How do you actually decide which forex pairs are worth trading? - page 2

New comment
 
Muhammad Minhas Qamar #:
[U]nless you are training on ticks, there's a high chance of overfit[t]ing.

I'm afraid that I don't follow. In my experience, backtesting EA's that control for OHLC prices on OHLC data is far more efficient than backtesting EA's that control for tick prices on tick data─especially terms of say, 30-year samples. The resolution of data required has more to do with the specific trading logic of an individual EA being backtested than anything else. Assuming that all of the EA's being backtested are coded properly and that an agent super-farm is used, data resolution is irrelevant to both over-optimization and overfitting.

Muhammad Minhas Qamar #:
[S]light changes should not affect drastic change in behavior. That's the point.

Such slight parameter changes are textbook over-optimization. Having said that, over-optimization can be one element of greater overfitting committed by a trader gone wild in other aspects of code development.

Muhammad Minhas Qamar #:
An event like Brexit can't be predicted anyways.

Just slap GBPUSD spot together with GBPUSD futures. They both got Brexitted.

 
Muhammad Minhas Qamar #:
What you are talking about is related to training sample. It's a workflow for ML. Overfiting is generally used to describe a model that does not generalize well. Whether or not you have more or less training sample will indirectly affect this, so it's not what overfitting is and by the way, you can pretty easily overfit "code" to even a century of data. Consider a modern ML architecture for transformers. It requires millions if not tens of millions of data points for it to reliably perform on out-of-sample tests. So unless you are training on ticks, there's a high chance of overfiting. 

Yes, but those slight changes should not affect drastic change in behavior. That's the point. 

You are correct. With my example for EURUSD and GBPUSD the biggest example of this was during Brexit. But again my point was that a strategy should work well enough on pairs that behave the same. An event like Brexit can't be predicted anyways. 
You guys are kinda saying the same thing tbh. Tons of data won’t save you if you’re curve fitting like crazy. If tiny tweaks wreck performance, that’s a red flag. Markets switch up too, so past “edge” can vanish real quick.
 
MarcosNavarro1 #:
You guys are kinda saying the same thing tbh. Tons of data won’t save you if you’re curve fitting like crazy. If tiny tweaks wreck performance, that’s a red flag. Markets switch up too, so past “edge” can vanish real quick.
Yeah, that's why I haven't replied 
 
Yeah I agree with a lot of what’s being said here, especially around volatility and overfitting.

From what I’ve been seeing in testing, a big issue is that strategies are often evaluated as if they should work consistently across all conditions, but in reality their performance is highly dependent on the environment.

The same system can behave completely differently depending on whether the market is trending, unstable, or lacking momentum.

That’s where most of the inconsistency seems to come from — not necessarily the logic itself, but when it’s being applied.

Lately I’ve been focusing more on breaking down performance by market condition rather than just overall results, and it’s been interesting to see how much of the drawdown comes from specific phases.

It changes how you look at strategy development completely.

I’ve been analyzing this more deeply using actual trade results, and it’s surprising how often losses are concentrated in just a few conditions.
12
New comment