Sat 10 Oct · 7:30 AM ET
ArsenalvLeeds
- Expected goals
- 2.08 – 0.64
- Likeliest score
- 2–0 14%
Model 02 · Premier League 2026–27 · live
The same Bayesian engine, re-tuned for club football: twelve seasons of results, a league Elo anchor, newly promoted clubs, and a simulator that plays the rest of the season 20,000 times.
This weekend · Matchday 6
Issued on Friday 9 October, using every result up to Thursday night. Each match was played out 20,000 times; the bar shows how often each side won.
Sat 10 Oct · 7:30 AM ET
Sat 10 Oct · 10:00 AM ET
Sat 10 Oct · 10:00 AM ET
Sat 10 Oct · 10:00 AM ET
Sat 10 Oct · 10:00 AM ET
Sat 10 Oct · 12:30 PM ET
Sun 11 Oct · 9:00 AM ET
Sun 11 Oct · 9:00 AM ET
Sun 11 Oct · 11:30 AM ET
Mon 12 Oct · 3:00 PM ET
Expected goals here are the model's average goals per side across those 20,000 runs, not shot-based xG. Even the likeliest exact score happens only about one time in eight. Percentages are rounded, so a row may add up to 99. Team news after Thursday is not included, and none of these forecasts has been scored yet.
Method
Every new issuance runs these steps from scratch, using only matches that had already been played.
Every Premier League match from 2014/15 up to the cutoff day, with older matches counting less: a year-old result counts half.
A running Elo rating for every club (K = 20). Newly promoted clubs start 75 points below the division average.
Each club's attack and defence start from its Elo rating (weight 0.6), then the goals pull them where they belong.
Dixon–Coles, fitted with ADVI over 30,000 iterations, keeping 1,000 plausible versions of every club.
An 11 × 11 score grid for every draw. Clubs the model knows little about get a wider, less confident grid.
Every remaining fixture is played out 20,000 times under Premier League tiebreak rules. The published numbers are those frequencies.
One forecast, start to finish
Coventry had never played in the model's twelve-season archive, so their strength came from the promoted-club rule and the model's own prior. Here is what the model published before kickoff, and what happened.
From 20,000 simulated seasons. Each figure carries a simulation error of about ±0.3 percentage points.
3–0
Arsenal won, the outcome the model rated most likely. The forecast's ranked probability score was 0.031, close to perfect (0).
Also on that opener board: Arsenal v Man City in November, priced at 44.5% · 28.7% · 26.8%.
Contributing factors
Follow one real match through the machine, from the data it reads to the forecast it publishes. Every number is a real setting from the published model.
Every fixture of the season
The forecast published on 25 August, after the opening weekend: every remaining fixture, home club down the side, away club across the top. Taller means a likelier home win. Hover any cell, or pick a club.
The season outlook
Before a ball was kicked on 21 August, the simulator counted how often each club finished first, in the top four or in the bottom three.
Median points needed: 86 for the title, 67 for fourth place, 38 to survive. Simulation counts, not guarantees: within one forecast, team strengths are held fixed for the rest of the season, so the simulation knows nothing of injuries or January transfers. They update when a later forecast re-fits on new results.
The live cycle
Each step only runs if the one before it passed. Results enter only when two independent sources agree exactly; any disagreement stops the machine.
Forecasts as evidence
Each issuance prices every remaining fixture and is sealed. Results are scored against the freshest forecast that still came before kickoff.
Every issuance
370
The 25 August issuance priced 370 fixtures, from days to nine months ahead, each stored with its time, inputs and model version.
After new results
+1
New results trigger a new issuance. It sits beside the old one and never replaces it.
If a cycle fails
×
When a run stops, nothing is filled in afterwards. The older forecast stays the one that gets scored.
After the final whistle
90′
Each result is scored against the most recent forecast issued before that kickoff, never one made later.
In the product plan
T−24h
The specification adds separate day-ahead and near-kickoff forecasts for every match. Planned, not yet live.
Experiments
Each experiment fixed its pass mark before seeing results, usually an improvement of 0.001 in RPS with an interval entirely below zero. None has passed yet, and every result is published.
Six seasons, 2,280 matches. The model was ahead, but the interval crosses zero, so the pre-written pass rule failed.
Δ −0.0012 · interval −0.0028 to +0.0005
Memory length, fixture congestion, a faster-moving home edge and post-window uncertainty: 47 runs on older seasons only. Shorter memory was worse every time.
best combination −0.00007 · bar −0.0010
2025/26, scored once. The model finished slightly behind Elo, inside the noise; one season can only detect a much larger gap.
Δ +0.0010 · interval −0.0026 to +0.0045
The estimate leaned towards fresher fits, but it missed the bar and its interval crossed zero. The owner switched to matchday refits anyway, recorded as a decision rather than evidence.
Δ −0.00022 · bar −0.00030
Each refit began from what opening prices implied about every club. The gain was a fifth of the bar.
Δ −0.00019 · bar −0.0010
The largest effect measured so far, on 85 matches involving clubs the model barely knew. One of its two intervals crossed zero.
Δ −0.0041 on 85 fixtures
Checked against 30 promoted clubs. It was closer on average, but better in only 15 of 30, and the interval crosses zero.
error 57.7 vs 64.3 Elo
The fit converged but failed its pre-set quality certificate, so no score was ever computed.
no score: fit failed
The only clean alternative had already measured worse on the tuning seasons. Running it could not have produced a usable answer, so it was not run.
−100 Elo seed: +0.00017 worse
Still running, or ended without a result
Results appear here once they are published in full. Δ is the change in mean RPS against the live model; negative is better.
The product it is becoming
The forecasts are being built into a match workspace. Its specification starts with three questions that are easy to blur together.
Question 1
Home, draw and away probabilities, expected goals and supported totals for each fixture, with the time the forecast was made.
Question 2
The whole season from matchday 1, including missing, late and refused forecasts. Never just the good weeks.
Question 3
A likely outcome and a well-priced outcome are different things. Without enough evidence, the answer stays “no”.
| Market | What the model outputs | Status |
|---|---|---|
| Home · draw · away | Three probabilities that always add up to one | In scope |
| Goals and margins | Expected winning margin and the chance of winning by 2, 3 or 4 or more | In scope |
| Spreads and handicaps | The exact goal-margin event a contract describes, settled by its own rules | In scope |
| Anytime scorer | Each player's scoring chance, with minutes and lineup status shown separately | In scope |
| First scorer | Needs its own event-order model and validation | Deferred |
| In-play | Needs live data, state changes and latency rules | Deferred |
Research roadmap
Each is tested on its own against the unchanged model. Only then do the winners get combined, and only after that does anyone look at prices.
The rules of the work
Work moves through five milestones, with two gates only the project owner can open.
M1
Research harnesses, tests and independent reviews.
Gate A
No experiment starts from a general “continue”.
M2
Frozen, authorised runs. Every verdict published.
Gate B
Two good parts do not automatically make a good whole.
M3
Market pages, plus scorer, goal and workload research.
M4
Paper decisions on real future prices, never hindsight.
M5
Only with an explicit decision and a rollback plan.