Honest question. Every ML trading story I find online is either a course ad or a hedge fund whitepaper I can't replicate.
Is there anyone here — retail, real account — running machine-learning signals profitably for 12+ months? Not looking for your alpha, just whether it's real and roughly what the setup looks like.
12 months live with a gradient-boosted classifier that ranks 30 forex pairs by 5-day forward return probability. It is profitable — but the edge is thin, roughly 0.3 Sharpe after costs.
The model isn't magic. It's a slightly better ranking filter feeding a boring momentum system. Anyone promising more than that, retail-side, is usually selling something.
The dirty secret of retail ML: the biggest win is risk management, not prediction. Regime classification and position sizing models beat signal generation models almost every time I've tested both. Prediction is the hardest possible problem to aim ML at.
Feature engineering beats model choice, every single time. My XGBoost with handcrafted features beats my friend's transformer fed raw OHLCV by a mile. The model is the easy part; knowing what to feed it is the job.
Counterpoint from the other side: I spent a year on LSTM price prediction. Conclusion: I built a very expensive random number generator. Respect to the people making it work — it's real, just rare and thin.