Start with the match, not a confident headline
A scorecard reports what has happened. A model estimate describes uncertainty before a match. FairPlay keeps these separate: source-confirmed scores belong to the match centre; statistical context belongs to the analysis view. A blank estimate means the evidence did not meet the model’s rules, not that the missing number can safely be guessed.
The internal model version is elo-v1-preview. It is an experimental, read-only method. It does not use a player profile as proof of selection, predict an exact score or replace a confirmed playing XI.
What goes into the calculation?
The analysis loader uses up to 24 months of normalized, decided match history. Each input needs an identifiable fixture, different home and away team IDs, a valid start time, a supported format group and a winner that matches one of those teams. Duplicate records are removed and results are processed in date order. Draws, no-results and unsupported formats are not treated as wins.
Short-format and one-day records are rated separately. The model starts each side at 1,500 and changes its rating after a result using an Elo update with K = 24 and a base-10 scale of 400. It applies a +30 adjustment to the first-named/home side. That is a declared prototype assumption, not evidence that every nominal home side has an actual venue advantage.
When are estimates withheld?
Each side needs at least eight prior rated results in the relevant format group. Confidence is pulled toward 50% until the side with fewer results has 20. The analysis view also requires a fresh source check and a passed validation result for that format; if history is unavailable, the check is delayed or either side lacks the minimum history, the estimate stays withheld.
A team name alone is not enough to join historical identities. Changing squads, neutral venues and inconsistent provider labels can limit the evidence even when the schedule is visible. Player form, pitch preparation, weather and late availability are not automatically solved by this team-history model.
How is validation assessed?
The model uses a chronological 70/30 history split. In the holdout portion, an eligible result is predicted before that result updates the ratings. The model records Brier score, log loss and classification accuracy. Its current pass gate requires at least 100 tested matches and a Brier score below 0.25. A pass describes that sample; it does not guarantee future performance or prove the method beats every simpler baseline.
Recent form and head-to-head panels use their own labelled result windows. Those panels should be read as limited samples, not as all-time records or an instruction to wager. Open the contributing scorecards to verify the underlying fixtures.
External back/lay rates are a different source
The FairPlay Rates desk displays quotes from the supplied external endpoint. Those quotes are not created by this Elo model and should not be confused with its probability estimates. The desk labels source status and update times and withholds unavailable or suspended prices. Rate cells have no bet-placement, wallet or deposit action.
Read responsibly and report errors
No model can make a cricket result certain. Use the analysis as an explanation of the evidence that is available, then verify the match status, date and lineup separately. For a wrong team identity, score or source attribution, use our corrections page. For provider freshness and the difference between cached and current observations, read the sports-data method.