Program Trading Models

Updated Oct 10, 2008

My first two purely blackbox trading models are Aggressive Hedge I and Aggressive Hedge II. These models are very simple to update (I just run a couple of computer programs on data files I regularly need to update anyway) and so I'll look in on them from time to time to see what they're doing. Most of the amazing performance of Agg Hedge II was during the backtest period - not suprising, as they were tuned to be good during creation and backtesting. Their good performance continued for some time thereafter, but in '05 looked toppy, and after that were downright dangerous. What is the message here? I think there are two: First and most obvious, backtesting a mechanical model you've settled on will virtually always produce more impressive results looking backward than forward. The second message is that program trading and the rise of hedge funds as major fractions of total market volume makes for strong competition and rapid discovery and exploitation of successful strategies. And an over-exploited pattern becomes a source of losses instead of gains. I am always going to worry about the continued success of these 'black box' models. And for these reasons it is hard to imagine that I'll choose to trade real money with them... for the moment.

But Then...

In mid 2004 I had an idea for a new black box model using Profunds and coded it up. It made profits during the '99 to '03 mostly bear market; but by late 2004 the bull market was well established and I lost interest in mechanical models. The code sat there for 4 years, un-looked at. Now with the bull market staggering as of late '07 onward and noting the clear failure of the pronouncements of the top rated market timers, I looked back and found this model and ran it on the historical data. It has performed reasonably well, and continues that good performance so far during the current bear market. . I call it "Model A" - date of inception= Oct 4, 2004. Encouraged, I devised a more sophisticated model which uses various data as input to the calculation of a "figure of merit" on which trading risk/reward and a trading allocation decision is made. This "Model C" - date of inception = Mar 25, 2008 - is doing much better, producing good profits which are robust against reasonable changes to the very limited number of free parameters. I'll be interested to see how it does going forward. None of these models include breaking news or other unpredictables, just historical data.

The message seems to be that models must be continually created, but the encouraging aspect is that a good strategy can continue to perform well for a decent interval of time after creation, before a new model must be devised. Most hedge fund quants will agree that this is the new world order.