Sunday, October 12, 2008

Systems Development - Keep it Simple Stupid

Earlier this year I had a conversation with a head trader at Eclipse Capital, a CTA in St. Louis, MO.  We were discussing trend trading models and the level of sophistication required to trend trade.  He had experience at an east coast firm full of PhDs and the midwestern firm.  

I noted that from testing a wide variety of models, the simplest models were the most effective.  He agreed, saying that he had a room full of PhDs and no alpha, or something to that effect.  The complex trend trading models were not necessarily robust enough to be significantly better than relatively simple models.  

As parameters are added to a backtested system, the robustness of the model decreases due to the smaller number of trades tested and subsequently future trades taken.  Also, models are not perfect representations of how markets will behave in the future, therefore an optimized model will not necessarily perform as well as desired.  

Optimization does give us a good place to start in model implementation because it is the best thing that we have to go on.  For example, on a simple price channel system, varying the channel breakout days in testing provides a large range of outcomes.  The cluster of breakout days near the optimal point should give a better indication of how markets tend to behave over long periods of time.

The model tends to work the best using breakouts between 20-30 days.  Testing this over different time periods also yields interesting results.   

Between 2000-2008 the results are similar altough, the best performing breakout range was extended further out.



In 2008 the system performed well in the 10-12 day breakout and 35-40 day breakout range with the worst performances in the 20-30 breakout range.  This is probably due to the high volatility that has been experienced this year.  


These findings suggest that trading the same model with differing breakout days may potentially smooth the equity curve.  So, rather than adding complexity to the model through more parameters, trading the same model with different inputs may provide an improved volatility adjusted return. A future article will focus on this topic.

No comments: