Model Methodology

1. Daily Pipeline

Every morning at 04:00 UTC the pipeline runs five sequential stages for all matches with kickoff that day:

1. 1
   Fixtures. Fetch all matches from API-Football (280+ leagues). Store in matches table with team IDs, league, kickoff, venue, referee.
2. 2
   Enrichment. Fetch standings, H2H records, team stats, and injury reports. Runs again at 12:00 and 16:00 UTC to pick up late injury news.
3. 3
   Odds. Fetch pre-match odds from 13 bookmakers via API-Football bulk endpoint. Runs every 2 hours 07:00–22:00 UTC. Closing line snapshot taken at 13:30 and 17:30 UTC.
4. 4
   Predictions. Run Poisson model + XGBoost ensemble for each match. Compute ensemble probability. Store to predictions table.
5. 5
   Betting. Apply calibration + Kelly sizing. 16 strategies in our ensemble evaluate their configured markets independently. Qualifying bets are stored to simulated_bets — each tagged with the originating strategy so per-strategy ROI + CLV remain auditable on /performance.

2. Prediction Model

2a. Poisson Model

Goals for each team are modelled as independent Poisson processes:

A Dixon-Coles correction (ρ = −0.13) reduces systematic overestimation of 0-0 and 1-0 draws. Attack and defence parameters are estimated from the last 10 matches with an exponential decay weighting (recent matches count more).

The Poisson model is computed for every match regardless of data availability — it serves as the baseline and as the fallback for leagues where XGBoost has insufficient training data.

2b. XGBoost Classifier

A gradient-boosted tree trained on 95,000+ historical matches predicts 1X2 and Over/Under outcomes. Training features include:

XGBoost is only deployed for Grade A/B matches (European top leagues + secondary leagues with sufficient historical data). Grade C/D matches use Poisson only.

2c. Ensemble Blending

The final ensemble probability blends Poisson + XGBoost using a tier-specific alpha parameter:

Lower alpha in efficient markets (Premier League) reflects the bookmakers' superior information. Higher alpha in less liquid markets (lower league) reflects that our Poisson model likely captures information the bookmaker's line doesn't fully price in.

2d. Platt Scaling (Post-hoc Calibration)

After blending, a Platt sigmoid correction is applied per market using parameters fit on settled predictions:

α and β are fit by minimising log-loss on settled 1X2 predictions every Sunday. Requires 100+ settled predictions per market. This corrects systematic over- or under-confidence that survives the blending step.

3. Signal Collection (58 per match)

Separately from the model, the pipeline collects up to 58 signals per match into a normalised EAV (entity-attribute-value) table. Signals inform the Intelligence Summary on the match detail page and feed the meta-model (training, not yet live).

4. Bet Sizing (Kelly Criterion)

Stakes are sized using fractional Kelly:

5. Track Record & CLV

The track record page shows every settled prediction — no cherry-picking, no removing losses. The two metrics we use to evaluate model quality:

Positive CLV over a large sample is a stronger predictor of long-term profitability than raw hit rate. Results take hundreds of bets to converge; CLV signal appears much sooner.

5b. What Variance Actually Looks Like

Every honest value-betting service shows drawdowns. WinnerOdds — one of the longest-running tracked services in this space — publicly discloses a worst drawdown of −39%in October 2019. A model that has never had a losing month is either cherry-picking its public record or hasn't been running long enough.

Our current drawdown profile (sampled live from simulated_bets as of 2026-06-05):

Important caveat: our public chain only has ~35 days of settled bets so far. Drawdown numbers have meaningful statistical weight only after 500+ settled bets. Numbers above will move materially as the chain extends — by design. The chain is published live and never edited; you can inspect every individual bet at /performance.

Honest disclosure on verification: these numbers are self-reported. We are the ones running the database the chain lives in. Independent third-party verification — via Bet-Analytix or Smart Betting Club — is on the roadmap and intentionally gated until our v2-model cohort has stabilised, because the first verified number we publish becomes a permanent public anchor. We'd rather publish a verified honest number than an unverified flattering one.

Why we publish this: drawdown isn't evidence the model is broken — it's the cost of being a +EV bettor. CLV is the leading indicator of edge; drawdown is the lagging price you pay for short-term variance. Services that hide drawdowns are either lucky-sampled or fabricated.

6. What We Don't Do