Nerds on Wall Street: Math, Machines, and Wired Markets

8.10. Genetically Optimized Forecasting Models in Hindsight

Looking back on these models many years after we developed them is highly instructive. We did get what we wanted from the GA. It allowed us to produce much better models using the criteria we chose than other previous methods had. It produced simpler, more reasonable models with better statistical and financial performance in the test periods.

We tried to avoid data mining, but the market has only one past, and we knew what it was. The bias introduced by that knowledge cannot be removed.

After they were developed, these models made money in some countries, but not in others. Arguably, there were structural changes in the markets that ran contrary to what any rearview model based on the past would predict. Long Term Capital Management, a veritable who's who on Wall Street with multiple Nobel laureates as founders, had similar, though more spectacular, troubles.

No matter how much we want to think otherwise, financial markets are not physics experiments. They reflect a shifting combination of economic forces and human emotion. Like all other facets of human behavior, they are a mixture of the rational and the irrational, not necessarily the best target for the approach described here.

Some of the risks in modeling markets, with or without the aid of evolutionary computation, are unavoidable. Others are not. There is no guarantee of future wealth. There are some basic caveats one should follow to avoid the known pitfalls ...