Methodology: how start probabilities are made

How we estimate who starts, how we update after news, and how we measure our own accuracy.

Start probability (p_start)

For each player we estimate the probability that he is in the starting XI for the next league match. Inputs include recent league starts and minutes, published predicted line-ups from several independent sources, injury and suspension lists, fixture congestion (a cup or European match within three days) and news from press conferences and official club channels.

Probabilities within a team are balanced so that they add up to about eleven: exactly one goalkeeper and ten outfield players. Until independent predicted line-ups are available, model-only probabilities are capped at 89%: a 90%+ rating needs supporting evidence such as a predicted XI or a press-conference confirmation.

Expected minutes

Expected minutes = p_start × typical minutes as a starter + (1 − p_start) × chance of coming off the bench × typical minutes as a substitute. A player who is 80% to start and usually plays 85 minutes gets about 70 expected minutes.

Rotation risk bands

• Low: start probability 85% or more.
• Medium: 60–84%.
• High: 30–59%.
• Unlikely: below 30%.
• Out: ruled out by the club or injury lists.

Where we have a specific reason, we show it.

Updates and sources

The model re-runs at scheduled times around each gameweek (currently 07:00, 11:00, 13:00, 16:00, 18:00 UK time) and immediately after major press conferences. Each change is logged with one of these source types:

• Press conference – what the manager said, linked to the conference.
• Official club channel – club website or official social accounts.
• Journalist report – a named reporter with a track record on that club.
• Training photos – who was pictured in training.
• Model update – the model re-ran on new data, e.g. a new predicted line-up signal, an injury-list change or a recalibration.

Before matchday we publish an early forecast; on matchday a matchday model takes over, and when the official team sheet is out we mark the line-up as official. Switching between these is itself logged.

Calibration

Raw model probabilities are recalibrated against previous seasons and earlier gameweeks so that "70%" means roughly seven in ten such players start. Calibration is reviewed as more gameweeks are evaluated and the version in use is recorded with every run.

How we score ourselves

See the accuracy page. In short: we take the last prediction published before kick-off and before the official team sheet, and compare it with the real starting XI (correct of 11, Brier score, log loss, calibration and minutes error). Every team we predicted is included.

Data we do not use or publish

We do not use data from the official fantasy game, and we do not publish ownership percentages, price-change predictions or betting odds. LikelyXI is independent and not affiliated with the Premier League or Fantasy Premier League.

Last reviewed 28 October 2026.