How we calculate it 🧠

Each national team has an ELO (relative strength, adjusted with international history). The ELO difference + the host's real home advantage is converted into expected goals (λ) for each team.

With those λ we build a Poisson matrix (probability of each exact scoreline) and sum: home win, draw and away — plus Over/Under 2.5.

Backtest: log-loss 0.97 vs 1.10 for chance and 1.05 for the baseline. Near-perfect calibration by buckets.

Each team has an offensive rating (how many corners it generates) and a defensive one (how many it concedes), estimated by normalizing for the opponent's strength — not raw averages, which mislead.

Expected match corners → Poisson over the total → probability of each Over/Under line.

We model two things per team: how much indiscipline of its own it has (cards it receives) and how much it provokes in the opponent.

Expected total ≈ K·(atk_home·def_away + atk_away·def_home) → Poisson → Over/Under. Validated out-of-sample (trains on the old, tests on the recent).

From the model's probability we get the fair odds = 1 ÷ probability. If a bookmaker pays more than those fair odds, the market mispriced it: that's where the value is.

value = house_odds × model_probability. We only flag a pick when value ≥ 1.06 (≥6% mathematical edge). Below that, it's not worth the risk.

Compares the odds you bet at vs the market's closing odds. If you got a better price than the final market (CLV+), your edge is real — it's what best correlates with winning over the long run; match by match is noise.

A sharp line move is usually professional money coming in. We monitor it: if the line moves toward your pick, it confirms the edge; if it moves against, it reevaluates it.

The size of each bet comes from the Kelly criterion (optimal proportion given your edge and the odds), in a fractional version to smooth variance. Discipline over gut feeling — that's how a real edge turns into money without blowing up the bankroll.