Red Half Lab: a Liverpool supporter's data project

3. The team and player data

The current REDS match system has 15 tasks: 2 for the team, 10 for player contributions and 3 for goalkeepers. Its report directory also provides match context, including recent form, head-to-head results and upcoming fixtures.

The team score model supports several related views, including the score, result and clean-sheet outlook. Those views come from the same score distribution; they are not separate models merely because they appear as different questions.

Player coverage includes Alisson, Virgil van Dijk, Jeremie Frimpong, Florian Wirtz, Dominik Szoboszlai, Ryan Gravenberch, Cody Gakpo, Alexander Isak. The saved directory has 32 profiles, covering established players, newer arrivals and younger players in the project record.

The fan notebook contains 32 questions. Examples include Van Dijk's duels across different periods, Wirtz and Szoboszlai in the number-ten role, and Frimpong at right-back and in a right-sided attacking role. These comparisons give the project a way to investigate specific football questions beyond a single match.

Coverage varies. A profile does not mean every task has a supported answer for that player, and some saved questions use older evidence. Appearance conditions, positional references and evidence dates remain part of the individual report. The player directory provides those details.

4. Progress and baseline checks

REDS has progressed from model development to a working interface for selecting questions, reading match and player reports, and following post-match checks. The content service supplies the selected answer and its chart instructions; the website presents that report.

The current system contains 15 match tasks. Of these, 14 have passed the current label-quality checks on the outcomes used to train and evaluate them. Attacking contribution remains experimental because those checks are unresolved for that task. The task count therefore describes the system's scope, not fifteen equally established findings.

A simple baseline provides a reference for judging whether the model adds anything. The two examples below come from the stored historical forward checks: the score-and-result task was checked across 444 matches, and the starting task across 555 player appearances. In the latter, the question is whether a player starts given that the player appears.

Both stored model scores are lower than their baselines. The score-and-result advantage is modest, while the starting comparison shows a larger difference on its own task. These results support continued development; they do not establish a live success rate or guarantee the next forecast.

New evidence also has to be checked before a model changes. A candidate that fails the replacement criteria leaves the current model in place. Missing post-match evidence remains pending, and later improvements belong in future forecasts without rewriting earlier records.

The next stage is to build a longer record of genuine pre-match forecasts and later checks, and to improve reports where the available evidence is incomplete. The full baseline report and development log keep those comparisons and changes accessible.