Is curve fitting less likely on intraday strategies (1min-5min) over muti decades with a large sample size?

[A G]

Hello,

We understand that curve fitting is clearly an issue.

But might it be harder to curve fit a strategy that trades on the 1-5 minute timeframe over say 10 years if it's trading 1000s of times.

Isn't curve fitting more likely on the higher timeframes with fewer sample sizes?

[Lorentzos Roussos]

A G:

Hello,

We understand that curve fitting is clearly an issue.

But might it be harder to curve fit a strategy that trades on the 1-5 minute timeframe over say 10 years if it's trading 1000s of times.

Isn't curve fitting more likely on the higher timeframes with fewer sample sizes?

It is likely if whatever is learning has the capacity to "memorize" conditions for trades . 

In machine learning however if you memorize the same thing many times it becomes "knowledge" in a way , sorta , kinda .

Naturally the question that follows is : "the memorization capacity of the genetic algorithm of the tester does not transcend the test i perform" and that is correct.

This means that eventhough the genetic algorithm may have the capacity to store things while searching it will essentially find the "best fitting" approach for the time period 

you have requested . What the tester outcome of an optimization (in mt5) means is :

"These parameters (your ea inputs) are the best for trading in this period"

Then it is your job to decide if this is something generic or something ultra specific . 

So with this optimization method the agent that the overfit emerges with is you , the coder.

However , the more "generalized" the data you give the tester is -> the more generalized the optimization may be , and , the genetic algorithm is less likely to "cheat" via memorization (like neural networks do for instance). In genetic algorithms you are likely to be fooled by the "absolute" nature of the method where it gives you the exact outcome you asked for but the one that is easier to attain. For example if you ax an algorithm to find the best strategy for gbpusd for 2021-2022 it will just sell , if you optimize by p+l  , if you optimize by % it will perform one trade and it will be a winning one .etc 

[A G]

Lorentzos Roussos #:

It is likely if whatever is learning has the capacity to "memorize" conditions for trades . 

In machine learning however if you memorize the same thing many times it becomes "knowledge" in a way , sorta , kinda .

Naturally the question that follows is : "the memorization capacity of the genetic algorithm of the tester does not transcend the test i perform" and that is correct.

This means that eventhough the genetic algorithm may have the capacity to store things while searching it will essentially find the "best fitting" approach for the time period 

you have requested . What the tester outcome of an optimization (in mt5) means is :

"These parameters (your ea inputs) are the best for trading in this period"

Then it is your job to decide if this is something generic or something ultra specific . 

So with this optimization method the agent that the overfit emerges with is you , the coder.

However , the more "generalized" the data you give the tester is -> the more generalized the optimization may be , and , the genetic algorithm is less likely to "cheat" via memorization (like neural networks do for instance). In genetic algorithms you are likely to be fooled by the "absolute" nature of the method where it gives you the exact outcome you asked for but the one that is easier to attain. For example if you ax an algorithm to find the best strategy for gbpusd for 2021-2022 it will just sell , if you optimize by p+l  , if you optimize by % it will perform one trade and it will be a winning one .etc 

Yep, I kind of agree with what you're saying... If we keep the strategy generic enough, and have a big enough sample size, timeframe, and market conditions we're less likely to curve fit - which is something we want to try and avoid as best we can. 

Thanks for your detailed response.