What this shows: Every brain prediction captured by the continuous-learner that crossed a conviction threshold. We treat each as a paper trade: ENTRY at the price-at-capture, EXIT at the actual price 1 day later. R-multiple comes from the realized return divided by an ATR proxy.
Different from walk-forward backtest: walk-forward replays HISTORICAL bars against the CURRENT model. This page shows the brain's REAL predictions made at the time the data arrived. It's a true forward record — no look-ahead, no cherry-picking.
Different from autopilot paper trader: autopilot uses high-severity findings. This uses every continuous-learner capture that met conviction. It's a complete record of every confident call.
For each continuous-learner journal entry that:
• Has predProb such that |predProb - 0.5| ≥ threshold (conviction filter)
• Was resolved at the short (1-day) horizon (has realizedRet)
• Has outcome ∈ {correct, wrong} (skip flat)
We compute:
• side = predProb ≥ 0.5 ? LONG : SHORT
• R = signedReturn / max(0.005, atrPctProxy) — capped at ±3
• cumulative R sum forms the equity curve
Sharpe: daily R's mean / std × √252. Max DD: deepest peak-to-trough on the equity curve. Profit factor: sum of positive R / |sum of negative R|. A profit factor > 1.5 is decent, > 2 is good.