🔄 Continuous Learning
Walk-forward self-learning loop · every page visit captures predictions · auto-labels them after 24h · --
How this is different from emit→outcome: The old loop only trained on explicit "high conviction" findings emitted by the brain — maybe 5/day. This loop captures predictions for all 22 symbols on every page visit and auto-resolves them by checking the realized 1-day return. Net: ~200+ training examples per day instead of 5.

How concept-drift adaptation works: Rolling accuracy over the last 50 resolved predictions is monitored. If it drops below 45% (worse than coin flip), the learning rate doubles for the next training batch, making the model react faster to regime change. Returns to normal once accuracy recovers above 55%.

Refusal to learn from fake data: Only predictions captured against quotes with priceSource ∈ {stooq, coinbase, finnhub, polygon, alpaca, tradier} are eligible. Stale-seed or mock prices produce no training signal.
Captured today
0
Unresolved (waiting 24h)
0
Total trained
0
Rolling accuracy
—
📈 Rolling accuracy over time
📊 Captures per symbol
TIME
SYM
ENTRY
EXIT
PRED
RET
OUTCOME
🧪 Verification
You can confirm the loop is actually working by:
1. Open this page now. Click 📸 Capture now. Captured-today count goes up.
2. Come back 24h later. Click ⚖ Resolve now. Unresolved count goes down, trained count goes up.
3. Watch the rolling accuracy chart fill in. If it's above 50% with enough data, the brain is actually learning predictive patterns.
4. Check Brain Health — training row count should be climbing day over day.