Sticks V6 — Known Limitations & Realism Check
Model Limitations
Small Sample Size
- Backtest: 28 bets at +3.9% ROI — statistically meaningless
- Need 500+ bets to confirm edge exists (3-4 months of paper trading)
- Current "profitable" result could be pure variance
xG Model is Simple
- Uses team goals scored/conceded averages with 50% regression to league mean
- Real xG models use shot location, shot type, body part, game state
- Our model will never be as accurate as bookmaker models with proprietary data
- The edge comes from finding MISPRICED markets, not better predictions
Limited Historical Odds
- API-Football only serves odds for recent/upcoming fixtures (not full season history)
- Backtest only covers ~130 fixtures with real odds across 5 leagues
- Can't test across multiple full seasons
Away Win Model is Weak
- 29% win rate vs model prediction — consistently overestimates away team probability
- Home advantage adjustment (8% boost / 7% reduce) may be insufficient
- Away teams face crowd pressure, travel fatigue, pitch unfamiliarity
Operational Risks
API Dependencies
- API-Football downtime = no analysis, no settlement
- Betfair downtime = no odds, no bet placement
- ESPN downtime = no live scores (fallback only)
- Rate limits: Ultra = 450 req/min, 75k/day — sufficient but not unlimited
Settlement Reliability
- Primary: API-Football fixture_id lookup — reliable but costs 1 API call per bet
- Fallback: ESPN fuzzy name matching — fragile for non-English team names
- Corner/shot markets require additional API call for fixture stats
- Some markets can't auto-settle (e.g. player-specific markets)
Betfair Specifics
- Exchange liquidity varies — some markets have thin order books
- Odds can move between model check and bet placement
- Session tokens expire — need re-login handling (not implemented for long runs)
Paper vs Live
- Paper mode has no execution risk (slippage, rejected orders, liquidity)
- Live mode would face: bet acceptance delays, partial fills, account limits
- Betfair takes 5% commission on net winnings (not modelled in paper mode)
What This System CAN Do
- Systematically scan 7 leagues for statistical edges
- Compare model probabilities to real bookmaker odds
- Only bet when edge exceeds bookmaker margin (8%+)
- Track everything: bets, P&L, CLV, XP, history
- Settle automatically using fixture results
- Run 24/7 with 60-second heartbeat
What This System CANNOT Do
- Guarantee profits — no betting system can
- Beat the market consistently long-term without continuous improvement
- Account for injuries, suspensions, team news (unless Claude filter is active)
- Replace human judgment for qualitative factors
- Work on illiquid/exotic markets
Path to Production
- Run paper mode for 3+ months (target: 500+ bets)
- Track CLV — if consistently positive, edge is real
- If ROI > 5% over 500+ bets, consider live with minimum stakes
- Start live at 25% of paper stake sizes
- Scale up only after 1000+ live bets confirm the edge
- Monitor for model degradation — bookmakers adapt