Notebook FCSearch
eve∴
A field of
possible futures.

What could happen when these teams meet?

Houston Dash v Washington Spirit

Club pairingModels trained to 15 Dec 2024Houston Dash attacking

256retained futures
20 contain a shot
Transformer V1, new combinations of playAttacking →
Selected futureShot
20 / 256 draws match this filterAt least one shot event, own goals excluded.

Selected possession 01

Ends with a shot in the sequence

3passes1shot events5actions

Ask these 256 draws

Passes per draw

Transformer V1, darker where the current search matchesTransformer V2, same context

The selected path has 3 passes; 95 of 256 draws have more. Shares of saved draws, not forecast probabilities. Choose a bar to open a matching draw.

The whole simulated match

Every possession, counted

256 saved possessions for each team from Transformer V1. Each figure is a share of those draws, with its 95% sampling margin (±, in points); none is a forecast probability.

Home attacking

Houston Dash

Possessions with a shot
7.8%± 3.3history 9.8%
Reach the box
18.4%± 4.7history 23.0%
Goal events
0.4%± 0.8
Actions per possession
4.1

Away attacking

Washington Spirit

Possessions with a shot
8.2%± 3.4history 10.9%
Reach the box
18.8%± 4.8history 21.1%
Goal events
2.3%± 1.9
Actions per possession
4.1

Where each team wins the ball

Houston Dash

Washington Spirit

Where possessions begin, both teams attacking right. Houston Dash start 7% of possessions in the final 30% of the pitch; Washington Spirit 9%.

Where the shots come from

22 Houston Dash shots attacking right, 23 Washington Spirit shots attacking left. A ringed dot is a goal event.

How each team moves the ball

Houston Dash

Washington Spirit

Passes between 20 zones. Forward moves in the team colour, backward moves pale, sideways moves white; line weight is the share of all passes.

How long possessions run

Houston DashWashington Spirit

Share of possessions by number of actions. Most end within three.

Longer possessions shoot more

Houston DashWashington Spirit

Of the possessions with each length, the share with a shot. Gaps mark lengths with too few draws to rate.

Where play slows down

Houston DashWashington Spirit

Mean seconds between actions in each zone. Wide areas of the final third are usually slowest.

How each method sees this match

Houston DashWashington Spirit

Share of possessions with a shot under each way of making possibilities, with its 95% sampling interval. Bars whose whiskers overlap cannot be told apart.

Every possession at once

1 still going 20 have shot

All saved possessions for one team, unfolding together one action at a time. The newest action is brightest; yellow dots are shots, white rings goal events.
Inspect the actions ↘
Selected possession
StepSelected future
10s Keeper pickup at 12.5, 47.5
21s Pass, complete at 12.5, 57.5 → 22.5, 42.5
32s Pass, complete at 27.5, 42.5 → 32.5, 37.5
46s Pass, complete at 32.5, 32.5 → 47.5, 2.5
512s Shot saved or blocked at 82.5, 32.5
What EVE is, and where this lab fits ↘

EVE simulates football as sequences of events. A transformer predicts the next action from the ones before it, the action is appended, and it predicts again. Earlier EVE models generated whole NWSL matches this way. This lab works one possession at a time, so every step can be drawn, measured and improved, in all three leagues.

Read it at three scales: an action (where a pass starts and ends, or where a shot is taken), a possession (the sequence until it ends), and a population (many sampled possessions, the faint field and the bars above).

What this page does today
OperationHereStatus
Search saved drawsExplore, and the questions aboveWorking. A search changes what is shown, not the model's probabilities
Continue from a real beginningWhere futures divergeWorking on saved continuations from 18 real possessions
Generate for a match contextPossession futuresSaved draws for 812 pairings of learned clubs and three retained fixtures; choosing clubs loads a saved collection, it does not run the model
Change an assumptionNot offeredNeeds a model that takes lineups or tactics as inputs and tests of its response

Toward unplayed matches, a design not yet running

  1. Open a future fixture and see the context the model knows before kickoff. Needs a new export or service
  2. Draw a versioned set of possessions for that context. Needs a new export or service
  3. Search, compare and follow paths. Working today on saved draws
  4. Inspect the model and assumptions behind each path. Partly here: source, sample size and the evidence dialog
  5. After the match, compare the frozen pre-match set with what happened. Needs a new export or service

Whole simulated matches need more than possessions placed end to end: a model of how possessions, scores and the clock evolve together. That is modelling work still ahead.

Saved draws from models trained on 2022 to 2024 matches, to 15 Dec 2024: not current form, lineups or a whole match. Each draw starts from the restart and location of a training possession; the transformer generates every action from there, the first included. Neither transformer cleared the improvement it was required to show over both baselines. Event locations, not player tracking. Playback traces actions at a visual pace; it is not a reconstruction of ball speed. Violet paths show up to 40 matching alternatives.

EVE / Possession labExperimentalSaved model draws

The test of a possible future

Does the model know the game?

One plausible possession proves very little. We test the distribution of possibilities against what happened across 182 NWSL matches, with 512 sampled possessions per side.

Pre-match errors, lower is better
ModelPass count (CRPS)Shot chance (Brier)Ending (Brier)
Historical paths2.120840.093490.19064
Transition model2.141660.093540.19078
Transformer V12.125000.093560.19083
Transformer V22.125260.093540.19076

V2 is a useful experiment. It is not a clear win.

V2 improves local spatial likelihoods. None of the primary pass, shot or ending comparisons establishes an improvement over V1. Futures still tend to be too short, and the gaps between actions can be too large.

  • Pass count (CRPS): V2 minus V1 0.00026; 95% interval [-0.00212, 0.00260].
  • Shot chance (Brier): V2 minus V1 -0.00003; 95% interval [-0.00014, 0.00009].
  • Ending (Brier): V2 minus V1 -0.00006; 95% interval [-0.00030, 0.00016].

These are reused development matches, not an untouched final test. Intervals use 2,000 paired match resamples, conditional on one fit and finite draws. They omit training-seed uncertainty and repeated-development correction.

What you are seeing

Passes and shots from event data. There are no off-ball player positions. The possession endings scored above are no shot, shot without goal, goal and own goal. A natural model stop is distinct from reaching a sampling limit.

“Possession futures” shows saved draws for a chosen pairing of clubs the models learned, or a retained fixture. “Where futures diverge” gives a transformer the actual beginning of a possession. That is a diagnostic of continuations, not a pre-match forecast. The six observed possessions per league were selected chronologically, without choosing by outcome.

The interface browses saved draws. Another future selects another draw; it does not trigger a new inference. Filtering a finite sample changes what is shown, not the model's probability. Older historical paths do not retain the own-goal flag, so their goal events remain ambiguous.