Topic: How to avoid curve fitting over optimization

Hi,

did you have some tricks or settings to avoid over optimization and or curve fitting?

Re: How to avoid curve fitting over optimization

You can do double OOS validation

Cut the historical data, for example the end date would be 12-31-2018, and then you can set OOS percentage and set the acceptance criteria for the OOS data.

After you get the strategies you can revalidate them again with uncut/whole historical data.

Re: How to avoid curve fitting over optimization

Hi double OOS don´t help so  much.

For example:
I generate 1000 Strategies.
The first OOS filter remove 500 Strategies 1000-500=500
The second OOS filter half of this strategies 500-250=250

In the rest pool are too much curvefitted strategies.

The best way is to find a good generation setting. Choose the optimal indicators and SL TP settings.
OOS filter, Walkforward and Montecarlo are tools to decrease the curvefitting a little, but don´t solve this problem.

I think the best way is to use an optimal setting.
Choose strategies with many trades, more trades means less curvefitting.

for example:
You have an trading idear, you can set your indicators
You need good knowledge in forex and trading to use optimal settings.

Re: How to avoid curve fitting over optimization

Some generals tips to avoid curve fitting. Not all tools are applicable for all types of trading/strategies. Use when required. Some may or may not be available in EA Studio

- Monte Carlo testing
- Walk Forward Testing
- Increase sample size
- Reduce the degrees of freedom of steps for SL, TP, indicators etc. For example, minimum ATR step of 0.5 (0.5, 1, 1.5 etc).
- Round values to nearest 5, 10, logical number etc. For e.g, if a SL is 197, round it to 200. Is it still profitable?
- Apply strategy to multiple time frames, broker data and instruments.
- Use (and understand how to use) OOS/IS.

Ultimately it is very hard to avoid curve fitting and generally you will require further analysing post-generation in the live environment.