RESEARCH / FORMERLY PROJECT B
Match Pulse.
Liverpool’s form, cards and match style. In research, with a target release in Q1 2027.
THE RESEARCH RECORD
Follow the work before the launch.
We are testing what can be learned about single-match form, three-match form, red and yellow cards, and patterns of play. We continue to improve the models and check new matches. The release date is an estimate, subject to review.
Historical replays use rebuilt data and are separate from genuine forecasts recorded before kick-off. Live reliability is not established. Counts below belong to individual research tracks and overlap across matches.
| Research track | Historical replay | Simple baseline | Historical confidence | Real pre-match checks |
|---|---|---|---|---|
| Three-match formHistorical forward replay; within ±6 state points | 8/10 hitsHit criterion stated in the track definition | 3/10 hitsMost frequent state band in the training data | 58%Mean probability of landing within ±6 points | 0/1 complete1 awaiting outcomes or labels |
| Single-match formHistorical forward replay; within ±6 state points | 5/10 hitsHit criterion stated in the track definition | 3/10 hitsMost frequent state band in the training data | 41%Mean probability of landing within ±6 points | 1/1 complete0 awaiting outcomes or labels |
| Match styleHistorical forward replay; next-match class | 3/10 hitsLog loss 1.269 | 4/10 hitsLog loss 1.458. Training-period class frequencies | Not reportedNo comparable confidence figure published | 1/1 complete0 awaiting outcomes or labels |
| Red-card riskLatest historical replay; labels required before scoring | 0 scoredAwaiting labels | Not scored yetTraining-period league event rate | Not reportedNo comparable confidence figure published | 1/2 complete1 awaiting outcomes or labels |
| No-yellow probabilityLatest historical replay; labels required before scoring | 0 scoredAwaiting labels | Not scored yetTraining-period league event rate | Not reportedNo comparable confidence figure published | 0/1 complete1 awaiting outcomes or labels |
| Yellow-card countLatest historical replay; labels required before scoring | 0 scoredAwaiting labels | Not scored yetTraining-period league mean count | Not reportedNo comparable confidence figure published | 1/2 complete1 awaiting outcomes or labels |
How to read the comparison
Three-match windows share matches, so ten windows are not ten independent trials. Historical confidence is the probability the model assigned in those replays. It is not a live success rate.
For match style, compare log loss alongside top-class hits. Lower log loss can accompany fewer hits; both are shown. The groups describe learned behaviour and have not been independently verified as tactical labels.
What is still being checked
Cards are scored only when labels are available. Missing comparisons stay unreported until a valid sample is available. Genuine pre-match checks are still too few to establish reliability.
A controller heartbeat confirms that the local process checked in. It does not mean the evidence or scores changed. Records over 24 hours old are marked in the timestamp row.
RESEARCH STATUS DATA
A small, inspectable dataset.
Download the aggregate record behind this page, including its timestamps and comparison rules. The dataset is shared under CC BY 4.0. Match forecasts, raw provider data and model files remain private.
Download status JSON ↗The verified research snapshots are now available on GitHub ↗. Each snapshot includes its evidence date. Local development reads the verified research record; a production build serves a dated snapshot.
We plan to open-source the models on GitHub once they are stable and have passed review.