Let’s examine two systems with the over the data set January 2000 – December 2009.
"Static results" (below) shows the results of a standard trend following model utilizing two exponential moving averages (XMA) with periods 20 and 50. If the 20 XMA crosses above the 50 XMA a long signal is triggered, and if the 20 XMA crosses below the 50 XMA, a short signal is triggered. The in-sample data range is January 2000 – December 2003 and the out-of-sample is January 2004 – December 2009. As you can see, it is not a terrible system over the life of the entire data set, but there is a better way.
"Rolling with the market" (below) shows the results of a trend following model with the same in-sample and out- of-sample data set. While this is a single system, it processes signals by analyzing 20 unique sub-systems. Each sub-system is independent, meaning it has no bearing on the other systems within the model. The model then chooses what position it will take based on what the sub-systems are doing as a whole. No single sub-system is the driver for entry and exit signals. Similarly the logic behind each sub-system is generally simplistic in nature. As noted in the performance info, this style of system performed much better over the same data set.
One of the main problems with systems is the propensity for them to be curve fit or for the logic to be too specific. Allowing trading models to have a certain amount of freedom by requiring them to process a larger number of market generalities, as opposed to specifics, will allow the strategy to be more adaptable in the long run. This should allow the system developer to stay in the game much longer than the one who believes that specific market actions beget specific reactions.
Sam Butler is president of Altaloma Asset Management LLC and Troy Asset Management LLC. E-mail him at firstname.lastname@example.org.