Negative Binomial Distribution explained for betting

In one sentence

The negative binomial is a count distribution with a separate dial for spread, so it can handle totals that swing more widely than a Poisson would predict.

How it works

Poisson forces the variance to equal the mean. If a match averages 10.5 corners, Poisson insists the variance is 10.5 too. In real data corner totals often have a variance nearer 15 or 16, because some matches are one-way sieges and others are cagey. This is called overdispersion.

The negative binomial fixes it by assuming the underlying rate itself varies from match to match. Picture each game drawing its own "corner rate" from a hat, then Poisson doing the rest. Mixing Poisson over a gamma-shaped spread of rates produces exactly the negative binomial.

The practical result: same average, fatter tails. Big totals and very small totals both become more likely, while totals near the average become a little less likely.

The maths

P(X=k) = \binom{k+r-1}{k}\,p^{r}(1-p)^{k}

\text{mean} = \mu, \qquad \text{variance} = \mu + \frac{\mu^{2}}{r}, \qquad p = \frac{\mu}{\text{variance}}

X is the count (corners, cards, shots).
μ is the expected count.
r is the dispersion or "size" parameter; a small r means lots of extra spread, and as r grows large the model becomes Poisson.
p is a probability parameter linked to μ and r.

In words: you get to set the average and the spread separately, which Poisson does not allow.

Worked betting example

Football corners, illustrative inputs. Your data suggests total corners in a fixture average 10.5 with a variance of 16.

Matching the negative binomial: p = 10.5 ÷ 16 = 0.6562 and r ≈ 20.0 (from r = μ × p ÷ (1 − p)).
P(over 12.5 corners): Poisson 25.80%, negative binomial 28.61%. Fair odds 3.88 versus 3.49.
P(under 8.5 corners): Poisson 27.94%, negative binomial 33.16%. The low tail grows too.
The market offers 3.70 on over 12.5. Backing £10 under Poisson: 0.2580 × £27 − 0.7420 × £10 ≈ −£0.45. Under negative binomial: 0.2861 × £27 − 0.7139 × £10 ≈ +£0.59, or about +£0.43 after 2% commission.

The difference is purely in the shape of the tails. If your variance estimate is wrong, so is this edge.

Where it's good

Corners, cards, fouls and shots, which are almost always overdispersed.
Total runs or wickets counts in cricket formats where scoring is streaky.
Tail markets (high corner lines, many cards) where Poisson underprices both extremes.
As a regression model (negative binomial GLM) when team and referee features drive the mean.

Limitations and pitfalls

You need a solid variance estimate. With 20 or 30 matches, the sample variance is noisy and the tail probabilities swing a lot.
Overdispersion in pooled data may just reflect different teams having different means. Model the mean properly first and some of the "extra spread" disappears.
It still assumes events are counted independently within a match; it does not capture game-state effects such as a team chasing an equaliser.
Referee assignments and team news can change card expectations late; a stale model misses this.
It does not help when the issue is too many zeros specifically. That calls for a zero-inflated model.

How to build it

scipy.stats.nbinom (note its n and p parameters) or statsmodels NegativeBinomial / GLM with a negative binomial family.
R: MASS::glm.nb.
Data: per-match counts with team, venue and referee; compare mean and variance before choosing.
Tip: test whether variance clearly exceeds the mean after modelling team effects; if not, stay with Poisson.

Poisson: the special case with no extra spread.
Zero-inflated models: for an excess of zeros rather than general spread.
Generalised linear models: the regression framework for fitting it.
Gamma distribution: the mixing distribution behind it.
Calibration: to check whether tail prices are right.