Knowns and Unknowns through the Crunch

I have been reading Cassola et al.'s A research perspective on the propagation of the credit market turmoil from the most recent ECB research bulletin. The paper concentrates on information issues in the credit crunch. Here's how I see matters there.

Before the crunch investors had some information on asset prices as many ABS were (or seemed to be) liquid. On the other hand there were two classes of information failure relating to asset risk: ignorance (where a risk taker simply had not tried to assess risk); and model failure (where the risk taker had tried, but had got the wrong answer because their model was mis-calibrated and/or failed to include all the pertinent risk factors).

After the crunch in contrast, investors had less information on asset prices but more information on the risk of ABS. Rapid asset prices changes combined with illiquidity caused rapid selling, model recalibration, and risk takers had more information.

Then however, understanding that ABS were illiquid, that prices were both falling and uncertain, and that it was unclear who had what, investors in financial institutions became concerned about their ability to assess bank credit quality. As Cassola et al put it:

the actual extent of exposures to the problematic instruments and the health of specific financial institutions becomes only gradually more known

Hence the interbank market freeze up, amongst other things.

It seems then that a good understanding of the interplay between what's known, what's known to be unknown, what's thought to be known, and what you don't even know you need to know is important in explaining the Crunch. (Why did Rumsfeld get so much grief for this? I think it was the smartest things he has ever said.) In particular new information can suddenly throw the spotlight on something you thought you knew but in fact you didn't. Caveat emptor.

Black box trading and information fusion

According to Wikipedia, Information Fusion

refers to the field of study of techniques attempting to merge information from disparate sources despite differing conceptual, contextual and typographical representations

The convention is to keep the term data fusion for the situation where all information is quantitative, and use information fusion for the broader problem of integrating quantitative and qualitative data.
Another authority says that data fusion

takes isolated pieces of sensor output and turns them into a situation picture: a human-understandable representation of the objects in the world that created those sensor outputs.

Basically then, whenever you have diverse data which you have to try to turn into a coherent picture, you are performing data or information fusion.

Unsurprisingly much of the academic interest in this area occurs in limited problem domains: figuring out where the planes are from radar and visual data, for instance, or combining multiple different sonar sources to get a more complete picture of what's swimming around you. Many quantitative trading models of this class: they take feeds of market data and transactions and attempt to form a picture of where the market will go next. One simpler class of models, for instance, are basically trend followers. Often the idea of momentum is used: when markets are rising on increasing volume with low volatility, the models pile in, perhaps intensifying the rise. Decreasing volumes and/or rising volatility are sometimes used as triggers to reduce the size of the trade.

Many quantitative models, then, implicitly have a confidence estimate built in. When they strongly believe in their own predictions, they put a trade on. When they either don't believe in them, or they cannot make a prediction, the trade is taken off.

This feature is important: quantitative trading has been described as picking up pennies in from of a steam roller, and certainly many trading strategies act like short gamma positions, making a little money when they work, but losing a great deal when they are wrong. A false positive - a trade that you don't think will work and so don't make but in fact would have been profitable - is a lot less bad than a false negative - a trade you do think will work but turns out not to. The magnitude of this issue can be seen from Morgan Stanley's $480M one day quant trading loss.

For this reason, some quant traders use multiple models and only trade when all of them are giving the same signal. If the models are sufficiently different and do not share common assumptions this helps to reduce model risk.

It occurred to me recently that another approach might be to cast quantitative trading as an information fusion rather than a data fusion problem. That is, is there non-quantitative information that might be useful, in particular in avoiding false negatives by making the model more doubtful in situations where more care is needed? One of the anticedants here is the theory of prediction markets: when a large number of independent people have an opinion, a suitable weighting strategy can often lead to better predictions than any individual pundit. Note that I am not discussing analysts opinions here - there are clearly institutional biases at work there, and the history of collective analyst predictions is not that promising. Rather I am suggesting trying to use the commentariat, ideally as large a body of it as possible, as a signal akin to rising volatility. When enough blogs start to discuss a possible crash, that is a sell signal akin to rising volatility or rising market risk premiums. Such an information fusion based quantitative trading model would be of more use in global macro than in very short term applications like index arb, but the idea of using rising worry as a deleveraging signal could be interesting. Or it could just be a heap of potatoes.