Premier League
Matchweek 9Sat, Oct 31 – Mon, Nov 2times in ET · updated Oct 29, 8:48 PM ETTeam match model — tested blind on past seasons. Probability distributions, not picks. Paper-only and educational.
What these numbers are
Tested blind on nine past Premier League seasons (2013-14 to 2021-22, 3,420 matches) it was never fit on, where it beat the previous model and a plain rating system. Its total-goals numbers follow the league's recent scoring rate, so they are the same for every match — in that test they were more accurate than the previous model's team-by-team totals. Its live record this season is still being graded. 40 Premier League matches have been graded under this model — far too few to support any accuracy claim. No live win rate or accuracy figure is quoted until its season record is large enough to mean something. 36 matches were graded under the model this one replaced; that record is kept separately and is not evidence for this model.
These are the model’s own probability distributions, published so you can see what it says — not picks, not advice, and not compared against any price.
Fixtures
10 scheduled · 10 with a report · 10 with a supported read
- Sat, Oct 31 · 8:30 AM ETUpcomingv
Chelsea(model favourite)
Manchester UnitedChelsea · 38%model forecast · match resultChelsea 38%Draw 25%Manchester United 37% - Sat, Oct 31 · 11:00 AM ETUpcomingv
Manchester City(model favourite)
Brighton & Hove AlbionManchester City · 63%model forecast · match resultManchester City 63%Draw 21%Brighton & Hove Albion 16% - Sat, Oct 31 · 11:00 AM ETUpcomingv
Bournemouth(model favourite)
Leeds UnitedBournemouth · 52%model forecast · match resultBournemouth 52%Draw 24%Leeds United 24% - Sat, Oct 31 · 11:00 AM ETUpcomingv
Coventry City
Sunderland(model favourite)Sunderland · 40%model forecast · match resultCoventry City 35%Draw 25%Sunderland 40% - Sat, Oct 31 · 11:00 AM ETUpcomingv
Hull City(model favourite)
Ipswich TownHull City · 50%model forecast · match resultHull City 50%Draw 24%Ipswich Town 26% - Sat, Oct 31 · 11:00 AM ETUpcomingv
Brentford(model favourite)
Nottingham ForestBrentford · 52%model forecast · match resultBrentford 52%Draw 24%Nottingham Forest 24% - Sat, Oct 31 · 1:30 PM ETUpcomingv
Tottenham Hotspur(model favourite)
Crystal PalaceTottenham Hotspur · 40%model forecast · match resultTottenham Hotspur 40%Draw 25%Crystal Palace 35% - Sat, Oct 31 · 4:00 PM ETUpcomingv
Aston Villa(model favourite)
FulhamAston Villa · 53%model forecast · match resultAston Villa 53%Draw 24%Fulham 24% - Sun, Nov 1 · 11:30 AM ETUpcomingv
Liverpool
Arsenal(model favourite)Arsenal · 42%model forecast · match resultLiverpool 33%Draw 25%Arsenal 42% - Mon, Nov 2 · 3:00 PM ETUpcomingv
Newcastle United(model favourite)
EvertonNewcastle United · 46%model forecast · match resultNewcastle United 46%Draw 25%Everton 29%
Expected total goals is 2.79 in every match (over 2.5 goals: 52.9%): the model sets each match’s total from the league’s recent scoring rate and splits it between the two sides by strength. Each report shows its own scorelines and each side’s goals.
Also forecast with this model: Ligue 1 match forecasts →
If he starts — likeliest scorers
Shot-on-goal and scoring chances — conditional on STARTING until lineups post (~1h before kickoff).
- Erling HaalandManchester City · Manchester City v Brighton & Hove Albion127 appsSOG 81.3%Score 52.9%
- Alexander IsakLiverpool · Liverpool v Arsenal86 appsSOG 69.4%Score 41.2%
- Igor ThiagoBrentford · Brentford v Nottingham Forest38 appsSOG 64.4%Score 36.4%
- Chris WoodNottingham Forest · Brentford v Nottingham Forest75 appsSOG 59.3%Score 35.4%
- Callum WilsonBrentford · Brentford v Nottingham Forest43 appsSOG 61.2%Score 35.3%
- Cole PalmerChelsea · Chelsea v Manchester United91 appsSOG 65%Score 33.4%
- Hugo EkitikeLiverpool · Liverpool v Arsenal21 appsSOG 57.7%Score 32.3%
- Taiwo AwoniyiCoventry City · Coventry City v Sunderland35 appsSOG 53.3%Score 31.9%
- Yoane WissaNewcastle United · Newcastle United v Everton83 appsSOG 58.8%Score 31.6%
- Jean-Philippe MatetaCrystal Palace · Tottenham Hotspur v Crystal Palace89 appsSOG 59.8%Score 31.5%
- Zian FlemmingIpswich Town · Hull City v Ipswich Town21 appsSOG 57.7%Score 31%
- Bryan MbeumoManchester United · Chelsea v Manchester United127 appsSOG 59.7%Score 30.4%
Player model · not validated — the holdout test
Fitted on 45,788 appearances, then tested on a season it had never seen — 11,567 player-matches, log loss 0.25062 against 0.26043 for a position-only baseline, and it predicted 1040 scorers where 987 actually scored. Once an eleven is named, a side’s goals are shared across the men actually playing — the rows add back to what the match simulation expects that team to score.
- Three markets were measured under identical bars and REJECTED — plain shots on calibration, assists and cards on margin. They are deliberately absent rather than shown with a warning.
- Injury and suspension news is not joined into these lists yet, so an unavailable player can still appear in a conditional list.
- The source carries no minutes — participation is a discrete state (started or substitute), not an expected-minutes term.
- Conditional rows state P(scores | he starts). They are not a claim that he will start.
Past matchesEach played fixture: the forecast published before kickoff, and the result once graded.
Every fixture we published a forecast for
What the model said before each kickoff, kept exactly as it was published. These are past fixtures; nothing here is a current read. Results are read against the official score after full time — a log far too short to say anything about accuracy; the model's test is the blind run on past seasons.
Player markets: 3904 projections graded, 3818 void because the player did not take the field the projection assumed — a condition that did not hold is never scored as a miss.
How the model's picks actually turned out
Match-result forecasts: the outcome the model gave the most probability to, graded against the official full-time score.
76 graded picks is still a small sample. A hit rate over this many outcomes has a wide enough margin of error to be consistent with almost any true accuracy.
This model has cleared no preregistered bar and publishes distributions rather than picks — the 'predicted' column is simply its likeliest outcome.
| Date | Event | Model said | What happened | Result |
|---|---|---|---|---|
| 2026-10-25 | Sunderland v Leeds United | home win · 39.4% | draw (0-0) | MISS |
| 2026-10-25 | Crystal Palace v Newcastle United | home win · 38.1% | away win (0-2) | MISS |
| 2026-10-25 | Hull City v Brentford | away win · 46.6% | draw (1-1) | MISS |
| 2026-10-25 | Liverpool v Brighton & Hove Albion | home win · 48.4% | home win (2-0) | HIT |
| 2026-10-25 | Manchester United v Bournemouth | home win · 49.5% | away win (0-3) | MISS |
| 2026-10-24 | Chelsea v Tottenham Hotspur | home win · 58.6% | away win (0-2) | MISS |
Showing the 6 most recent of 76 graded.
Model health · Premier League
open for the live record behind each callMatch model Tested on past seasonsForward test Forward test runningLive record Holding
Model status · Premier League
What the statuses mean →- Match modelTested on past seasonsTested blind on nine past seasons
Scored on 3,420 Premier League matches from seasons it was never fit on: log loss 0.974, against 0.983 for a plain rating system and 0.994 for the previous model (lower is better). The gain over the rating system held in a season bootstrap (95% interval −0.018 to −0.001 per match). Its totals follow the league's recent scoring rate, so over-2.5 is the same for every match.
What does this mean?
Before it went live, this model was scored on past seasons it had never seen and cleared bars set in advance against simpler baselines. That is a bar, not a promise, and not a claim that it beats sportsbook prices.
- Forward testForward test runningForward test · 40 of 60 matches
Every graded match scores the live model beside the one it replaced. A verdict needs 60 matches; 40 are in.
What does this mean?
The model is being graded on new games as they happen, against the model it replaced. If it falls significantly behind, the previous model comes back automatically.
- Live recordHoldingHolding against even odds
Measured by log loss (lower is better) over 76 graded: 1.030 for this call against 1.099 for even odds, a difference of −0.068 (95% interval −0.168 to +0.037). The interval includes zero, so this call cannot yet be told apart from even odds.
What does this mean?
Over the games graded so far, this call has done at least as well as its floor (a coin flip, or the model it replaced).
Full 2026-27 fixture listEvery fixture, whether or not the model can price it.
2026-27 fixture list
Every fixture, whether or not the model can price it.
Saturday, October 31
- Manchester UnitedatChelseaSat, Oct 31, 8:30 AM ET
- Brighton & Hove AlbionatManchester CitySat, Oct 31, 11:00 AM ET
- Leeds UnitedatBournemouthSat, Oct 31, 11:00 AM ET
- SunderlandatCoventry CitySat, Oct 31, 11:00 AM ET
- Ipswich TownatHull CitySat, Oct 31, 11:00 AM ET
- Nottingham ForestatBrentfordSat, Oct 31, 11:00 AM ET
- Crystal PalaceatTottenham HotspurSat, Oct 31, 1:30 PM ET
- FulhamatAston VillaSat, Oct 31, 4:00 PM ET
Sunday, November 1
- ArsenalatLiverpoolSun, Nov 1, 11:30 AM ET
Monday, November 2
- EvertonatNewcastle UnitedMon, Nov 2, 3:00 PM ET
Saturday, November 7
- Hull CityatArsenalSat, Nov 7, 10:00 AM ET
- Manchester CityatNottingham ForestSat, Nov 7, 10:00 AM ET
Source: openfootball (public domain) · captured Oct 24, 10:16 AM ET · showing the next 12 of 300
Technical detailsInternal model identifiers behind the numbers on this page.
Player model: epl-player-v2-shrunk-rate
Paper-only and educational. Not betting advice. Updated Oct 29, 8:48 PM ET from pregame inputs only; the model never sees a result from a match it is forecasting.