Prem Predict
NextMW 7
Model Labmodel_v3

How good is the model?

Every forecast is graded against what actually happened. See the hit rate, whether the model beats simply picking home wins, how well its confidence tracks reality, and its biggest misses and boldest correct calls.

Evaluation window2026/2027

Grading covers every finished 2026/2027 Premier League match with a stored pre-match forecast (60 matches). Honest out-of-sample numbers — earlier matchweeks are retrospective replays, not picks published at kickoff time.

Season accuracy so far
50.0%
30 of 60 finished matches called correctly · updates each matchweek
Home-team baseline
38.3%
What picking the home side every week would score
Model improvement
+11.7 pp
Accuracy gained over always picking home wins
Test log loss
1.027
From 42 held-out test matches: 40.5% accuracy, Brier 0.617. Lower log loss is better.

Can you trust a 70% prediction?

Calibration checks whether the model's confidence matches what actually happens.

0%0%25%25%50%50%75%75%100%100%0–10%: predicted 2.8%, happened 14.3% (7 forecasts)10–20%: predicted 15.1%, happened 10.0% (20 forecasts)20–30%: predicted 25.8%, happened 33.3% (69 forecasts)30–40%: predicted 34.6%, happened 27.8% (36 forecasts)40–50%: predicted 44.6%, happened 34.8% (23 forecasts)50–60%: predicted 53.9%, happened 50.0% (10 forecasts)60–70%: predicted 63.2%, happened 57.1% (7 forecasts)70–80%: predicted 72.1%, happened 100.0% (4 forecasts)80–90%: predicted 84.0%, happened 50.0% (2 forecasts)90–100%: predicted 92.7%, happened 100.0% (2 forecasts)Predicted probabilityActual frequency

Every forecast across the season is binned by predicted probability (all three outcomes per match). Dots sitting on the dashed diagonal mean the model's confidence matches reality; dot size tracks sample count.

Accuracy by matchweek

How the hit rate moved across the 2026/2027 season.

Each column is the share of that matchweek's matches the model called correctly. The dashed line is the season average.

Biggest upsets

Results the model rated least likely before kickoff.

  1. NFO0 – 1COV
    MW 5 · 19 Sept 2026Gave COV win 1.2% · it happened
  2. CRY2 – 3IPS
    MW 4 · 12 Sept 2026Gave IPS win 15.2% · it happened
  3. COV4 – 1NEW
    MW 6 · 12 Oct 2026Gave COV win 17.1% · it happened
  4. LIV0 – 0FUL
    MW 4 · 12 Sept 2026Gave a draw 21.2% · it happened
  5. LIV2 – 2NFO
    MW 2 · 29 Aug 2026Gave a draw 21.5% · it happened
  6. BHA3 – 0ARS
    MW 5 · 19 Sept 2026Gave BHA win 22.2% · it happened

Best calls

Bold predictions the model got right against the odds.

  1. NEW2 – 1HUL
    MW 5 · 19 Sept 2026Called at just NEW win 36.8% · and it landed
  2. CHE2 – 2HUL
    MW 4 · 12 Sept 2026Called at just a draw 37.1% · and it landed
  3. CRY3 – 1NFO
    MW 6 · 11 Oct 2026Called at just CRY win 37.3% · and it landed
  4. SUN1 – 0FUL
    MW 2 · 30 Aug 2026Called at just SUN win 37.9% · and it landed
  5. BRE3 – 0CHE
    MW 5 · 18 Sept 2026Called at just BRE win 38.7% · and it landed
  6. EVE2 – 0CRY
    MW 1 · 22 Aug 2026Called at just EVE win 39.7% · and it landed