Methodology

1. Data sources

Match results, scores and historical seasons come from the open football-data.co.uk CSVs that are ingested weekly. Fixtures, kick-off times, live results and market odds are synchronised from ESPN. Chance quality comes from real expected goals (xG) scraped from FlashScore and imported FootyStats/FBref files — we distinguish real provider xG from fallback estimates, and every public page states which one is being used.

Data freshness is reported on the site banner (xG coverage, last xG match, stale/fresh status). If a source is unavailable, the affected stage is skipped rather than filled with invented numbers.

2. Team ratings

Every club carries two walk-forward rating systems: an Elo rating with margin-of-victory scaling and tier multipliers, and a Glicko-2 rating with a rating deviation, which makes the model care about how certain a rating is. Ratings are computed from every finished match in order, so each snapshot only ever uses information available before kick-off — no look-ahead.

3. Match model

Fixtures are modelled with a Dixon-Coles bivariate Poisson model (attack/defence parameters, home advantage, a low-score correction rho constrained to a stable range) wrapped in an XGBoost ensemble that also sees market-implied probabilities, xG form, rest days, venue splits and head-to-head results. The ensemble is walk-forward validated and probability-calibrated, and every published probability is accompanied by a per-match “Why this prediction?” breakdown on the fixtures page.

4. Season simulations

Every remaining fixture is replayed thousands of times per run (20,000 simulations in the weekly refresh, plus a lighter daily run so probabilities move after midweek games). Each simulation samples a scoreline for every fixture from the match model, applies Greek Super League rules, and produces a final table. Dixon-Coles parameter uncertainty is captured with a parametric bootstrap, and a Glicko rating blend is applied to outcome probabilities (conditional scoreline shapes stay Dixon-Coles).

The Greek two-phase format is simulated exactly, because it changes what “a good season” means:

Projected points are the mean across simulated seasons; the “10–90%” column is the band the middle 80% of simulated seasons land in. Extreme best/worst outcomes are available in tooltips but grow with the number of simulations, so they are never presented as the expected range.

5. Targets and race charts

Probability checkpoints are recomputed as of each completed matchday, replaying the remainder of the season with only the information available at that point. That is what powers the title-race and relegation trend charts, the matchday movers, and the deltas shown next to every probability.

6. Honesty about uncertainty

Outputs are statistical probabilities, not predictions of certainty, and nothing on this site is betting advice or a tip. We publish model performance openly on the track record page — hit rates, log loss, Brier score and probability calibration — including periods where the model underperformed. When data is missing or a source is stale, the site says so instead of guessing.

Club names and crests are used for identification only and remain the property of their owners. This site is not affiliated with any league, club or governing body.