The Quant Model Effect

Posted on Mi 15 August 2007 in misc • 2 min read

  • Neural Nets tags: [] meta: dsq_thread_id: '181042049' author:

I was reading an article in the WSJ yesterday about how smaller investors have been dabbling in arcane hedging strategies just like big time Wall Street institutions. They’re trading future options, dabbling in currencies, and even selling short puts. Just like the pro’s, they figure that the markets can be analyzed using rational mathematical strategies and just like the pro’s they’re feeling the pain right now.

The talk on the street is that a lot of the recent market volatility is due from sophisticated quantitative models breaking down because they couldn’t adjust fast enough and value their positions. This, in my opinion, is leading to a lot of assets being mis-priced in this turbulent market but that’s another post altogether.

Quantitative models are fallible in turbulent markets because they can’t adapt fast enough to changing conditions. Their fallibility lies in the mathematical analysis that assumes turbulent markets can be rationally modeled. Everything works fine as long as the market operates fairly smoothly and during low volatile periods. This sure sounds like the markets we’ve had over the past few years! Taleb would say that we model the markets with the belief that we live in mediocristan but in reality we live in extremistan.

What about neural net models? Are they fallible too or can they adapt fast enough in a turbulent market? The answer is, “it depends.” Normally, neural net models will be unable to “turn on a dime” but they work relatively well if the market action remains within the bounds of their training data. By this I mean if your training data had captured similar volatile events during the construction of the model, you’ll be able to model possible outcomes. If the events are outside the training data range, then you have to continually re-optimize the model.

For the record, when I write words like “predict”, or “predictive” I usually put them in the context of a type of directional/trend neural net model which has some sort of nonlinear relationships between the inputs and outputs. It is foolish to use neural nets to predict the price of XYZ for Friday’s close because its like shooting in the dark, you don’t know if you’ll hit that target or not. I’m more interested in the curve plot that shows the relationships of the inputs to the outputs than predicting an exact price.

Neural nets are fantastic for predicting the direction of market trends rather than a specific closing price. Using pattern recognition or a basic back propagation neural nets, you can identify the market conditions and predict whether or not you’ll see an increase/decrease in volatility or if that crazy sell off is just a dip buying opportunity or the possible end to the trend!