ML Filter
Definition
An ML filter is a machine-learning model (typically a neural network or boosted-tree classifier) that scores rule-generated trade signals and rejects low-quality ones. Modern most-profitable scalpers use ML filtering to raise win rates without changing underlying strategy logic — keeping rule-based auditability while adding ML-driven quality screening.
In-depth: ML Filter
ML filters emerged as the dominant architecture for ML-augmented EAs around 2023-2024 because they pair the strengths of rule-based systems (transparency, predictability, auditability) with the strengths of machine learning (pattern recognition in noisy data, adaptive thresholding).
Mechanism:
1. **Rule-based signal generation**: the EA's underlying strategy generates candidate trade signals via traditional rule logic — moving-average crossover, indicator combinations, session-timing filters, etc. This produces a flow of "candidate trades" that meet basic rule criteria but vary widely in quality 2. **Feature extraction**: for each candidate trade, the EA computes a feature vector — typically 20-50 features describing current market structure (volatility, recent trend strength, distance from key levels, volume ratios, session timing, news-event proximity, etc.) 3. **ML model scoring**: a trained ML model (neural network, gradient-boosted trees, or similar) takes the feature vector and outputs a quality score — typically a probability that the trade will be profitable 4. **Threshold filtering**: trades with scores above the threshold are taken; trades below are rejected. The threshold is calibrated to maximise expectancy: too low and low-quality trades dilute the edge; too high and not enough trades are taken to compound the edge
Why ML filters work better than pure rule-based or pure ML alternatives:
• **Better than pure rule-based**: rule layers alone can't easily learn the complex feature interactions that distinguish high-quality from low-quality candidate signals. ML filtering raises win rate by 5-15 percentage points on rule-based strategies without changing position sizing or fundamental strategy logic • **Better than end-to-end ML**: pure ML systems are black boxes — when they fail, the failure mode is hard to diagnose. ML filters preserve the auditable rule layer, so failures can be attributed to (a) rule layer producing too few or too many candidates, or (b) ML layer mis-scoring candidates. This makes operational maintenance materially easier • **More resilient to ML model decay**: when the underlying strategy edge survives without ML augmentation (just with lower win rates), an ML model that has decayed produces a graceful degradation. End-to-end ML strategies have no fallback if the model decays
Examples of ML filter architecture in production EAs:
• **Scalperology AI**: rules-based M1 XAUUSD scalper with neural-network entry filter. Backtest of the rule layer alone shows ~55% win rate; with the ML filter, ~70% win rate. The rule layer remains transparent; the ML adds precision • **Smart Robot AI**: pattern classifier on multi-asset trend-following signals; ML scores each candidate trend signal and adjusts position size based on confidence • **Phalanx Neural AI**: regime classifier that doesn't directly filter trades but disables trading entirely when the regime classifier's confidence drops below threshold — a coarser ML filter pattern
For EA buyer evaluation: vendors using ML filtering should document the underlying rule layer (so buyers know what the strategy is doing fundamentally) and the ML training cadence (so buyers can assess whether the ML layer is being maintained). Vendors marketing "AI EA" without documenting either are likely using ML filtering as a marketing layer without serious engineering investment in the underlying rule architecture.