π Prediction Log β get smarter over time
Every game prediction logged, every outcome scored, every cited factor tracked. After 30-50 games real signal emerges and the model gets sharper.
π How this works
Each prediction is committed to the static manifest predictions.json. Outcomes + scoring live in KV (prediction-outcome-{game_id}) so they update without a git push. The "Update with result" button on each row scores the prediction and accumulates factor-level accuracy stats β across many games, factors that consistently predict correctly bubble up; noise factors fade. Sam (or planning chat) updates after each game.
Foundation: All 3 backtests side-by-side Β· individual BT1 NBA Β· BT2 MLB Β· Strategy v2/v3. 40 games tested, +v2 retroactive on NBA showed +59% (13 bets β likely variance). Tracking the next 100 to find out where (or whether) real edge exists.
π― Dual-tier tracking Β· new
Every prediction now carries an optional tier field (tier_1 for safe $5 bets Β· tier_2 for refined longshots) and a methodology field (e.g. method_1_correlated, talent_gap, situational) so the dashboard can separate ledgers. Existing entries default to tier_1 Β· talent_gap. Read the full framework at /bets/dual-tier-strategy/. Schema documented in predictions.json.
π₯ Blowout Correlation Edge tracker
Bets placed: 0
Phase 1 target: 30 bets
Trigger opportunities scanned: β
Hit rate: N/A
Cumulative ROI: N/A
β Full framework at /bets/edges/blowout-correlation/. BCE bets stored under KV key prefix bce-bet-, indexed at bce-feed, phase state at bce-phase-state.
π Overall accuracy
Pending
β
games yet to play
Calibration
β
predicted % vs reality
π Pending predictions
β
Completed predictions
π― Factor tracking
Each cited factor accumulates a record across all scored predictions. After ~20-30 games, high-signal factors (consistently validated) get more weight in future analyses; noise factors get dropped.
No factors scored yet. After the first game outcome is recorded, accuracy starts accumulating.