How WinPIQ reads a football match
No more guesswork. Football decisions built on data.
1. Information
Results, goals, xG where it exists, market prices — from authorised sources only. Each row records where it came from, and that origin is stored with it.
2. Ratings
Elo with a per-league home advantage, attack and defence strengths that decay over time, and form adjusted for the strength of the opposition faced.
3. Models
Elo, Poisson, Dixon-Coles and gradient boosting — weighted by log-loss on earlier seasons only, then isotonically calibrated.
4. Against the market
The bookmaker's margin removed, a fair price derived, and the gap expressed in probability points. The published probability is the model's alone.
5. Confidence and information quality
Confidence is not probability. The honest states are named: no edge at this price, waiting on squad news, no suitable opportunity right now.
6. Proof
Every prediction gets an identifier, a timestamp and a SHA-256, is sealed before kickoff and kept permanently — the misses included.
What the backtest shows
| Season | Matches | Model accuracy | Market accuracy | Model log-loss | Market log-loss |
|---|---|---|---|---|---|
| 2223 | 1826 | 52.8% | 54.1% | 1.007 | 0.978 |
| 2324 | 1752 | 53.3% | 55.0% | 0.985 | 0.956 |
| 2425 | 1752 | 52.0% | 53.6% | 0.991 | 0.964 |
| 2526 | 1752 | 52.3% | 53.4% | 0.998 | 0.978 |
In all 4 completed seasons the market price alone was both more accurate than the model alone and scored a lower log-loss. We do not claim otherwise. These are historical model-only and market-only figures; what the published calls actually did is in the record.
Source: services/ml/artifacts/backtest.json · isotonic calibration iso_v2 · updated 2026-08-31 · season 2627 excluded: no settled market prices for this season yet
The evidence
Every settled prediction: the verified record · where the information comes from · editorial policy.
Built like a fintech, not a tipster
- No guaranteed profits
- No edited history — every version stays
- No operator influence, no affiliate money
- Transparent methodology, model version on every prediction
Questions
- No. WinPIQ is decision intelligence: probabilities, fair prices, risk, uncertainty and a verified record. It also tells you when to pass.
- Not yet on closing lines: backtest log-loss 0.99 vs market 0.96–0.98. We publish this because you should know it. Calibration is our first target; edge is the second.
- The whole verified record, the Match of the Day before kickoff, decision states for it, the daily Time Machine and a limited Ticket X-Ray.
- One click, immediately, no questions required. Annual plans are paid up front for the season.