Model · Learning System v2
COLLECTING DATALoading...
Training Data Progress0 / 80 closed trades with features

Need 80 more closed trades. At ~15/day, ready in ~6 days.

Paper TradingCurrent system
P&L+$0.00
Trades0

Composite score + heuristic

ML ModelPending

Awaiting 80 trades

BaselineAlways NO > 0.80

No resolved baseline contracts yet

Feature ImportancePending

Feature importance appears after the ML model trains

Prediction CalibrationPredicted vs Actual
🧠

How the new model works

The system now extracts 30+ features for every contract opportunity (time to resolution, price momentum, volume, wallet behavior, category type, etc.) and stores them as a snapshot with each paper trade. Once 80 closed trades accumulate feature snapshots, a LightGBM gradient boosting model trains automatically using walk-forward cross-validation. It predicts P(WIN) directly from features, replacing the old heuristic. The model retrains weekly as new data arrives. If the ML model can't beat the "always bet NO > 0.80" baseline after 200+ trades, the honest conclusion is there's no exploitable edge.

What changed from v1

Stop-loss disabled for contracts resolving within 7 days (they're binary outcomes — let them resolve). Entry prefers contracts resolving within 48 hours (fastest learning feedback). Weight oscillation stopped (frozen). Scoring threshold tiered: 70 for < 2 days, 75 for < 7 days, 80 for longer. Factor weights no longer retrain daily — the ML model replaces them entirely once it activates.