The brain is graded on data it never trained on (walk-forward + random split). This is the standard "is it real signal or noise?" test — and it's already statistically significant.
Walk-forward accuracy
—
vs coin flip (50%)—
p-value—
sample (n)—
Held-out accuracy
—
Brier—
calibration (ECE)—
p-value—
Each trading day we log every directional call + its entry price, then grade it against the real 5-trading-day move. No cherry-picking, no hindsight. This is the receipt that decides whether paid options-flow data is worth it — it fills out over the coming weeks.
How to read this. Backtest = how the brain scored on history it never saw (proves the signal is real, not curve-fit). Forward test = the live, no-excuses receipt: did today's calls actually work? "Beats coin flip" needs a one-sided z > 1.64 (95% confidence) — which needs enough graded calls, so early on it correctly says "too few." A real edge here is small by nature (markets are efficient): even 53–55% directional with positive expectancy is a genuine, monetizable edge when sized right. If the confluence leg (brain + insiders agreeing) beats the brain-only leg over time, that's the proof that fusing more signals — including paid options flow — adds real edge.