How the predictions desk builds a call

The model

The model used by the predictions desk is a structured weighting system, not a machine-learning black box. The model takes six inputs: recent form (weighted 20 per cent), venue form (weighted 15 per cent), head-to-head (weighted 10 per cent), availability (weighted 20 per cent), opposition match-up (weighted 20 per cent) and conditions (weighted 15 per cent).

The weights are not arbitrary. The weights are calibrated against the previous 18 months of completed fixtures across the five sports the desk covers. The calibration is updated every quarter. The current calibration is v3, in use since the start of the year.

The inputs

Recent form is the win-loss-draw record of the side across the last five matches in the same competition. Venue form is the win-loss-draw record at the same venue across the last three matches. Head-to-head is the win-loss-draw record of the two sides against each other across the last five meetings.

Availability is the confirmed list of players named in the squad for the upcoming match, adjusted for the role weight of each player. A side missing its captain is downgraded more heavily than a side missing a rotation player. Opposition match-up is the structural comparison between the two sides — which side has the advantage at each position on the pitch.

Conditions are the weather forecast, the pitch report and the referee. The conditions input is the most volatile. The desk updates the conditions input twice in the 24 hours before kick-off.

The confidence labels

Every prediction carries a confidence label: Strong, Moderate or Lean. Strong means the model gives the call a probability of 65 per cent or higher. Moderate means the model gives the call a probability between 55 and 65 per cent. Lean means the model gives the call a probability between 50 and 55 per cent.

The labels are not marketing. The labels are the desk’s public commitment to the strength of the underlying evidence. A Strong call is one where the evidence converges. A Moderate call is one where the evidence is mixed but tilts one way. A Lean call is one where the evidence is split.

The verification

Every prediction is verified against the actual outcome within 24 hours of the match. The verification is published on the predictions page. The verification includes the model’s probability, the actual outcome, and the size of the gap between the two.

The verification is not edited. If a Strong call was wrong, the verification shows the call as wrong. The desk does not re-label calls after the fact. The desk does not retire predictions that did not land. The desk shows the full record.

What the desk does not predict

The desk does not predict individual player performance within a match. The desk predicts the outcome of the match. The desk does not predict the order of goals, wickets or points. The desk does not predict the minute of the first goal or the over of the first wicket.

The desk does not predict friendly matches. The desk does not predict matches where the availability list is not confirmed. The desk does not predict matches where the venue is not confirmed.

How to read a prediction

A prediction is not a tip. A prediction is a structured call with a named probability and a named confidence label. Read the inputs before you read the call. Read the conditions. Read the availability. The call is the last paragraph, not the first.