How our forecasts work

Every forecast is a set of probabilities made by statistical models from free, checked football data. It is published before kickoff, never edited, and scored in public once the result is verified.

Last updated 10 min read

What a forecast is

A forecast gives the chance of each result before a match: a home win, a draw and an away win, which always add up to 100%. It also gives the expected number of goals for each team and the chance of every exact score up to 10 goals each.

A probability is not a prediction of one result. When we give a home win 49%, we expect the home team to win about 49 of 100 matches like this one, and not to win the other 51. A good forecaster is right about how often things happen, not about every single match.

From raw data to a scored forecastThe same steps run for every match in every covered competition.
ForecastFC forecast pipelineBefore kickoff: free sources, checks, a frozen data snapshot, the models, then the published forecast at 24 hours, 3 hours and 45 minutes. After the match: result reports, two sources agree, the match is settled, the forecast is scored, and the score appears on the performance page.Before kickoffFree sourcesAPIs, open dataChecksNames, consensusSnapshotFrozen, hashedModelsElo, Dixon-ColesForecast24h, 3h, 45 minAfter the final whistleResult reportsFrom several sourcesTwo agreeFinal score verifiedSettledWithin 4 hoursScoredRPS, log loss, BrierPerformancePublic, every matchForecastFC forecast pipelineBefore kickoffAfter the matchFree sourcesAPIs, open dataChecksNames, consensusSnapshotFrozen, hashedModelsElo, Dixon-ColesForecast24h, 3h, 45 minResult reportsFrom several sourcesTwo agreeFinal score verifiedSettledWithin 4 hoursScoredRPS, log loss, BrierPerformancePublic, every match
View as a list
  1. Free sources. Fixtures, results, line-ups and context are collected from free sources whose terms allow it.
  2. Checks. Team names are matched to one identity and facts are compared across sources.
  3. Snapshot. The data the model may see is frozen at a cutoff time and given a hash.
  4. Models. Elo and Dixon-Coles turn the snapshot into a full score grid.
  5. Forecast. Published at 24 hours, 3 hours and 45 minutes before kickoff and never edited.
  6. Result reports. After the match, several sources report the final score.
  7. Two agree. The score is marked final when two independent sources agree.
  8. Settled. The stored forecast is matched with the verified result, normally within 4 hours.
  9. Scored. RPS, log loss and Brier score are computed for each forecast time.
  10. Performance. Every score, including the misses, is added to the public performance page.

Data we use

We only use free sources whose terms allow automated use on a public website, and we check those terms again every quarter. Each source, its licence and its current health are listed on Data sources.

  • Fixtures and results for every covered competition, going back more than ten seasons.
  • Goals and goal minutes, cards and line-ups from a licensed free feed.
  • Availability facts: suspensions we compute from card counts and each league's rules, and injuries only when an official source or two independent sources agree.
  • Schedule: rest days, matches in a short period, travel distance between stadiums and international breaks.
  • Weather at the stadium, with a low weight. It matters more in V.League 1 than in Europe.
No betting odds and no expected goals (xG) from other companies. We do not collect, store or use betting odds anywhere, and no free source licenses xG for a public site. Forecasts come from football data alone.

Models

All models are trained and run on ordinary servers without graphics cards. A forecast takes milliseconds once the data is ready, so every match gets the same care.

Elo rating

We compute our own Elo rating for every club from results. A win against a strong side moves the rating more than a win against a weak one. The rating includes a home advantage and moves part of the way back to the average between seasons.

Dixon-Coles with time decay (current model)

The Dixon-Coles model estimates how strong each team is in attack and in defence, and from that the expected goals for each side. It corrects the plain Poisson model's known weakness with low scores such as 0–0 and 1–1. Recent matches count more than older ones: their weight fades with time at a rate we tune on past seasons. The model produces the full grid of exact scores, and the home, draw and away chances are sums over that grid.

Challengers

New models compete with the current one on the same past seasons, match by match, without seeing the future. Candidates include Dixon-Coles with Elo-adjusted strengths and a gradient-boosted model that predicts goal rates from schedule and availability features. A challenger replaces the current model only when it scores better with a confidence interval that excludes zero and no league gets worse beyond a set tolerance. The switch is a manual decision that is logged, and the new version name appears on every forecast it makes.

Calibration

Raw model chances can be slightly too confident or too cautious. A final calibration step, fitted on past seasons the model did not train on, adjusts them so that events we call 30% happen about 30% of the time. The calibration chart shows how well this holds.

Baselines we must beat

Before a model is used at all, it has to beat two simple baselines over at least three past seasons in every league: the league's long-run share of home wins, draws and away wins, and a forecast from Elo alone.

Forecast times

Each match gets up to three forecasts. Each one uses only data that existed at its cutoff time.

Free

24 hours before

Ratings, form, schedule and known suspensions. Home, draw and away chances, expected goals and the three most likely scores.

PRO

3 hours before

Adds injuries and suspensions confirmed since the day before, and the expected line-ups.

PRO

45 minutes before

Adds the confirmed starting line-ups and how they differ from each team's usual eleven.

If a forecast cannot be made in time, for example because a source is down, the match page says that forecast was skipped and why. We never publish a late forecast in its place. When a match is postponed, its forecasts are withdrawn and new ones are made before the new kickoff.

Never edited, always traceable

Forecasts are stored once and never changed. A new forecast time adds a new record; it does not overwrite the earlier one. Every forecast carries three labels you can see on the match page:

  • Model version: which model and release made it.
  • Data as of: the cutoff time of the data the model was allowed to see.
  • Snapshot hash: a fingerprint of the exact inputs. The same inputs and model always give the same forecast.

Automated tests feed the system data that appeared after a cutoff and check that no earlier snapshot or forecast changes. This is how we know a forecast did not see the future.

Pro members see each probability to one decimal place with a small pattern unique to their account: the last digit can differ from normal rounding by at most 0.1 percentage point, and the three results still add up to 100%. It lets us trace copies made in breach of our terms. The stored forecast never changes, and every score on the performance page uses the stored values.

How we score forecasts

Once two independent sources agree on the final score, the match is settled and each of its forecasts is scored. We use three standard measures. For all three, lower is better and a perfect forecast scores 0.

  • Ranked probability score (RPS), our main measure. It respects the order of results: calling a home win when the match was drawn is a smaller miss than calling a home win when the away team won.
  • Log loss looks only at the chance we gave the result that happened, and punishes confident misses hard.
  • Brier score adds up the squared gaps between each chance and what happened (1 for the actual result, 0 for the others).

Worked example

Arsenal beat Chelsea 2–1. Our 24h forecast gave a home win 49.0%, a draw 26.6% and an away win 24.4%.

RPS compares running totals in order. The chance of "home win" was 0.490 and the chance of "home win or draw" was 0.490 + 0.266 = 0.756. Both events happened, so each target is 1:

RPS = ((1 − 0.490)² + (1 − 0.756)²) ÷ 2 = (0.260 + 0.060) ÷ 2 = 0.160

Scores of the 24h forecast, lower is better
MeasureActual result: home win 2–1If Chelsea had won
RPS0.1600.406
Log loss0.7131.411
Brier score0.3900.882

The same forecast would have scored much worse if Chelsea had won, because it gave that result only 24.4%. One match says little: the performance page shows averages over every settled match, including the misses, next to the two baselines and with the number of matches behind each figure.

How explanations are written

The short explanations on match pages are not written by AI. Each sentence is a fixed template, translated by people, and filled with checked facts. A sentence appears only when all of its facts exist; if a fact is missing, the sentence is left out rather than guessed.

Template

{team}'s rating is {points} points above {opponent}'s once home advantage is included.

On the match page

Arsenal's rating is 41 points above Chelsea's once home advantage is included. Source: ForecastFC Elo, updated after every match

Each sentence links back to the facts and sources it used. Editors can add a short note to a match; it is labelled as an editor's note.

Limits

  • A forecast is a probability-based estimate, not a guarantee. Unlikely results happen every week, and they should.
  • Events after the last forecast, such as a red card, an injury in the warm-up or the weather turning, are not in it.
  • We have no shot or expected goals data, so a team that creates many chances but scores few can be underrated for a while.
  • Line-ups come from one licensed feed, and injury news is used only when it is official or confirmed by two sources, so late news can be missing.
  • Free sources can be late or down. When a competition has fewer than two healthy result sources, settlement waits rather than guessing.
  • V.League 1 has the thinnest free data. Its forecasts appear only after it passes the same data and backtest checks as the European leagues; until then its pages show fixtures, results and tables.
  • ForecastFC is football analytics. It is not betting advice and must not be used as such.

Spotted something wrong? Report a data error.