The projection behind the next round's probabilities and the projected finish, read against what happened: this season as it goes, from forecasts written down before each match, and the past seasons it could not have been tuned to. Nothing here is a judgement of the model: those are made by the project's test harness and recorded in its model log. These are the same forecasts, read plainly.
Each fixture's forecast of record is the last one this site made before its kickoff; the record began in October 2026, so earlier rounds have none. The score is the log loss: for each match, minus the logarithm of the chance the forecast gave to the result that happened, averaged over the matches. Lower is better, and 1.099 is the score of giving every outcome a third. The yardstick is the league's own frequencies over the previous eight seasons (44% home wins, 23% draws, 33% away wins), forecast for every match alike.
Over the 5 matches scored so far the favourite won 40% of the time. The forecasts score 1.180 and the frequencies 1.138. The frequencies are the better by 0.042 per match. In a run of matches where the favourites lose, a forecast that backs them pays more than one that never does, and 5 matches say little either way: over the three seasons before these forecasts went live, 2023/24 to 2025/26, they beat the frequencies by 0.086 per match (the project's final test of the projection, run r049). The market's closing price (Bet365, the bookmaker's margin taken out) scores 1.186 on the 5 of those matches that have one, worse than the forecasts by 0.007; over the seasons measured the market is the better by about 0.011 per match.
| Round | Dates | Matches | Favourite won | Log loss | The frequencies | The market | Season so far: favourite won | Season so far: log loss | Season so far: the frequencies | Season so far: the market |
|---|---|---|---|---|---|---|---|---|---|---|
| Round 6 | 2026-10-10 to 2026-10-12 | 5 | 40% | 1.180 | 1.138 | 1.186 | 40% | 1.180 | 1.138 | 1.186 |
| Match | Score | Forecast, home / draw / away | The market, home / draw / away | Favourite | Given to the result: ours | the market | Forecast written |
|---|---|---|---|---|---|---|---|
| Man Utd v Tottenham | 1-1 | 54% / 21% / 24% | 56% / 23% / 21% | home | 21% | 23% | 2026-10-10 16:29 |
| Crystal Palace v Nottingham Forest | 3-1 | 40% / 25% / 35% | 33% / 29% / 38% | home (won) | 40% | 33% | 2026-10-11 11:46 |
| Hull v Everton | 2-2 | 25% / 25% / 50% | 23% / 28% / 50% | away | 25% | 28% | 2026-10-11 11:46 |
| Liverpool v Man City | 0-3 | 35% / 22% / 43% | 27% / 25% / 48% | away (won) | 43% | 48% | 2026-10-11 11:46 |
| Coventry v Newcastle | 4-1 | 30% / 24% / 46% | 26% / 25% / 49% | away | 30% | 26% | 2026-10-12 06:21 |
Before a match the projection gives three probabilities: home win, draw, away win. The favourite is the outcome with the largest of the three. The reading asks one question in several ways: when the projection says 60 per cent, does the thing happen 60 per cent of the time?
Over the 1,514 Premier League matches of 2019/20 to 2022/23, the favourite won 52% of matches against a mean forecast of 51%: the projection names the right side a little more often than it says it will.
| Season | Matches | Favourite won | Favourite's mean forecast | Home wins, forecast / actual | Draws | Away wins |
|---|---|---|---|---|---|---|
| 2019/20 | 379 | 50% | 52% | 48% / 45% | 23% / 24% | 29% / 31% |
| 2020/21 | 379 | 48% | 51% | 47% / 38% | 24% / 22% | 29% / 40% |
| 2021/22 | 376 | 53% | 50% | 47% / 43% | 24% / 23% | 30% / 34% |
| 2022/23 | 380 | 54% | 50% | 46% / 48% | 23% / 23% | 31% / 29% |
| The favourite was given | Matches | Mean forecast | It won | Of which home favourites |
|---|---|---|---|---|
| 33% to 40% | 184 | 38% | 37% | 57% |
| 40% to 45% | 350 | 42% | 39% | 67% |
| 45% to 50% | 290 | 47% | 47% | 72% |
| 50% to 55% | 215 | 52% | 58% | 79% |
| 55% to 60% | 192 | 57% | 58% | 80% |
| 60% to 70% | 203 | 64% | 68% | 95% |
| 70% to 100% | 80 | 75% | 82% | 100% |
Above about 50 per cent the favourite wins more often than its forecast says; below it, about as often. The projection's ordering is right and its numbers are too cautious at the top, which is the one thing the harness has measured about it that a later change can fix.
Each row is a tenth of the matches, ordered by the forecast: the mean forecast against how often the outcome happened.
| Home wins | ||
|---|---|---|
| Forecast | Happened | Matches |
| 25% | 15% | 152 |
| 33% | 25% | 151 |
| 37% | 34% | 152 |
| 41% | 36% | 151 |
| 44% | 33% | 151 |
| 48% | 42% | 152 |
| 52% | 53% | 151 |
| 56% | 59% | 152 |
| 62% | 62% | 151 |
| 71% | 79% | 151 |
| Draws | ||
|---|---|---|
| Forecast | Happened | Matches |
| 17% | 10% | 152 |
| 20% | 29% | 151 |
| 22% | 20% | 152 |
| 23% | 15% | 151 |
| 24% | 22% | 151 |
| 24% | 25% | 152 |
| 25% | 30% | 151 |
| 26% | 25% | 152 |
| 26% | 22% | 151 |
| 28% | 32% | 151 |
| Away wins | ||
|---|---|---|
| Forecast | Happened | Matches |
| 12% | 10% | 152 |
| 17% | 17% | 151 |
| 21% | 18% | 152 |
| 24% | 22% | 151 |
| 27% | 30% | 151 |
| 30% | 38% | 152 |
| 33% | 36% | 151 |
| 37% | 49% | 152 |
| 43% | 50% | 151 |
| 53% | 66% | 151 |
Draws are forecast within a narrow band for nearly every match and happen about as often as forecast over the seasons: the projection says little about which matches will be drawn.
The ratings carry over from earlier seasons, so the first matches are not forecast blind. The stage is the home team's matches played so far that season.
| Stage of the season | Matches | Favourite's forecast | Favourite won | Draws, forecast / actual |
|---|---|---|---|---|
| matches 1 to 5 | 201 | 51% | 48% | 23% / 25% |
| 6 to 10 | 200 | 51% | 50% | 23% / 26% |
| 11 to 19 | 363 | 51% | 51% | 23% / 23% |
| 20 to 29 | 395 | 51% | 52% | 24% / 24% |
| 30 to 38 | 355 | 51% | 54% | 23% / 19% |
The results the projection gave the least chance to. A season has about 380 matches, so a result forecast at 10 per cent is expected about 38 times a season: a list of misses is not a list of errors.
| Season | Date | Match | Score | Forecast, home / draw / away | Given to the result |
|---|---|---|---|---|---|
| 2019/20 | 2019-10-06 | Man City v Wolves | 0-2 | 77% / 14% / 9% | 9% |
| 2022/23 | 2022-11-12 | Man City v Brentford | 1-2 | 75% / 15% / 10% | 10% |
| 2021/22 | 2021-10-30 | Man City v Crystal Palace | 0-2 | 74% / 16% / 10% | 10% |
| 2020/21 | 2021-01-21 | Liverpool v Burnley | 0-1 | 73% / 17% / 10% | 10% |
| 2020/21 | 2021-04-03 | Chelsea v WBA | 2-5 | 71% / 18% / 11% | 11% |
| 2022/23 | 2022-10-29 | Liverpool v Leeds | 1-2 | 72% / 16% / 11% | 11% |
| 2020/21 | 2021-03-07 | Liverpool v Fulham | 0-1 | 71% / 18% / 12% | 12% |
| 2020/21 | 2021-04-10 | Man City v Leeds | 1-2 | 71% / 17% / 12% | 12% |
| 2020/21 | 2021-02-03 | Liverpool v Brighton | 0-1 | 70% / 18% / 13% | 13% |
| 2020/21 | 2021-01-27 | Man Utd v Sheff Utd | 1-2 | 66% / 20% / 14% | 14% |
| 2020/21 | 2020-12-15 | Man City v WBA | 1-1 | 79% / 14% / 7% | 14% |
| 2022/23 | 2022-12-31 | Man City v Everton | 1-1 | 77% / 14% / 9% | 14% |
| 2019/20 | 2020-01-18 | Man City v Crystal Palace | 2-2 | 76% / 14% / 9% | 14% |
| 2020/21 | 2020-12-27 | Liverpool v WBA | 1-1 | 77% / 15% / 8% | 15% |
| 2019/20 | 2019-12-26 | Chelsea v Southampton | 0-2 | 67% / 19% / 15% | 15% |
2023/24 to 2025/26 were read once, as the projection's final test, out of time: 1,140 matches. The score the harness judges on (log loss, lower is better) was 0.988 against 1.074 for always forecasting the league's own frequencies.
| Season | The projection | The league's frequencies |
|---|---|---|
| 2023/24 | 0.935 | 1.054 |
| 2024/25 | 0.994 | 1.082 |
| 2025/26 | 1.034 | 1.085 |
| Home wins (slope 1.40) | ||
|---|---|---|
| Forecast | Happened | Matches |
| 22% | 15% | 114 |
| 30% | 22% | 114 |
| 36% | 26% | 114 |
| 40% | 28% | 114 |
| 44% | 54% | 114 |
| 48% | 38% | 114 |
| 52% | 50% | 114 |
| 57% | 59% | 114 |
| 63% | 67% | 114 |
| 72% | 73% | 114 |
| Draws (slope 1.39) | ||
|---|---|---|
| Forecast | Happened | Matches |
| 16% | 16% | 114 |
| 19% | 19% | 114 |
| 20% | 25% | 114 |
| 21% | 26% | 114 |
| 22% | 27% | 114 |
| 22% | 25% | 114 |
| 23% | 25% | 114 |
| 24% | 30% | 114 |
| 24% | 25% | 114 |
| 26% | 28% | 114 |
| Away wins (slope 1.39) | ||
|---|---|---|
| Forecast | Happened | Matches |
| 12% | 8% | 114 |
| 18% | 15% | 114 |
| 22% | 18% | 114 |
| 25% | 27% | 114 |
| 29% | 29% | 114 |
| 32% | 26% | 114 |
| 36% | 39% | 114 |
| 41% | 45% | 114 |
| 47% | 47% | 114 |
| 56% | 69% | 114 |
The slope is the harness's calibration measure: 1 is calibrated, above 1 too cautious. The same shape as on the earlier seasons: right in its ordering, too cautious at the top.
At each stage of the seasons 2019/20 to 2022/23, the projected finish was recomputed as it would have stood on that date and compared with the actual final table, beside a naive projection: points so far plus the current points per match over the matches left. "In range" is the share of teams whose final points, or position, fell inside the projection's 10th-to-90th range; a complete range would give 80 per cent. "Ratings fixed" is a simulation that varies only the match results, and its shortfall is the error in the ratings themselves. The league page's range includes that error: in each of 100 draws every team's two ratings are also moved by their own random factors with a spread of 0.24 (the standard deviation of the factor's logarithm), the spread set so that the range holds 80 per cent of final points over every team at every stage of these seasons (it holds 89%, against 72% with the ratings fixed).
| After the median team had played | Matches left | Points error, projection | Points error, naive | Position error, projection | Position error, naive | Final points in range | Final points in range, ratings fixed | Final position in range | Final position in range, ratings fixed | Champion's title chance | Relegated teams' chance |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 5 matches | 331 | 8.0 | 16.4 | 2.8 | 3.7 | 89% | 62% | 94% | 76% | 40% | 34% |
| 10 matches | 282 | 7.2 | 11.0 | 2.4 | 2.9 | 91% | 65% | 95% | 85% | 41% | 45% |
| 19 matches | 192 | 4.5 | 5.7 | 1.6 | 1.9 | 91% | 82% | 94% | 92% | 66% | 56% |
| 29 matches | 94 | 3.6 | 4.1 | 1.3 | 1.4 | 84% | 77% | 90% | 82% | 75% | 66% |
On the 1,513 Premier League matches of 2019/20 to 2022/23, the projection as this site shows it against Pinnacle's closing price with the bookmaker's margin taken out. "Ours minus the market's" is our probability for an outcome less the market's; a positive mean means we give it more. The market's log loss is lower (better) on the result by 0.0233 per match, with a standard error of 0.0054, and on the total by 0.0039 (0.0038). The two name a different favourite in 16% of matches.
| Outcome | Our mean | The market's mean | Happened | Ours minus the market's: mean | Mean absolute | Median | 10th to 90th | Correlation |
|---|---|---|---|---|---|---|---|---|
| Home win | 47% | 43% | 44% | +0.036 | 0.081 | +0.046 | -0.091 to +0.144 | 0.95 |
| Draw | 23% | 24% | 23% | -0.006 | 0.029 | -0.013 | -0.045 to +0.045 | 0.81 |
| Away win | 30% | 33% | 33% | -0.029 | 0.070 | -0.012 | -0.153 to +0.064 | 0.95 |
| Over 2.5 goals | 53% | 53% | 52% | -0.004 | 0.050 | -0.002 | -0.089 to +0.075 | 0.70 |
| The home team's match number | Matches | Mean absolute disagreement, home / draw / away / over | Our log loss | The market's | Difference (market minus ours) | Standard error |
|---|---|---|---|---|---|---|
| 1 to 5 | 201 | 0.076 / 0.028 / 0.065 / 0.050 | 0.9984 | 0.9779 | -0.0205 | 0.0172 |
| 6 to 10 | 199 | 0.069 / 0.028 / 0.058 / 0.048 | 1.0059 | 0.9823 | -0.0236 | 0.0124 |
| 11 to 19 | 363 | 0.083 / 0.029 / 0.072 / 0.051 | 1.0033 | 0.9786 | -0.0247 | 0.0129 |
| 20 to 29 | 395 | 0.082 / 0.029 / 0.071 / 0.047 | 0.9896 | 0.9664 | -0.0232 | 0.0094 |
| 30 or more | 355 | 0.089 / 0.030 / 0.077 / 0.056 | 0.9720 | 0.9484 | -0.0236 | 0.0130 |
Each team over its matches home and away: our chance of the team winning less the market's (positive: we rate the team higher than the market), and the log loss difference on its matches (market minus ours, positive when ours is better).
| Team | Matches | Our win chance, mean | The market's | Ours minus the market's | Mean absolute | Difference on its matches | Standard error |
|---|---|---|---|---|---|---|---|
| 168 | 75 | 26% | 20% | +0.067 | 0.073 | -0.0648 | 0.0224 |
| Bournemouth | 76 | 30% | 23% | +0.063 | 0.069 | -0.0334 | 0.0170 |
| Nottingham Forest | 38 | 27% | 21% | +0.060 | 0.067 | -0.0057 | 0.0395 |
| 175 | 38 | 25% | 20% | +0.052 | 0.070 | +0.0098 | 0.0568 |
| Crystal Palace | 151 | 31% | 27% | +0.043 | 0.064 | -0.0195 | 0.0155 |
| 184 | 112 | 29% | 24% | +0.043 | 0.062 | -0.0321 | 0.0159 |
| 27 | 76 | 29% | 25% | +0.041 | 0.067 | -0.0342 | 0.0286 |
| Brentford | 76 | 34% | 30% | +0.037 | 0.058 | +0.0038 | 0.0161 |
| Everton | 152 | 36% | 33% | +0.033 | 0.069 | -0.0083 | 0.0108 |
| Newcastle | 151 | 33% | 30% | +0.032 | 0.081 | -0.0278 | 0.0141 |
| 29 | 152 | 37% | 34% | +0.029 | 0.070 | -0.0177 | 0.0140 |
| Fulham | 76 | 30% | 27% | +0.029 | 0.060 | -0.0476 | 0.0178 |
| 14 | 152 | 41% | 38% | +0.027 | 0.077 | -0.0402 | 0.0167 |
| 18 | 150 | 32% | 30% | +0.027 | 0.061 | -0.0075 | 0.0157 |
| 161 | 151 | 35% | 33% | +0.025 | 0.065 | -0.0187 | 0.0146 |
| 163 | 75 | 28% | 26% | +0.017 | 0.067 | -0.0193 | 0.0326 |
| Aston Villa | 152 | 34% | 32% | +0.016 | 0.065 | -0.0162 | 0.0167 |
| Leeds | 114 | 33% | 32% | +0.016 | 0.063 | -0.0327 | 0.0239 |
| Tottenham | 150 | 46% | 47% | -0.011 | 0.076 | -0.0039 | 0.0152 |
| Brighton | 152 | 34% | 36% | -0.021 | 0.071 | -0.0122 | 0.0148 |
| Arsenal | 152 | 45% | 49% | -0.040 | 0.082 | -0.0425 | 0.0158 |
| Chelsea | 151 | 50% | 55% | -0.046 | 0.086 | -0.0218 | 0.0245 |
| Man Utd | 152 | 46% | 51% | -0.048 | 0.093 | -0.0273 | 0.0192 |
| Liverpool | 152 | 56% | 64% | -0.076 | 0.103 | -0.0368 | 0.0241 |
| Man City | 150 | 61% | 74% | -0.122 | 0.128 | -0.0210 | 0.0256 |
Where our probability for an outcome is above the market's by more than a threshold, and where it is below: the mean of each forecast and how often the outcome happened. If what happened sits at the market's figure, the market was right and our disagreement carried nothing; between the two, we had something.
| Threshold | Outcome | Ours higher: matches | Ours | Market | Happened | Ours lower: matches | Ours | Market | Happened |
|---|---|---|---|---|---|---|---|---|---|
| 0.05 | home | 728 | 39% | 27% | 27% | 296 | 61% | 71% | 71% |
| 0.05 | draw | 128 | 20% | 13% | 17% | 98 | 23% | 29% | 27% |
| 0.05 | away | 285 | 21% | 14% | 16% | 549 | 39% | 52% | 53% |
| 0.05 | over | 298 | 54% | 46% | 41% | 349 | 53% | 62% | 56% |
| 0.1 | home | 430 | 37% | 23% | 22% | 128 | 64% | 77% | 72% |
| 0.1 | draw | 8 | 21% | 10% | 0% | 0 | |||
| 0.1 | away | 23 | 25% | 13% | 9% | 334 | 43% | 58% | 56% |
| 0.1 | over | 66 | 56% | 44% | 42% | 106 | 54% | 67% | 61% |
A regression of each outcome on the market's log odds and ours: a weight on ours above zero by more than two standard errors means our forecast adds to the market's.
| Outcome | Weight on the market | Standard error | Weight on ours | Standard error |
|---|---|---|---|---|
| home | 1.161 | 0.183 | -0.390 | 0.304 |
| draw | 0.702 | 0.327 | 0.370 | 0.610 |
| away | 0.988 | 0.192 | -0.157 | 0.315 |
| over | 0.752 | 0.201 | 0.454 | 0.278 |
Descriptive: nothing here judges the projection or changes it. The judged number is the harness's run r053: the projection is 0.0234 log loss per match behind the market on these seasons. Whether blending the two beats the market is stage 9's question.
It judges nothing: every number is from forecasts the harness has already scored. It does not say how well the current season is being forecast; the next round's probabilities are the same forecasts, and their reading waits for the matches. It prices nothing against the bookmakers.