Adrian Reid explains how traders destroy their systems through overfitting. Instead of tweaking parameters endlessly, step back and add missing rules or new data types like volatility and volume. Redundant indicators are the trap; diverse data sources protect against curve-fitting.
How do I prevent overfitting while still making the system better?
Now this is a really great question because you can overfit in so many ways and all of them are bad, but you can still improve your system in a good way. So how do we do it? The first thing is have a back testing process like I talked about earlier. What are the steps you follow in order, in sequence to backtest your system correctly? Do it the same way every time.
When you're trying to improve your system, the best way to do it is not by optimizing, not by varying the parameters and trying to find better parameter combinations.
So don't try and improve your system just by optimizing. Instead, think. All right, step back and look at the rules and say, what is it missing? What market behavior could hurt this system?
What market behavior could this system benefit from that it's not currently catching? What type of conditions.
Generate bigger returns or smaller returns for this strategy? What market conditions are my worst losses in all my best gains? And try and think about some filters or some like a regime filter that might turn it on or off at a smarter moment to better capture the edge. Or try and think about a different type of data to help you refine. If you've got three moving averages and a MACD and a rate of change and you throw a Ballinger band at it.
It'S not going to help because all of those things are basically measuring the same thing.
So instead, think of what other types of data can I get? Maybe it's volatility, maybe it's relative strength, this versus the market, this versus other stocks. Maybe it's candlestick patterns, maybe it's volume. Basically you're looking for other types of data. Maybe it's the performance index, the performance of the currency of the market you're trading, the performance of gold or the underlying commodity of the that the companies are using. Think about other types of data that you can incorporate because the more you use the same indicators on open, high, low, close, the more you're just overfitting to exactly what happened in the past. You want to have different angles to try and improve your system.
I don't want too many rules that use the same data in my system because all it is overfitting. Instead, try and bring some diversity into the system decision making and you'll find that you'll get much better results, more robust rules.
