FootballUEFA Europa LeagueChecking data
Models

How The Forecasts Work

weights from UEFA Europa League
Every sport

One Model For Each Sport

each game forecast from earlier games only

Each sport has its own model, built for how that sport is scored. Every one is tested the same way: each past game is forecast using only the games before it, then compared with what happened and with a simple yardstick. Lower error is better; beating the yardstick means the model knows more than the basic rate.

Football

21 club competitions

  • Eight models (Elo, strength ratings, a goals model with the Dixon-Coles low-score correction, regression and more) each forecast every match.
  • They are blended, each weighted by how well it predicted that competition's past matches, then calibrated.

52% of 1,138 graded UEFA Europa League matches called right.

Also: correct scores, both teams to score, handicaps, halves, half-time/full-time and season simulations.

American football

NFL

  • Elo ratings give each game's expected margin, adjusted for home field, rest after a bye and a backup quarterback starting.
  • Points ratings for attack and defence give the expected total, lowered in strong wind.
  • Win chances, spreads and totals come from normal spreads around those numbers.

64.8% of 7,167 games called right, 2000 to 2026. Error 0.628 against 0.613 for bookmakers' closing line.The closing line is the market's final word, the hardest yardstick there is: few models beat it, and coming close shows the model is in the right range.

  • Halves and quarters: periods tied far more often than a smooth spread gives, so ties are fitted per period.
  • Season futures: playoffs, division, conference and Super Bowl from 10,000 simulated seasons.
Basketball

NBA

  • Elo ratings give the expected margin, with home court and a penalty for the second night of a back-to-back.
  • Points ratings around the league's moving scoring rate give the expected total.
  • Win chances, spreads and totals come from normal spreads around those numbers.

65.9% of 19,391 games called right, 2011-12 to 2026-27. Error 0.615 against 0.683 for the home side's usual win rate.

  • Halves and quarters: a share of the margin and total per period (only 15% of the margin in the fourth quarter).
  • Season futures: play-in, playoffs, division, conference and title from 10,000 simulated seasons.
Ice hockey

NHL

  • Attack and defence ratings for every team give each side's expected goals over 60 minutes, with home ice.
  • A Poisson grid of scores gives the result, puck line, totals and final scores.
  • A tie after 60 minutes goes to overtime and the shootout as a coin flip, which is how they have played out.

58.1% of 19,604 games called right, 2011-12 to 2026-27. Error 0.673 against 0.690 for the home side's usual win rate.

  • Periods: each period's goals as a share of the expected goals (the third has the most).
  • Season futures: points totals, playoffs, division, conference and Stanley Cup from 10,000 simulated seasons.
Baseball

MLB

  • Team batting and run-prevention ratings, each starting pitcher's own rating and a factor for every ballpark give each side's expected runs.
  • The ninth inning is played out: the home side does not bat when ahead, and a walk-off ends the game.
  • A tie after nine goes to extra innings.

56.5% of 37,995 games called right, 2011 to 2026. Error 0.679 against 0.691 for the home side's usual win rate.

  • First five innings: the starting pitcher counts more and the bullpen less.
  • Season futures: win totals, playoffs, division, league and World Series from 10,000 simulated seasons.
International football

International

  • World Football Elo ratings built on every international since 1872 give each side's expected goals.
  • Home advantage counts only away from neutral grounds, and competitive matches move the ratings more than friendlies.
  • A Dixon-Coles score grid gives the result, goals and correct scores.

59.0% of 26,025 games called right, 2000 to 2026. Error 0.891 against 1.045 for past home, draw and away rates.

  • Double chance, handicaps, team goals and winning margins from the same score grid.
College football

College Football

  • An Elo rating per team for the margin (home edge, a margin-of-victory update and a pull towards the average each season as rosters turn over), and points-scored and points-allowed ratings for the total.
  • Margin and total are each normal around the model's expectation, which gives the win chance, lines and totals.

73.2% of 11,980 games called right, 2013 to 2026. Error 0.518 against 0.662 for the home side's usual win rate.

  • Alternate spreads and totals, and each team's points, from the same margin and total.
  • Quarters: each period's share of the margin and total, fitted from past period scores.
College basketball

College Basketball

  • An Elo rating per team for the margin (home edge, none on neutral courts such as March Madness, a margin-of-victory update and a pull towards the average each season), and points ratings for the total.
  • Margin and total are each normal around the model's expectation, which gives the win chance, lines and totals.

73.0% of 41,984 games called right, 2019-20 to 2026-27. Error 0.524 against 0.636 for the home side's usual win rate.

  • Alternate spreads and totals, and each team's points, from the same margin and total.
  • Halves: each period's share of the margin and total, fitted from past period scores.
Rugby union

Rugby Union

  • An Elo rating per team for the margin (home edge, a margin-of-victory update and a pull towards the average each season; clubs meeting in the Champions Cup share their league ratings), and points ratings for the total around each competition's scoring rate.
  • Margin and total are each normal around the model's expectation, which gives the win chance, lines and totals.
  • A draw is the chance the margin rounds to zero.

72.5% of 10,123 games called right, 2010-11 to 2026-27. Error 0.549 against 0.637 for the home side's usual win rate.

  • Alternate spreads and totals, and each team's points, from the same margin and total.
  • Halves: each period's share of the margin and total, fitted from past period scores.
Rugby league

Rugby League

  • An Elo rating per team for the margin (home edge, a margin-of-victory update and a pull towards the average each season), and points ratings for the total. Golden-point extra time is played out in the score, so a draw is rare.
  • Margin and total are each normal around the model's expectation, which gives the win chance, lines and totals.
  • A draw is the chance the margin rounds to zero.

66.5% of 1,646 games called right, 2019 to 2026. Error 0.621 against 0.687 for the home side's usual win rate.

  • Alternate spreads and totals, and each team's points, from the same margin and total.
  • Halves: each period's share of the margin and total, fitted from past period scores.
Australian rules

AFL

  • An Elo rating per team for the margin (home-ground edge, a margin-of-victory update and a pull towards the average each season), and points ratings for the total.
  • Margin and total are each normal around the model's expectation, which gives the win chance, lines and totals.
  • A draw is the chance the margin rounds to zero.

67.4% of 1,648 games called right, 2019 to 2026. Error 0.591 against 0.684 for the home side's usual win rate.

  • Alternate spreads and totals, and each team's points, from the same margin and total.
  • Quarters: each period's share of the margin and total, fitted from past period scores.
Cricket

Cricket

  • An Elo rating per team in each competition (a side's Test, ODI and T20 strength differ) from every match since 2001 in Cricsheet's records.
  • A side is at home on the grounds where it plays most of its matches; neutral grounds favour nobody.
  • In Tests, the draw follows the recent draw rate and shrinks when the sides are far apart.

66.2% of 2,486 games called right, ODIs 2003 to now. Error 0.613 against 0.683 for the home side's usual win rate.Tests: 68.6% of decided matches called; franchise T20 leagues are close to a coin toss.

  • Tests: the draw as a third outcome, with its own fair odds.
MMA

UFC

  • An Elo rating per fighter from every UFC fight since 2010: a finish moves the ratings more than a decision.
  • A long lay-off pulls a rating back towards the average; the red corner gets a small edge for what the matchmakers know.
  • The chance a fight goes the distance is how often fights of its weight class and length went to the judges.

57.5% of 7,837 games called right, 2011 to 2026. Error 0.674 against 0.686 for always backing the red corner.Fights are hard to call from results alone: the edge over the red corner is real but small.

  • Goes the distance: decision or finish, from past fights of the same weight class and length.
Football in detail

How A Forecast Is Made

every match, every competition
  1. 1Rate the teams

    Every past result updates each team's ratings; a result from a year ago counts half as much as one from today.

  2. 2Ask eight models

    Eight different models each give their own home, draw and away chances for the match.

  3. 3Blend them

    The chances are combined, each model weighted by how well it has predicted this competition's past matches.

  4. 4Check against reality

    The blend is calibrated so that, say, 60% calls come true about 60% of the time.

  5. 5Work out the goals

    A Dixon-Coles goals model gives expected goals, likely scores, over/under and both teams to score, recalibrated against how many goals were actually scored.

Inside the blend

The Eight Models

weight in UEFA Europa League
  • Dixon-Coles goals13%

    The goals model corrected for low scores (0-0, 1-1) with recent matches counting more.

  • Goals model13%

    Expected goals for each side from attack and defence, turned into a chance for every score.

  • Evidence update13%

    Starts from the goals models' average and shifts it for evidence such as rest and form.

  • Tactical match-up13%

    Adjusts for how the two styles meet; neutral when there is no match-up signal.

  • Match simulation13%

    Plays the match thousands of times from the expected goals and counts the results.

  • Elo rating12%

    A rating that moves after every result, by more when a team beats a stronger side.

  • Strength ratings12%

    One strength number per team, fitted so stronger teams win their meetings more often.

  • Form regression12%

    Learns how gaps in form, attack, defence and ratings have turned into results.

The weights are close because the models have predicted these matches about equally well; no single model drives the forecast. Weights are relearned from each competition's results; pick another league in the sidebar to see its blend.

One match

Eight Opinions

UEFA Europa League
ModelGNK Dinamo Zagreb (CRO)DrawBayer Leverkusen (GER)
Dixon-Coles goals22%29%49%
Goals model21%31%48%
Evidence update26%26%48%
Tactical match-up26%26%48%
Match simulation20%31%49%
Elo rating41%27%33%
Strength ratings39%28%33%
Form regression34%19%46%
Blended forecast28%26%46%

GNK Dinamo Zagreb (CRO) v Bayer Leverkusen (GER), Thu 26 Nov: the next round's biggest match by points. Bold is each model's pick. Where they disagree, the blend leans on the ones with the best record.

Evidence

Why These Settings

re-run on every finished match

The settings in use called 53.1% of results against 52.6% for the basic blend, got 11.2% of exact scores against 10.0%, and were closer to the truth on 430 of 742 matches.

SettingResults calledExact scoresHonesty gapBetter / worse
Basic goals blend52.6%10.0%3.2 pts–
Basic + goal calibration52.6%9.7%3.2 ptssame
Dixon-Coles goals53.1%12.3%2.0 pts430 / 312
Dixon-Coles + goal calibrationin use53.1%11.2%2.0 pts430 / 312

Every finished match (742) re-forecast under each setting from only the matches before it; run Sun 18 Oct, 22:15 UK.

Track record

How Far To Trust It

1,138 graded
Results called52%596 of 1,138Always backing home44%the bar to beatHonesty gap2.5 ptschances vs what happenedExact scores10%112 matches
Best
UEFA Champions League 67% · Primeira Liga 57% · Serie A 55%
Hardest
Scottish Premiership 40% · Premier League 48% · Ligue 1 49%