How SportSignals Publishes Football Predictions

What does the current system publish?

The production integration retrieves fixture prediction probabilities from the Sportmonks Football API when they are available. Sportmonks documents winner, correct-score, over/under and both-teams-to-score probabilities and notes that some fixtures do not have enough data to be marked predictable. See the Sportmonks prediction documentation.

SportSignals turns available provider output into a consistent product view. Coverage therefore depends on fixture mapping, provider availability, market mapping and current odds data. A missing prediction is not silently replaced with a claim of equal confidence.

Who is responsible for each part?

How are team and player pages built?

Team and player identity, competition, standings, fixture and match-stat records are ingested from mapped football-data feeds, primarily Sportmonks. SportSignals normalises provider identifiers, removes placeholder records and reconciles aggregates against the match rows rendered on the public page.

Coverage varies by competition and provider feed. Each entity section states its season or date range, sample size, source class and freshness where available. Corrections are applied to the canonical page rather than published as a separate crawler-only version.

How does a probability become a pick?

1. Match and team identifiers are matched to a covered fixture.
2. Available provider prediction types are read for that fixture.
3. Those types are mapped to SportSignals market and outcome labels.
4. Probability units are normalised for the application.
5. Eligible current prices are selected from available odds data.
6. The model probability is compared with raw price-implied probability.
7. After settlement, an eligible published pick can enter the accuracy record.

This is an integration, comparison and reporting pipeline. It is not evidence that SportSignals independently trained the provider's machine-learning models.

How is the probability difference calculated?

Raw implied probability is 1 / decimal odds. In an illustrative example, decimal odds of 2.50 imply 40%. If the supplied model probability is 48%, the difference is 8 percentage points.

That comparison does not remove the full market margin and does not prove profit. The model can be wrong, the price can move, the market mapping can differ from settlement, and the quoted price may not be accepted.

How do the separate team ratings fit in?

SportSignals also exposes a separate overall team rating and specialist indices for both teams to score, over 2.5 goals, attack and defence when rating records are available. The current read-side implementation labels the underlying method Elo-based and can display match counts, update times and rating history. These fields support team comparison; they are not evidence that the ratings generated the provider-supplied fixture probabilities.

When a comparison row is absent, the current read path uses 1500 as a neutral fallback. That display behaviour does not establish that the write-side process initializes every new team at 1500. The current public and read-side code also does not substantiate an exact update equation, K-factor, home adjustment, goal-margin multiplier or seasonal regression constant, so SportSignals does not claim those values here.

FIFA's men's ranking procedure is an official example of an Elo-style method with its own documented components. It illustrates the method family; it is not the SportSignals equation. See the verified team-rating scope for the maintained field-level boundary.

How accurate are settled SportSignals picks?

The figures below come from the same settled-results queries used by the public accuracy page. Read each percentage with its denominator and market type.

How should probabilities be evaluated?

A headline hit rate is not enough. Check sample size, market type, calibration, scoring rules, price availability and time order. The scikit-learn calibration guide explains how forecast probabilities should correspond to observed frequencies. Its TimeSeriesSplit guide explains why later observations should be used to test a process trained on earlier observations.

What does this methodology not claim?

SportSignals does not currently claim a proprietary XGBoost, LightGBM or neural-network ensemble; daily in-house model retraining; human analyst approval of every pick; direct Opta or StatsBomb prediction-model partnerships; a fixed accuracy range; guaranteed value; or profitable outcomes.

Any future architecture or performance claim must be supported by maintainable internal documentation and reproducible public results before it appears here.

How should the information be used?

Treat each probability as an uncertain estimate, not a command. Check the market wording, current price and your own constraints. A high displayed probability can still lose, and a positive difference can disappear after model error, margin, movement or settlement differences.

Sources and evidence boundaries

• Sportmonks prediction probabilities supports provider attribution, markets and predictability metadata.
• Probability calibration supports the interpretation of probabilistic forecasts.
• TimeSeriesSplit supports time-ordered validation guidance.
• FIFA men's ranking procedure supports the generic description of an Elo-style method; it is not cited as the SportSignals update equation.

Sources checked 14 July 2026.

Frequently asked questions