is legitimate (simple random sample, assessment of normal population if sample size small, known standard deviation, etc.)
If most days they can get away with ignoring the model's assumptions, then folks wind up doing so, and then that knowledge degenerates. Ultimately the exceptional day that they need that skill, they don't have it. People function very, very poorly in relation to very infrequent (once a generation?), catastrophic events.
We know where leadership by an anti-intellectual "strongman" who scapegoats minorities and likes boisterous rallies goes