- Fix streak signal: was giving 100% to streak chair after normalization
(non-streak chairs were 0), now properly distributes probability with
streak chair getting less as streak grows (actual mean reversion)
- Change recent window from 20 to 50 games
- Reweight signals based on backtest: base_rate 0.15→0.20 (best performer),
recent 0.10→0.15, streak 0.10→0.05, balance 0.15→0.10
- Move Live Market Sentiment above Signal Breakdown
- Bet impact simulator on /predictions shows rank headroom and safe bet amounts
- Password-protected /visitors page with visitor log table and stats
- Console now logs real visitor IPs instead of Cloudflare tunnel IPs
Balance signal (15% weight) favors under-represented chairs over last 50
games. Visitor middleware captures real IPs from CF headers, batched into
ClickHouse with 90-day TTL.
- Whale/public picks now track 2nd pick and score SEMI (0.5 pts) like model
- Prediction table expanded from 20 to 50 rows
- "View All History" modal with pagination (50/page), fetches up to 500
- Accuracy rows use semi-win scoring for all three columns
- Add _compute_whale_public_picks() to reconstruct whale/public picks from historical bets
- Merge whale_pick, public_pick, whale_hit, public_hit into last_20_predictions
- Add get_prediction_history(limit) for lightweight prediction+accuracy data
- Add /api/prediction-history endpoint (default 100, max 500)
- Add Whale and Public columns with HIT/MISS to Last 20 table in frontend
Prediction hero now shows ranked TOP PICK and 2ND PICK with EV per unit
bet (P(win)*2.9 - 1). Bet Size Advisor panel shows Kelly criterion
fraction (capped 25%), best chair with confidence, and historical bet
rank insight (how often lowest/highest-bet chair wins).
Live Market Sentiment section tracks whale trend (top 5 bettors by amount)
and public trend (total pool distribution) in real-time via WebSocket,
mirroring the live dashboard. Notes highlight agreement/divergence between
model pick and crowd favorite.
Historical crowd analysis cards show how often the most-bet, mid-bet, and
least-bet chairs actually won across all games.
Round result flash now includes whale/public pick accuracy alongside the
model prediction result. user_bet WebSocket events are tracked to build
per-round bettor profiles for whale analysis.
- WebSocket connection shows live game state (round #, phase, bets per
chair, pot) in a persistent bar at the top of predictions page
- Prediction cards now display current bet amounts per chair
- Round results flash HIT/MISS against the Bayesian prediction
- New "Last 20 Predictions vs Actual" table with per-game probabilities,
predicted vs actual winner, and running accuracy
- Predictions auto-refresh after each round ends
- Fix winning cards chart: use taller container (480px) and dedicated
scales config for horizontal bar rendering
- Add _last_n_predictions() helper to db.py for detailed per-game
prediction history with game numbers
Bayesian next-chair predictor (Markov chains, base rate, streak regression),
statistical tests (chi-squared, runs test, autocorrelation), theory
backtesting with rolling accuracy, and card-level analysis (value/suit
distribution, face card frequency, top winning cards).
New /patterns page with 9 analyses: chair win bias, bet rank
correlations, hand type distributions, pot size buckets, streaks,
hourly patterns, and recent-vs-overall comparison. Also adds a
standalone analyze.py CLI script for terminal output.
CHAIRS mapping was {1:A, 2:B, 3:C} but the API's country field has
1=C and 3=A. Fixed the mapping in backend and both frontends, added a
startup migration to swap A↔C columns in existing DB data, corrected
multiIf SQL queries that hardcoded the wrong winner→column mapping,
and moved _save_round() to the ENDED status block so hot/cold stats
reflect the latest round immediately.
