ML Filter

Overview

VECTRA's ML Signal Filter uses a Gradient Boosted Decision Stump model trained on historical trade outcomes to predict whether a new signal will be profitable. It acts as the penultimate gate before AI validation.

Feature Extraction

The ML model receives normalized features (all scaled 0-1):

Feature	Source	Normalization
Confidence	Signal confidence score	Direct (already 0-1)
Direction	LONG=1, SHORT=0	Binary encoding
Risk:Reward	SL/TP distance ratio	Divided by 5 (5× = 1.0)
RSI	14-period RSI	Divided by 100
ADX	Trend strength	Divided by 60
ATR Ratio	Current vs. average ATR	Log-scaled
Regime	Market regime encoding	Ordinal encoding
Funding Rate	Current funding rate	Scaled
Volume Ratio	Current vs. average volume	Divided by 3

Training Process

1

Data collection

Train on completed trades with known outcomes. Features are extracted at signal generation time.

2

Labeling

Outcome labeled as WIN (positive PnL) or LOSS (negative PnL).

3

Model training

Gradient Boosted Decision Stumps trained incrementally.

4

Persistence

Model persisted to disk for restart survival.

Decision Logic

The ML filter outputs a probability score (0-1). If the score falls below the threshold, the signal is rejected with a "ML_FILTERED" tag. The threshold is adaptive based on recent model accuracy.

Custom Training (INSTITUTIONAL)

Institutional tier users can retrain the model with custom feature sets and tune hyperparameters for their specific trading style and pair preferences.