One of the biggest appeals of mechanical trading strategies is that there are an endless number of parameters and indicators that a developer can add to any system. The first instinct of every new system trader is to attempt to improve on a basic system by adding another component to it. This usually results in a system that looks better in backtesting, but fails to perform well moving forward.
As system traders continue to experiment with different strategies, many come to appreciate that over-complicated systems are more prone to curve-fitting bias. Simpler strategies might look less impressive on backtests, but they are generally more robust options to trade moving forward.
Comparing System Trading To Cooking
The authors of GESTALTU published a post where they compared trading system development to cooking. The similarity that they identified in both fields is that more ingredients does not necessarily equate to a better overall result.
The article explains that adding more ingredients to a recipe will likely detract from the way the original ingredients worked together. 同様に, adding more indicators to a strategy will likely influence how that strategy performs in certain market environments.
The author describes a strategy that he developed early in his career that contained 37 different parameters. With that many different inputs, finding an optimized version of the strategy become nearly impossible. さらに, any backtesting results using that many variables are likely to be exposed to curve-fitting bias.
Limiting Degrees of Freedom
In yesterday’s post we touched on the different ways that price action and technical indicators can be exposed to curve-fitting. In each case, the key was to limit the degrees of freedom in the system.
The fact that price action generally used less parameters than technical indicators makes price action less likely to be exposed to curve-fitting. Applying that logic more broadly, the more parameters a strategy has, the more likely it is to contain some degree of curve-fitting.
Backtesting with Multiple Parameters
While backtesting strategies with large numbers of parameters is certainly possible, it is exponentially more difficult. Increasing the number of parameters will require a much larger sample size in order to produce backtesting results that are worthwhile.
Because backtesting results from systems with more degrees of freedom are more prone to curve-fitting, we can reasonably assume that results from systems with less parameters are more sound. したがって, we can have more confidence that lower parameter systems will continue to produce similar results moving forward.