Profit from the Bayesians
The markets do not only trade on participants' assessment of the value of securities. They also trade on participants' assessment of the beliefs of other participants about the value of securities. If you think everyone is going to sell a stock tomorrow, you'll sell it today, regardless of your assessment of its long term value, because if need be you can buy it back more cheaply the day after tomorrow.
If you can successfully determine what market participants will do, arbitrage opportunities sometimes result. At the moment there may be one concerning the quantitative equity funds. We know roughly what these funds do - they try to exploit market dislocations. The fund has an idea of what is a dislocation because it has a model of what relationships 'should' hold. And these models are often calibrated using some form of Bayesian network.
Now here's the interesting bit. It seems many of these funds have broadly the same position, or at least the same kinds of position. We suspect this because when one quant fund liquidates it appears that some of the others take a bath. The fund being liquidated had longs and shorts based on their model and when these longs are sold, causing these stocks to fall in price, and the shorts are bought back, causing them to rise, presumably the opposite behaviour to the one predicted by the model is observed. Since funds share models or at least modeling assumptions, this hurts a number of quant funds simultaneously.
So, here's the plan. First figure out what that model portfolio looks like, roughly. It should not be too hard to make a plausible guess of the basic structure of the Bayesian model some of the funds use: just run it, and get a portfolio - we'll call this portfolio A. Then figure out a portfolio that mostly is market neutral (and in particular does not lose much money on those days that the market relationships predicted by the model do hold) but makes a lot of money if portfolio A has a really bad day. You can think of this as a far out of the money put on the model's correctness. Wait for next quant fund liquidation (which is bound to come as someone is always over-leveraged) then buy a yacht. Or some kind of boat, anyway.
If you can successfully determine what market participants will do, arbitrage opportunities sometimes result. At the moment there may be one concerning the quantitative equity funds. We know roughly what these funds do - they try to exploit market dislocations. The fund has an idea of what is a dislocation because it has a model of what relationships 'should' hold. And these models are often calibrated using some form of Bayesian network.
Now here's the interesting bit. It seems many of these funds have broadly the same position, or at least the same kinds of position. We suspect this because when one quant fund liquidates it appears that some of the others take a bath. The fund being liquidated had longs and shorts based on their model and when these longs are sold, causing these stocks to fall in price, and the shorts are bought back, causing them to rise, presumably the opposite behaviour to the one predicted by the model is observed. Since funds share models or at least modeling assumptions, this hurts a number of quant funds simultaneously.
So, here's the plan. First figure out what that model portfolio looks like, roughly. It should not be too hard to make a plausible guess of the basic structure of the Bayesian model some of the funds use: just run it, and get a portfolio - we'll call this portfolio A. Then figure out a portfolio that mostly is market neutral (and in particular does not lose much money on those days that the market relationships predicted by the model do hold) but makes a lot of money if portfolio A has a really bad day. You can think of this as a far out of the money put on the model's correctness. Wait for next quant fund liquidation (which is bound to come as someone is always over-leveraged) then buy a yacht. Or some kind of boat, anyway.
Labels: Markets, Model risk
posted by David Murphy at 21:46
2 Comments:
Hmmm. Would the law of large numbers come into this? Even if, astoundingly, some of those Bayes nets modellers had discovered life outside Gaussian assumptions and used mixed distributions, intervals, fragments/plates etc, wouldn't the fact that you were modelling across a whole collection of these models give you a roughly gaussian spread across their collective behaviours? Forgive my apparently naive question: I'm a little spaced-out this morning. I think what I'm trying to say is that you may only need to do sensitivity analysis on a normative gaussian model to make this leverage work. Sensitivity analysis on bayes nets with long reasoning chains has been well studied, so finding the weak points in complex models may turn out to be easier; will think more about this when brain returns to normal.
The lovely thing about this idea is that you don't need to worry about what the clever people do as they won't be getting in your way. It suffice that there are some (sufficiently big and sufficiently leveraged) gaussian guys. And I'm reasonably sure that is true.
Whether you would want to do the sensitivity analysis in the model or non-parametrically is a different question. My suspicion is that the anti-A portfolio is best constructed entirely without (joint) distributional assumptions. It's like dispersion trading - you don't care what the copula is, just that it isn't the one that other people think it is.
Post a Comment
<< Home