forex-scalping.com

Machine learning can find patterns in data, but forex data is noisy and non-stationary. Expectations should be modest.

A sensible benchmark is a simple rule-based strategy on the same data. If the machine learning model cannot beat it clearly after costs and out of sample, the extra complexity is probably not justified.

Feature importance is worth checking as well. If a model relies mostly on one or two features, a simple rule based on those features may achieve most of the benefit with far less risk of overfitting.

Challenges:

• Overfitting: complex models fit noise.
• Changing markets: patterns may disappear.
• Transaction costs: small predicted edges vanish after spread and commission.
• Data leakage: using future information by mistake.

Approaches that help:

• Simple models first (logistic regression, decision trees).
• Strict out-of-sample and walk-forward testing.
• Features based on market logic, not random combinations.

Have you tried machine learning for trading, and what did you learn?

The benchmark point is the one I'd underline. When I tried a gradient boosted model on EURUSD M5 a couple of years ago, it beat a plain Asian range breakout rule in sample by a nice margin and then lost to it out of sample once costs were in.

The reason was boring. The predicted edge was around 0.3 pips per trade before costs, and my round trip with commission was close to 0.8 pips. The model was guessing direction slightly better than a coin and paying a toll on every guess.

I'd add one item to your leakage list that caught me. My labels used the bar close while some features used the same bar's high and low, which looked brilliant in testing and was useless live, because in real time you don't know the high until the bar is finished.

What has worked better for me is using a model as a filter rather than a signal. A small classifier that says "skip this one" on days with abnormal spread or a very compressed range removed some of my worst trades without inventing new ones.

Has anyone here had a model survive a walk-forward across a real regime change, like the 2022 dollar run? That's the test I'd trust most.

FXS Prop Desk, one item I would add to your challenges list is overfitting the test set itself. If you try 100 model settings and keep the best walk-forward result, about 5 of them will clear a 5% significance bar by luck alone. The walk-forward then stops being out of sample in any useful sense.

What fixed that for me was a final holdout I only touch once. The last six months of data sit in a separate file, and the model gets one run on it after every choice is frozen. If it fails there, the project goes back in the drawer rather than into another tuning round.

Label design matters as much as features. Predicting next bar direction gave me 52% accuracy and nothing tradeable. Predicting whether a trade hits a 1.5R target before a 1R stop within 30 minutes matched how I actually trade, and the result was easier to judge after costs.

Accuracy alone also misleads on scalping labels. If 70% of setups are losers by design, a model that always says skip is 70% accurate and useless.

How many configurations do you allow yourself before you call a feature set dead? I cap myself at twenty and log every one of them.