TradingAgents/docs/iterations/scanners/ml_signal.md

ML Signal Scanner

Current Understanding

Uses a trained ML model to predict short-term price movement probability. Current threshold of 35% win probability is worse than a coin flip — the model needs retraining or the threshold needs raising to 55%+ to be useful. Signal quality depends heavily on feature freshness; stale features degrade performance.

Evidence Log

(populated by /iterate runs)

Pending Hypotheses

• Does raising the threshold to 55%+ improve precision at the cost of recall?
• Would retraining on the last 90 days of recommendations improve accuracy?