Methodology

The mechanism behind every claim on the landing page — model, tiers, comparison math, verification stack, code.

The model

Two predictors blended at training time:

• A Poisson goal-rate model (Dixon-Coles correction for low-scoring outcomes, per-league rho fitted from historical scoreline frequencies)
• An XGBoost gradient-boosted classifier trained on form, ELO, expected-goals, lineups, injuries, market movement signals
• Blended weights are optimised on a rolling holdout; weights refit weekly via workers/jobs/weekly_blend_refit.py

Retrained weekly on every completed match across 280+ leagues via workers/jobs/weekly_retrain.py (Sunday 03:00 UTC). Promotion stays manual — a new model only goes live after a comparison report (scripts/compare_models.py) and the operator flips the MODEL_VERSION env on Railway.

Calibration

The raw model outputs are not betting probabilities yet — they tend to be over-confident at the tails. We run two calibration stages:

• Shrinkage toward Pinnacle (sharp-book anchor) scaled by data tier
• Platt scaling per market (1X2, O/U 2.5, BTTS, AH, DC) fitted on settled bets; 2-feature logistic for O/U 2.5
• Isotonic regression as the final layer (one-time fit during training)

Platt weights refit weekly viascripts/fit_platt_live.py.

Bot strategies

Production runs ~16 paper-trading bots simultaneously, each with its own market / league / edge-threshold filter. They share the same prediction backbone; the filters differ:

• Calibrated — promotion-grade. The only tier that places real money via the Coolbet placer. ~6 bots today.
• Beta — paper-only, accumulating live evidence after a backtest passes. Auto-promoted by scripts/weekly_bot_review.py at ROI ≥ +10% / CLV ≥ +5% on 60+ settled.
• Active — paper-only, observation tier.
• Retired — failed experiments. Excluded from headline numbers but kept in shadow_bets so they continue to be evaluated — reactivation gated on shadow performance recovery.

How pick rows are computed

Every settled bet on the public ledger represents one row in our simulated_bets table. The pipeline writes a row before kickoff with:

• match_id + kickoff_utc — the fixture
• market + selection — what we're betting
• odds_at_pick + recommended_bookmaker — the price + book
• edge_percent + calibrated_prob — the model's edge estimate
• placed_at_utc — timestamp of the row insert

After the match completes, settlement adds result, pnl, closing_odds, clv, clv_pinnacle. All historical rows are immutable — we never edit a settled bet.

How competitor comparisons work

For each competitor we pull their public per-bet history, filter to the same time window + same markets + same settlement state, and compute ROI assuming a €10 flat stake on every side. Scripts live in the engine repo:

• scripts/production_audit_vs_winnerodds.py — WinnerOdds public GraphQL endpoint atapp.winnerodds.com:4000
• scripts/audit_vs_signalodds.py — SignalOdds paginated HTML at signalodds.com/predictions/past
• scripts/audit_vs_deepbetting.py — DeepBetting /backend/api/predictions-api.php?type=stats

The output of each audit lands at ledger/comparison_*.json — same hash committed by GitHub Actions, anyone can clone the repo and re-run.

Set your own stake

The landing comparison uses €10 flat as a clean apples-to-apples unit. Drag the slider below to see what the same matched-window ROI translates to at any per-bet stake.

Why flat stake, not Kelly

Conventional wisdom says Kelly sizing maximises bankroll growth when your edge estimate is well-calibrated. We checked. On the same n=1,181 production-cohort pre-match sample since 2026-05-04:

• Flat €10 stake — ROI +9.08% (€612 on €6,741 staked)
• Kelly-weighted — ROI +7.31% (same picks, sized by model edge × bankroll fraction)

Flat beatsKelly by ~1.8pp on our current sample. Reason: Kelly sizes up the highest-edge picks, and our highest-edge picks are exactly where the model's known longshot miscalibration hurts most (see "Conditional miscalibration at high odds" in MODEL_WHITEPAPER §11). Kelly amplifies the bets we're most overconfident on.

We surface flat €10 as the publishable headline because (a) it honestly matches the comparison baseline, and (b) it's currently our best strategy. Once the calibration fix lands (CAL-ALPHA-ODDS), we'll re-run this comparison — Kelly should win after the high-odds correction.

Competitor stake methods, as best we can tell from their public surfaces: WinnerOdds publishes "unit" stakes (Kelly-like internally); SignalOdds shows EV%-weighted; DeepBetting shows per-confidence-tier flat; Tipstrr shows "level stakes" (flat); Forebet doesn't publish stake — we apply €10 flat at our end to settle their picks. Translation: our €10-flat comparison is internally consistent on OUR side; per-competitor stake differences add minor noise we can't correct for from outside their books.

Per-competitor caveats

Each competitor row on the landing is the headline number. A few have noteworthy structure worth knowing:

• Forebet — algorithmic predictor with a free public history. Their +15.33% headline is driven entirely by Over/Under 2.5 (+35.67% on n=611 in this window); their 1X2 picks alone return essentially break-even at +0.23% on n=823. The OU lift is consistent with a low-scoring stretch of European fixtures during this audit window — likely to mean-revert as the new season starts. We show the headline honestly; the OU dependency is what to watch.
• Tipstrr — only monthly aggregates are public, per-bet detail is paywalled. The audit pools three active football tipsters (star-tips, mls-value, main-draws-model-top-euros) and covers all football bet types (1X2 + O/U + AH + BTTS) — slightly wider than ours (1X2 + O/U only). The wider scope works in their favour; even so, they're at -5.22% in the matched window.
• Betaminic — public strategies list is auth-walled (free signup gate plus paid tiers). Status isauth_required in ledger/comparison_betaminic.json with the activation lever documented (BETAMINIC_COOKIE env). Not displayed on the landing comparison until we have credible data.
• WinnerOdds + SignalOdds + DeepBetting — fully public per-bet history, no paywall, scope notes pinned to the same 1X2 + OU 2.5 cohort as ours.

Why the comparison count differs from the hero count

The landing page hero shows 1,181 settled bets at +9.08% ROI— that's our full production cohort across 1X2 + O/U 2.5 + BTTS since the calibrated tier launched.

The comparison card on the landing shows 989 bets at +11.91% ROI — a subset, because SignalOdds and DeepBetting both publish 1X2 + O/U only, not BTTS. To stay apples-to-apples we drop BTTS from BOTH sides. Dropping BTTS actually lifts our ROI from +9.08% to +11.91% because our BTTS model is weaker than 1X2/OU (see Known Limitations in MODEL_WHITEPAPER.md). Both numbers are real; they measure different cohorts.

Why the window starts 2026-05-04: that's the day our calibrated bot tier launched. Going further back would mix in pre-calibration bets we don't deploy real money on. The window will grow as more weeks of live evidence accumulate.

Verification stack

• Live API — /api/v1/track-record(settled) and /api/v1/upcoming (live picks). Public, no auth, JSON, CORS-open.
• Daily GitHub commit— every night at 22:45 UTC, a deterministic JSON of the day's settled bets is committed to ledger/YYYY-MM-DD.json and signed by github-actions[bot].
• Bitcoin blockchain anchor via OpenTimestamps. Each daily JSON is hashed and committed into a Bitcoin block. After ~1-6 hours the proof is permanent — anyone with ots verify can confirm "this exact JSON existed at this Bitcoin block height" without trusting GitHub or us.

Source code

Both repos are open source on GitHub:

• msellin/odds-intel-engine — the model, training, scheduler, scrapers, ledger
• msellin/odds-intel-web — this site, the picks page, the API routes

What we're not

This is a research project that publishes its track record. It is not financial advice, not a guarantee of future returns, and not a substitute for your own judgment. Past results are real (every row is in the public ledger), but the future is variance. Bet responsibly.