Football betting models

How to Build an Over/Under 2.5 Goals Model

The statistics, contextual factors and probability methods that can be used to estimate whether a football match is likely to produce at least three goals.

Expected goalsAttack and defence ratingsPoisson modelsFair oddsAcademic research
The model’s objective

An Over/Under 2.5 goals model should estimate two probabilities: the chance of a match producing three or more goals, and the chance of it producing zero, one or two goals. These probabilities can then be converted into fair decimal odds and compared with bookmaker prices.

What the model needs to predict

Over 2.5Three or more total goals

Examples include 2–1, 3–0, 2–2 and 4–1.

Under 2.5Two or fewer total goals

Examples include 0–0, 1–0, 1–1 and 2–0.

A strong model normally estimates expected home goals and expected away goals separately. Those two forecasts can then be combined to produce a full scoreline distribution.

Expected total goals = expected home goals + expected away goals

Core team-strength factors

The foundation should be ratings describing how effectively each team attacks and defends, ideally adjusted for the strength of its opponents.

Home attacking strength
Goals and expected goals created at home relative to the league average.
Away attacking strength
Goals and expected goals created away from home.
Home defensive strength
Goals and expected goals conceded at home.
Away defensive strength
Goals and expected goals conceded away.
League scoring environment
Average goals and home advantage within the competition.
Opponent adjustment
Performance corrected for the quality of previous opponents.

Raw goals alone are noisy. A team may score three times from a small number of chances, or fail to score despite creating several good opportunities. Expected-goals data can help distinguish repeatable chance creation from finishing variance.

Expected goals and chance quality

Expected goals, normally abbreviated to xG, estimate the quality of scoring chances. Useful inputs include:

  • Non-penalty xG for and against
  • xG per shot and shots conceded
  • Big chances created and allowed
  • Touches and passes inside the penalty area
  • Open-play xG, set-piece xG and penalty expectations
  • Post-shot xG, where reliable data are available
Why xG matters: recent goal totals can be heavily influenced by unusual finishing or goalkeeping. xG can provide a more stable estimate of the chances a team is regularly creating and conceding.

Shots and attacking process

Shots contain useful information even when xG is available. An Over/Under model can include:

Shot volume

Total shots, shots on target and shots inside the box.

Shot quality

xG per shot, clear chances and the location of attempts.

Pressure

Final-third entries, penalty-area touches and sustained attacks.

Research by Edward Wheatcroft has shown the potential value of predicted match statistics such as shots and corners as inputs to football forecasts. A related study models the probability that a team scores from a shot and examines whether this improves Over/Under 2.5 forecasts.

Recent form and time weighting

Team strength changes during a season, so matches should not all receive equal weight. Common approaches include exponential decay, rolling windows or dynamic ratings.

Weight of an older match = e-?t

Here, t is the age of the match and ? controls how quickly older results lose influence. This is preferable to relying only on an arbitrary “last five matches” figure.

PeriodPossible roleMain warning
Current seasonPrimary estimate of team levelSmall samples early in the season
Last 8–12 matchesCaptures recent tactical or personnel changesCan overreact to schedule strength
Previous seasonUseful prior informationSquads and managers may have changed

Match context and tactical factors

Expected match dominance
Asian handicap, win probabilities, Elo difference and expected possession.
Playing style
High pressing, direct play, transition frequency, defensive line and tempo.
Game-state tendencies
How teams change their attack rate when leading or trailing.
Formation and manager
Back-three systems, narrow formations and recent managerial changes.
Competition incentives
Cup ties, second legs, relegation matches and games where a draw is useful.
Home advantage
Venue-specific attacking and defensive effects.

Pre-match models cannot know the exact score path, but they can estimate likely scenarios. A strong favourite may dominate possession, while an early goal can either open the match or reduce the leading team’s attacking urgency.

Players, line-ups and availability

Team averages may become misleading when important players are absent. Consider:

  • Expected starting striker and principal chance creators
  • Finishing quality and shot volume of available forwards
  • Goalkeeper shot-stopping performance
  • Centre-back and defensive-midfield absences
  • Rotation, fixture congestion and likely minutes
  • Set-piece takers and aerial threats

A Bayesian player-ability study by Whitaker and colleagues extended a team scoring model with individual player information and tested it on the Over/Under 2.5 market.

Other factors worth testing

FactorWhy it may matterPriority
Rest days and travelCan affect intensity, pressing and team selectionMedium
WeatherWind, heat or heavy rain may alter chance creationLow to medium
RefereePenalties, added time and foul patterns may have a small effectLow
Pitch and venueSurface, dimensions and altitude can influence styleLow to medium
Market pricesGoal line, match odds and Asian handicap contain collective informationHigh as benchmark
Avoid adding variables simply because they sound plausible. Every factor should improve performance on matches that were not used to train the model.

Statistical modelling methods

Independent Poisson model

The simplest serious approach assigns separate Poisson distributions to home and away goals.

Home goals ~ Poisson(?H)    Away goals ~ Poisson(?A)

Once the two expected-goal rates are known, calculate the probabilities of every scoreline. The Over 2.5 probability is the sum of all scoreline probabilities where the two teams score at least three goals in total.

Dixon–Coles model

The Dixon–Coles method adds time weighting and a correction to low scorelines such as 0–0, 1–0, 0–1 and 1–1. It remains one of the best-known foundations for football score modelling.

Bivariate Poisson and dependent models

Independent Poisson models assume the two teams’ goal totals are independent after accounting for their expected rates. Bivariate Poisson and other dependence models relax that assumption and can improve the representation of draws and related score patterns.

Bayesian hierarchical models

Bayesian models estimate attack and defence strength while partially pooling information across teams. This can reduce extreme estimates when a team has played only a small number of matches and can naturally quantify uncertainty.

Machine learning

Gradient boosting, random forests, logistic regression and neural networks can predict Over/Under directly. They may handle nonlinear interactions effectively, but they still require careful calibration and chronological validation.

Turning probabilities into fair odds

Suppose the model estimates a 56% chance of Over 2.5 goals.

Fair Over odds = 1 ÷ 0.56 = 1.79

The Under probability is 44%, giving fair Under odds of approximately 2.27.

OutcomeModel probabilityFair odds
Over 2.556%1.79
Under 2.544%2.27

If the bookmaker offers 1.95 for Over 2.5, the model’s estimated expected value is:

Expected value = (0.56 × 1.95) - 1 = 0.092, or 9.2%

This is a model estimate, not a guaranteed return. Small errors in the estimated probability can remove the apparent advantage.

A practical model-building workflow

1
Build a clean historical dataset

Store match date, competition, teams, goals, xG, shots, odds, line-ups and relevant context.

2
Create pre-match features only

Every input must have been available before kick-off. Avoid accidental use of future information.

3
Estimate home and away scoring rates

Start with a transparent Poisson or Dixon–Coles model before testing more complicated methods.

4
Generate the score matrix

Calculate probabilities from 0–0 through a suitable maximum score, then sum all totals above and below 2.5.

5
Validate chronologically

Train on older seasons and test on later seasons. Do not randomly mix future matches into the training data.

6
Compare with bookmaker closing odds

Measure whether the model adds information beyond a strong market benchmark.

How to evaluate the model

Accuracy alone is not enough. A model that labels most matches “Over” may have a reasonable hit rate but poor probabilities.

Log loss

Penalises confident probabilities that are wrong.

Brier score

Measures squared error between probability and result.

Calibration

Checks whether events forecast at 60% occur about 60% of the time.

Closing-line value

Checks whether selections consistently beat the market’s later price.

Research on betting models increasingly stresses the importance of well-calibrated probabilities rather than selecting models only for classification accuracy.

Academic papers and further research

The following papers are useful starting points. Some links lead to journal abstracts, institutional copies or freely available preprints.

A Profitable Model for Predicting the Over/Under Market in Football

Edward Wheatcroft, International Journal of Forecasting, 2020

Directly examines the Over/Under 2.5 market and introduces Generalised Attacking Performance ratings. It tests value-betting strategies across European leagues.

View the paper via DOI ?

Modelling Association Football Scores and Inefficiencies in the Football Betting Market

Mark Dixon and Stuart Coles, Journal of the Royal Statistical Society, 1997

The classic dynamic Poisson model with recency weighting and a low-score correction. It remains a standard reference for football score forecasting.

View the paper via DOI ?

Bayesian Hierarchical Model for the Prediction of Football Results

Gianluca Baio and Marta Blangiardo, Journal of Applied Statistics, 2010

Uses hierarchical Bayesian methods to estimate team attack and defence strengths while accounting for uncertainty and partial pooling.

View the paper via DOI ?

Analysis of Sports Data by Using Bivariate Poisson Models

Dimitris Karlis and Ioannis Ntzoufras, The Statistician, 2003

Explores bivariate Poisson methods for related sports scores and discusses applications to football results.

View the paper via DOI ?

Forecasting Football Matches by Predicting Match Statistics

Edward Wheatcroft, 2020 preprint

Investigates whether predicted shots, corners and other match statistics can add useful information to pre-match football forecasts.

Read the free preprint ?

A Probabilistic Model for Predicting Shot Success in Football

Edward Wheatcroft and Ewelina Sienkiewicz, 2021

Models team-level shot conversion and tests whether predicted shot success improves match and Over/Under 2.5 forecasts.

Read the free preprint ?

A Bayesian Inference Approach for Determining Player Abilities in Football

Gavin Whitaker, Ricardo Silva, Daniel Edwards and Ioannis Kosmidis, 2017

Adds individual player abilities to a hierarchical scoring model and applies the output to Over/Under 2.5 predictions.

Read the free preprint ?

On the Dependence in Football Match Outcomes

Marco Petretta, Lorenzo Schiavon and Jacopo Diquigiovanni, 2021

Questions standard assumptions about dependence between home and away goals and proposes an alternative to the Dixon–Coles structure.

Read the free preprint ?

Extending the Dixon and Coles Model

Rouven Michels, Marius Ötting and Dimitris Karlis, 2023

Examines alternatives to Poisson marginal distributions and applies the extended framework to women’s football scores.

Read the free preprint ?

Recommended first version

A sensible minimum viable model would include:

  1. Home and away attack ratings
  2. Home and away defence ratings
  3. League average scoring and home advantage
  4. Time-decayed xG for and against
  5. Shot volume and shot quality
  6. Opponent-strength adjustment
  7. Confirmed line-up and key-absence adjustments
  8. A Poisson or Dixon–Coles score distribution
  9. Chronological out-of-sample testing
  10. Comparison with margin-free bookmaker probabilities
Best development path: begin with a transparent statistical model, establish a reliable benchmark, and add one group of variables at a time. Keep a new feature only when it improves calibration and out-of-sample performance.
Responsible gambling: Statistical models cannot guarantee a profit. Betting markets contain uncertainty, bookmaker margin and limits. Only gamble with money you can afford to lose, and treat model probabilities as estimates rather than certainties.