Premier League predictor

A running project · Started September 2026
Machine learningFootball

I'm building a model that predicts the outcome of every Premier League game each week — win, draw, or loss, with a probability for each. I'm building it by hand, from scratch, as a way to learn: no ML libraries doing the modeling for me, just NumPy, SciPy, and the math.

The code is public.

What this is

Every week, the model outputs three probabilities for each of the ten Premier League fixtures: home win, draw, away win. The goal is simple to state and hard to do: beat the bookmakers' implied probabilities, week after week, measured honestly on games the model has never seen.

Bookmaker odds are the benchmark, never an input — the model doesn't get to peek at the answer key. Promoted teams stay in; every game gets a prediction, no exceptions.

How the model works

The core is a Dixon–Coles model, the classic statistical approach to football scores. Each team gets an attack rating and a defense rating, there's a league-wide home advantage, and a small correction for low-scoring games. From those ratings the model computes the probability of every scoreline from 0–0 to 10–10, then adds them up into home/draw/away probabilities.

The first version worked but had a flaw: with only a few games of data, newly promoted teams got extreme ratings. After four losses, Coventry's attack rating was so negative the model gave them a 0.04% chance at Nottingham Forest. They won. That one game blew up the week's score.

Version 2 fixes this two ways. First, shrinkage: a gentle prior that pulls ratings toward average until a team has played enough games to earn an extreme one — the same idea as not judging a batsman by his first over. Second, time decay: recent games count more than old ones, with a 180-day half-life. Both settings were tuned on two full past seasons the model was never evaluated on.

Everything is validated walk-forward: to predict a game, the model may only use games played before it. No leakage, no peeking. An Elo ratings system runs alongside as a standing challenger.

How I'll use it

Each week follows the same loop:

  1. Before the matchweek — publish the model's win/draw/loss probabilities for all ten games.
  2. After the matchweek — compare predictions against results, write up what the model got right and wrong, and add it as a new section below.
  3. Iterate — feed what I learn back into the model. v3 so far: uncertainty quantification done; Elo ensemble built (55/45 v2/Elo blend tuned on 2024–26 — honestly the gain over v2 alone is within noise). Next: expected-goals features, then injuries as a v4 parameter. The full-season bookmaker comparison waits for 380 games.

Matchweek 11 November 21–23, 2026

Preview

The league is back after the international break, and the model returns with it. No change to the model this week: the home prior introduced in Matchweek 9 stays, and the blend (55% Dixon-Coles, 45% Elo) is unchanged. Matchweek 9 was the best accuracy week of the season at 6/10, but the market still won on log loss, so there is nothing here to fix and nothing to celebrate yet. Season tally: 23/50 against the market's 25/50.

GamePickHomeDrawAwayMkt HMkt DMkt A
Man City v Fulham Home win · 66% 66% 21% 13% — — —
Bournemouth v Nott'm Forest Home win · 54% 54% 25% 21% — — —
Aston Villa v Sunderland Home win · 57% 57% 24% 19% — — —
Chelsea v Leeds Home win · 45% 45% 27% 28% — — —
Coventry v Crystal Palace Home win · 45% 45% 27% 28% — — —
Tottenham v Ipswich Home win · 55% 55% 24% 21% — — —
Newcastle v Arsenal Away win · 59% 17% 24% 59% — — —
Hull v Brighton Away win · 47% 25% 28% 47% — — —
Liverpool v Man United Home win · 54% 54% 24% 22% — — —
Brentford v Everton Home win · 48% 48% 27% 25% — — —

Market columns: the bookmaker odds had not been posted to football-data.co.uk's fixtures feed when these picks were published, so there is nothing honest to put in those columns yet. The review will score the market on closing lines, as usual.

The headline is Sunday at Anfield: Liverpool against Manchester United, the model backs Liverpool at 54%. Arsenal away at Newcastle is the week's other away pick and the biggest away lean at 59%. Manchester City at home to Fulham is the banker at 66%. The two coin flips are Chelsea against Leeds and Coventry against Crystal Palace, both 45% home, 27% draw, 28% away. Eight home picks, two away, no draws on top — the same draw pattern as all season.

As ever, no injury or lineup overlay: the absence model stayed a validated null in testing and remains parked.

Matchweek 10 November 6–8, 2026

Skipped

No predictions were published for Matchweek 10. A timing issue meant the run didn't happen before the games kicked off, so there are no picks and no score to report. The season tally stays at 23/50, and the lesson is to run earlier next time.

Matchweek 9 October 31–November 2, 2026

Preview

After the worst week of the season (2/10 in Matchweek 8, season tally 16/40), the model gets one deliberate change. Matchweek 8 put the model's home lean on trial, so I reworked exactly that: the Dixon-Coles half of the blend now fits with a stronger prior pulling the home-advantage term toward zero. The change was chosen on the 2024–25 tuning window, not on this season — there it improved log loss from 0.9946 to 0.9910, the only home-shrink variant whose confidence interval excluded zero. In practice it trims about a point off every home price this week and flips no picks. The blend (55% Dixon-Coles, 45% Elo) is otherwise unchanged, and Elo's home prices were already close to the market's, so it was left alone.

GamePickHomeDrawAwayMkt HMkt DMkt A
Chelsea v Man United Home win · 42% 42% 26% 32% 52% 24% 24%
Bournemouth v Leeds Home win · 46% 46% 28% 26% 44% 28% 28%
Brentford v Nott'm Forest Home win · 51% 51% 27% 22% 46% 26% 28%
Coventry v Sunderland Home win · 35% 35% 31% 34% 31% 28% 41%
Hull v Ipswich Home win · 55% 55% 26% 19% 36% 28% 36%
Man City v Brighton Home win · 57% 57% 24% 20% 58% 21% 21%
Tottenham v Crystal Palace Home win · 48% 48% 27% 25% 61% 22% 17%
Aston Villa v Fulham Home win · 52% 52% 27% 21% 49% 25% 26%
Liverpool v Arsenal Away win · 44% 27% 29% 44% 32% 28% 40%
Newcastle v Everton Home win · 45% 45% 28% 27% 41% 27% 32%

Against the market, the model agrees on the top pick in nine of ten games. The one split is Coventry v Sunderland: the market makes Sunderland the favourite at 41%, while the model edges Coventry at 35% in the week's closest game. Everywhere else the two lean the same way, though the market is cooler than the model on Hull (36% vs 55%) and hotter on Tottenham (61% vs 48%).

Market columns (Mkt H/D/A): average pre-match odds across books from football-data.co.uk's fixtures feed, captured Friday, with the overround removed. The review will score the market on closing lines, as usual.

The headline is Sunday at Anfield: Arsenal away is the week's only non-home pick, at 44%, and both halves of the model agree on it. Saturday opens with Chelsea against Manchester United at Stamford Bridge, where the model backs Chelsea at 42% without much conviction.

The week's purest coin flip is Coventry v Sunderland: 35% home, 31% draw, 34% away. The Dixon-Coles half actually makes the draw its top outcome there; the blend edges home by less than a point. Man City at home to Brighton is the strongest lean of the week, and at 57% it is a modest one by recent standards.

And the honest caveat about the change above: it trims prices, it does not change behavior. The model still picks nine home winners this week. If home teams struggle again, the question comes back, louder.

As ever, no injury or lineup overlay: the absence model stayed a validated null in testing and remains parked.

Matchweek 8 October 23–25, 2026

2/10 correct

Two out of ten — the worst week of the season. The home lean we flagged in the preview got its trial, and the verdict is not kind.

2/10model's correct picks
1.190model log loss (lower is better)
2/10Elo challenger: 2/10 · 1.214
4/10bookmakers: 4/10 · 1.130

model got it right model got it wrong Draw the game was a draw

GamePredictedActualComments
Ipswich 1–1 Nott'm Forest Away win · 40% Draw Williams put Forest ahead; Enciso's equalizer landed the outcome neither side of the blend wanted. ✗
Aston Villa 2–1 Man City Away win · 52% Home win Jackson's brace ended City's perfect start. Villa Park was the priced trap, and it sprang. ✗
Arsenal 2–0 Everton Home win · 69% Home win The banker banked: two set-piece goals, Konsa and Gabriel. ✓
Coventry 0–1 Fulham Home win · 37% Away win Market split one goes to the market — Bobb's late winner settled it for Fulham. ✗
Chelsea 0–2 Tottenham Home win · 61% Away win The expensive one: Penders' own goal, Palmer sent off, Tel's late sealer. Sixty-one percent, gone. ✗
Crystal Palace 0–2 Newcastle Home win · 40% Away win Market split two — Osula and Barnes made it a comfortable away win. ✗
Hull 1–1 Brentford Home win · 37% Draw Market split three lands on neither: McBurnie's penalty, Ouattara's equalizer. Brentford's 56% didn't cash either. ✗
Liverpool 2–0 Brighton Home win · 40% Home win The form side got nothing at Anfield — Wirtz and Isak settled it. ✓
Man United 0–3 Bournemouth Home win · 51% Away win Three-nil at Old Trafford: Rayan, Tavernier, Evanilson. With Chelsea, the weekend's worst read. ✗
Sunderland 0–0 Leeds Away win · 37% Draw Goalless at the bottom. The draw was the model's second choice at 30% — second choice pays nothing. ✗

What I take from this week

  • The worst week of the season. Two of ten, a 1.190 log loss, and Elo level on accuracy (2/10, 1.214). Favorites had a bad weekend everywhere — the market managed only 4/10 on Friday's pre-match prices — but we had it worse.
  • The home-lean question is now the main question. The model picked the home team in seven games; home teams won three. All three model-versus-market splits were home picks against the market's away lean, and the market went two-for-three with a draw in the third. Whether the model's home advantage is simply too strong is top of the list for a rethink this week.
  • Confidence did the damage. Chelsea at 61% and United at 51% both lost heavily at home. Log loss punishes exactly that: after the banker, the week's two biggest leans both went the other way.
  • Draw watch rolls on. Three draws, none picked. Nine of the season's forty games have now drawn; the model has made a draw its top pick exactly once.
  • Season tally: 16/40. Forty games still say little about accuracy. The clearer signal so far is structural: a persistent gap between our home prices and the market's.

Matchweek 7 October 17–19, 2026

5/10 correct

Five out of ten again — but this time the Elo challenger takes the week on log loss.

5/10model's correct picks
0.874model log loss (lower is better)
5/10Elo challenger: 5/10 · 0.854
7/10bookmakers: 7/10 · 0.748

model got it right model got it wrong Draw the game was a draw

GamePredictedActualComments
Everton 0–2 Chelsea Home win · 41% Away win The model liked Everton at home; Neto's brace punished it. ✗
Brentford 1–2 Liverpool Home win · 44% Away win Liverpool came from behind; the model overrated Brentford's home edge. ✗
Fulham 1–0 Hull Away win · 34% Home win The coin flip that flipped the wrong way — 34.0% against 34.5%. ✗
Man City 3–2 Ipswich Home win · 85% Home win The banker banked, though 3–2 was nervier than 85% implies. ✓
Newcastle 0–1 Aston Villa Home win · 42% Away win Villa's winner made it three Saturday away victories against the model's home picks. ✗
Bournemouth 2–0 Sunderland Home win · 52% Home win A comfortable mid-range home call. ✓
Brighton 5–0 Crystal Palace Home win · 63% Home win The week's most emphatic scoreline matched its second-strongest lean. ✓
Leeds 1–1 Man United Home win · 38% Draw Derby ended level; the draw was the model's second choice at 29%. ✗
Nott'm Forest 0–1 Arsenal Away win · 54% Away win Arsenal ground it out; the road favorite delivered. ✓
Tottenham 3–1 Coventry Home win · 40% Home win The coin flip that came in — Spurs' first win of the season. ✓

What I take from this week

  • 5/10 again, but Elo takes the week. Same accuracy, better log loss (0.854 vs 0.874). Last week's first outright win over the challenger doesn't repeat; the gap between the two remains noise.
  • Saturday belonged to the road teams. Chelsea, Liverpool, and Villa all won away where the model had the home side, and Arsenal made it four on Sunday. The model picked the home team in three of the four.
  • Coin flips went 1-for-3. Only Tottenham came in. That is roughly what a 34–40% pick means, but calibration needs many more weeks to judge.
  • The blend keeps earning its keep. The ensemble beats v2 alone on log loss again (0.874 vs 0.894), even in a week it loses to Elo.
  • Season tally: 14/30. Thirty games say almost nothing; none of the gaps between the models are statistically meaningful yet.

Matchweek 6 October 10–12, 2026

5/10 correct

Five out of ten — and the first week the model beats the Elo challenger outright.

5/10model's correct picks
1.065model log loss (lower is better)
3/10Elo challenger: 3/10 · 1.078
n/abookmakers (odds feed not yet updated)

model got it right model got it wrong Draw the game was a draw

GamePredictedActualComments
Arsenal 2–1 Leeds Home win · 63% Home win The banker that banked — the model's strongest home lean of the week. ✓
Aston Villa 2–2 Brentford Home win · 39% Draw Model overrated Villa at home; the underlying numbers had Villa bottom of the xG-difference table and Brentford third in xG. ✗
Chelsea 5–1 Bournemouth Home win · 40% Home win The disagreement the model won — the Elo challenger had Bournemouth at 37%. ✓
Ipswich 2–1 Fulham Away win · 42% Home win Model leaned to winless Fulham away; xG had the sides nearly level and Portman Road should have tipped it. ✗
Sunderland 0–2 Brighton Away win · 42% Away win Brighton's attack did the business; the model's joint-strongest away call. ✓
Man United 1–1 Tottenham Home win · 66% Draw The week's biggest miss by confidence. xG agreed with the model — United rank fourth in xG difference — but the finishing problem looks chronic, not luck. ✗
Crystal Palace 3–1 Nott'm Forest Home win · 36% Home win A coin-flip game the model edged right; Elo had Forest. ✓
Hull 2–2 Everton Home win · 42% Draw Eighth vs fifth was always tight; the model's draw price (32%) was its second choice. ✗
Liverpool 0–3 Man City Away win · 45% Away win Right pick; Liverpool finished with ten men after Gravenberch's red card. ✓
Coventry 4–1 Newcastle Away win · 45% Home win The shocker. Promoted Coventry's second statement win in two weeks; Newcastle's underlying numbers were the league's worst and the finishing finally regressed. ✗

What I take from this week

  • 5/10, and ahead of Elo for the first time — on accuracy and log loss (1.065 vs 1.078). No bookmaker benchmark this week: the odds feed hadn't updated past matchweek 5, so there's nothing honest to report. Sixty games still say almost nothing; none of the gaps between the models are statistically meaningful yet.
  • Draw watch continues. Three draws, zero predicted. Draw prices sat between 20% and 32% all week — never the top pick, never far off. Long-run calibration still holds, but calibration over 760 games doesn't pay out in any given ten.
  • Strongest argument yet for xG features. Three of the five misses had the underlying numbers pointing the other way: Villa bottom of the xG-difference table vs third-place Brentford, Ipswich level with Fulham on xG with home advantage unpriced, and Newcastle's attack running on finishing variance (nine goals from 5.29 xG) with the league's leakiest underlying defense.
  • The United miss cuts the other way. xG agreed with the model — United rank fourth in xG difference — and the pick still lost. Persistent underperformance stops being luck and starts being finishing quality or goalkeeping at some point. Distinguishing that from noise is the harder problem, and it's on the list.
  • Coventry get their own note. Two weeks, two statement results for a promoted side with the league's lowest xG. Promoted teams are systematically weird for form-based models — no relevant history, high variance — which is why the excluded-promoted-teams ablation is still on the board.

Matchweek 5 September 18–20, 2026

4/10 correct

The first reviewed week. Ten games, and the model's first outing in its current form.

4/10model's correct picks
1.144model log loss (lower is better)
4/10bookmakers: 4/10 · 1.080
5/10Elo challenger: 5/10 · 1.066

model got it right model got it wrong Draw the game was a draw

GamePredictedActualComments
Brentford 3–0 Chelsea Home win · 47% Home win Strongest home lean of the week; the market slightly favored Chelsea. ✓
Tottenham 2–3 Aston Villa Away win · 35% Away win The model's only away pick of the week — and it disagreed with the market, which had Spurs at 48%. ✓
Brighton 3–0 Arsenal Away win · 45% Home win Model and market both liked Arsenal. ✗
Everton 1–0 Ipswich Home win · 60% Home win The comfortable one — everyone agreed on Everton. ✓
Newcastle 2–1 Hull Draw · 34% Home win Model read it as a coin flip (34/34/32). ✗
Nott'm Forest 0–1 Coventry Home win · 70% Away win Wrong pick, better calibration: v1 gave Coventry 0.04%, v2 gave 7%. ✗
Bournemouth 0–1 Liverpool Home win · 40% Away win Model overrated Bournemouth's strong start; the market favored Liverpool. ✗
Leeds 0–0 Crystal Palace Home win · 55% Draw Nobody saw the 0–0 coming; draws are the hardest outcome for every model. ✗
Man City 5–3 Sunderland Home win · 68% Home win The bankers of the week — City at home. ✓
Fulham 1–1 Man United Away win · 38% Draw Model leaned United (38%) in an open game; draw was its second choice (29%). ✗

What I take from this week

Four out of ten — level with the bookmakers on accuracy, a touch behind on log loss (1.144 vs 1.080), and behind the Elo challenger, which had the best week of the three at 5 of 10. An honest, average week.

A necessary caveat on all of the above, added after a review of this project: fifty games say almost nothing. The model's log-loss edge over the bookmakers across the five evaluated weeks is −0.015 per game with a standard error of 0.045 — statistically indistinguishable from zero. None of the gaps between the models clear that bar either. The honest reading is that the model is competitive with the bookmakers, and the ranking is unresolved until a full season of games accumulates.

The encouraging part: the two games where the model disagreed with the market, it was right both times — Brentford over Chelsea and Villa at Spurs. The draw story also got smaller on inspection: the evaluation window was unusually draw-heavy (32% vs ~26% long-run), and on 760 earlier games the model predicts draws at 25.1% against 25.9% actual — well calibrated. Draws are the hardest outcome for every model, but there's no structural blind spot to fix.

And the Forest–Coventry game did its job as a stress test. The old model's 0.04% was a failure of humility; 7% is a number a reasonable person could hold. Shrinkage works as advertised — though it's worth saying plainly that this one game accounts for nearly all of the measured v1-to-v2 improvement. The fix is principled; the victory lap is premature.

Code

Everything — model, validation, and tuning — is on GitHub.