How to Build an Over/Under Corners Betting Model

A useful corners model should not simply compare two teams' recent averages. It should estimate how many corners each team is likely to win, adjust those expectations for the opposition and match conditions, and then convert the expected total into probabilities and fair odds.

The basic modelling structure

A strong starting point is to predict the home and away teams' corners separately.

If the model estimates 6.15 home corners and 4.65 away corners, the expected match total is 10.80.

1. Team corner strength

The four most important starting statistics are:

• Home team corners won at home
• Home team corners conceded at home
• Away team corners won away
• Away team corners conceded away

A simple initial estimate is to average the relevant attacking and defensive figures:

This gives a useful baseline, but it should be adjusted for team quality, opposition, style and match context.

2. Recent form and recency weighting

Use several time windows rather than relying on the last few matches alone. A sensible set might include the last five, last ten and the full season.

Recent figures should also be opposition-adjusted. A team may appear strong because it has recently faced several weak defences.

3. Attacking style and chance creation

Corners are often produced by attacks ending in blocks, deflections or defensive clearances. Useful attacking variables include:

• Crosses attempted
• Shots and blocked shots
• Touches in the opposition penalty area
• Final-third entries
• Attacking-third possession
• Wide attacks and overlapping runs
• Progressive carries
• Passes into the penalty area

General possession is less informative than territorial possession. A side can dominate the ball in its own half without creating much corner pressure.

4. Defensive style and corners conceded

A team that defends deeply may allow repeated crosses and blocked attacks. Relevant defensive variables include:

• Opposition crosses allowed
• Opposition shots and blocked shots
• Defensive clearances
• Time spent defending in the penalty area
• Opposition final-third entries
• Low-block frequency
• Possession allowed in defensive areas

Historical corners conceded remains essential because two defensive teams can allow similar possession but very different numbers of corners.

5. Expected dominance and team strength

The stronger team will often control territory and generate more corners. Useful indicators include:

• Expected goals and expected goal difference
• Elo or power ratings
• Match-result probabilities
• Asian handicap line
• Expected possession
• Home and away strength

The relationship is not perfectly linear. A strong favourite may score early and reduce its attacking intensity, while the underdog is forced to attack more.

6. Game state and goal timing

Teams behave differently when winning, drawing or losing. A team that falls behind often attacks more aggressively, attempts more crosses and wins more corners. A side protecting a lead may sit deeper and concede territory.

A pre-match model cannot know the actual score path, but it can include:

• Probability each team scores first
• Expected goals for each side
• Probability of an early goal
• Probability of a close match
• Probability of a goalless match

An advanced model can simulate different score paths rather than assuming the same attacking rate throughout the match.

7. Match pace and attacking volume

Some fixtures naturally contain more attacking actions. Useful pace indicators are:

• Total shots
• Possessions and transitions
• Final-third entries
• Turnovers
• Direct attacks
• Pass tempo
• Ball-in-play time

Two direct, fast teams are generally more likely to create corner opportunities than two slow, low-risk possession teams.

8. Width, formations and tactical match-ups

Teams that attack through wide areas, use overlapping full-backs or cross frequently can have a very different corner profile from teams that create centrally.

Consider formation and tactical interactions such as:

• A strong winger facing a weak full-back
• Overlapping full-backs against a narrow defence
• A wide attacking side facing a low block
• Wing-backs in a 3-5-2 or 5-3-2
• Two teams that both prefer central attacks

Formation labels alone are not enough. A 3-5-2 can be highly attacking or extremely defensive depending on how the wing-backs are used.

9. Other match-context variables

10. Information from betting markets

The broader betting market contains information about expected dominance, scoring and game state. Potential inputs include:

• Match odds
• Asian handicap
• Over/under goals line
• Both-teams-to-score odds
• Existing corner line and prices

Exclude the bookmaker's corner price if you want a fully independent model. Alternatively, use it as part of a blended model or as a benchmark for measuring your edge.

Choosing a probability distribution

Poisson model

Poisson is the simplest starting point. If the expected total is 10.8, the model calculates the probability of every possible total and adds the probabilities for 11 corners or more.

The main limitation is that Poisson assumes the variance equals the mean. Corner totals often vary more than that assumption allows.

Negative binomial model

Negative binomial regression is often a better practical choice because it allows the variance to exceed the mean. This helps account for volatile matches, tactical changes, red cards and unusual game states.

Converting probabilities into fair odds

Suppose your model estimates a 54% probability of Over 10.5 corners.

The Under probability is 46%, giving fair odds of approximately 2.17.

If a bookmaker offers 2.00 on Over 10.5, its implied probability is 50%. Your estimated edge is four percentage points.

This does not guarantee an 8% return. It is only useful if the 54% estimate is well calibrated and genuinely more accurate than the market.

Minimum viable dataset

The variables to prioritise first

1. Home corners won at home
2. Away corners won away
3. Home corners conceded at home
4. Away corners conceded away
5. Opposition-adjusted corner ratings
6. Crosses for and against
7. Shots and blocked shots for and against
8. Expected possession or attacking-third possession
9. Asian handicap or expected strength difference
10. Expected goals and total-goals line
11. Recent form with recency weighting
12. League average and home advantage

These variables should provide most of the initial predictive value. Add detailed weather, referee and player-level inputs only when they improve out-of-sample performance.

How to test the model properly

Do not judge the model only by the percentage of winning Over or Under selections. Measure:

• Log loss
• Brier score
• Probability calibration
• Return on investment
• Closing-line value
• Results by probability band
• Performance on completely unseen seasons

Frequently asked questions