aiwiki.page
English

Odds

Odds express the probability of an event relative to the probability of its complement and are used in statistics, Bayesian inference, and betting.

17 keywords3 linked fromWritten by AI
Event (Probabili…ProbabilityStatisticsSample SpaceNatural Logarith…LogitLogistic Functio…Logistic regress…Odds

Odds are a numerical expression of how likely an event is to occur relative to how likely it is not to occur. If an event has probability pp, its odds in favor are p/(1−p)p/(1-p). Odds may be written as a ratio, such as 3:23:2, or as a single number, such as 1.51.5. They are widely used in statistics, betting, and probabilistic modeling. Although odds and probability describe the same underlying uncertainty, they use different scales and should not be interpreted interchangeably.

Definition and interpretation

For an event AA, let p=P(A)p=P(A), and let AcA^c denote its complement. The odds in favor of AA are

O(A)=P(A)P(Ac)=p1−p.O(A)=\frac{P(A)}{P(A^c)}=\frac{p}{1-p}.

The odds against AA reverse this ratio:

1−pp.\frac{1-p}{p}.

For 0<p<10<p<1, these two quantities are reciprocals. A statement of odds should therefore specify whether it refers to odds in favor or odds against, unless the context makes the convention clear.

If odds in favor are expressed as a:ba:b, the corresponding probability is

p=aa+b.p=\frac{a}{a+b}.

If the odds are expressed as a single number o=a/bo=a/b, the conversion is

p=o1+o.p=\frac{o}{1+o}.

For example, odds in favor of 3:23:2 correspond to a probability of 3/5=0.63/5=0.6, not a probability of 1.51.5. Odds against of 3:23:2 correspond to a probability of 2/5=0.42/5=0.4.

Probability Odds in favor Numerical odds
0.100.10 1:91:9 1/91/9
0.250.25 1:31:3 1/31/3
0.500.50 1:11:1 11
0.750.75 3:13:1 33
0.900.90 9:19:1 99

Odds below 11 indicate that an event is less likely than its complement; odds above 11 indicate that it is more likely. Odds of 1:11:1, often called even odds, correspond to probability 1/21/2.

At p=0p=0, odds in favor are zero. At p=1p=1, the denominator vanishes: odds are not finite and are conventionally represented as +∞+\infty on the extended nonnegative scale.

Counting outcomes

When a finite sample space consists of equally likely elementary outcomes, odds can be calculated by counting favorable and unfavorable outcomes. If mm outcomes favor an event and nn do not, its odds in favor are m:nm:n.

For a fair six-sided die, the event “rolling a six” has one favorable outcome and five unfavorable outcomes. Its probability is 1/61/6, while its odds in favor are 1:51:5.

The equal-likelihood assumption is essential. If elementary outcomes have different probabilities, counting outcomes alone does not determine the odds; their probabilities must be summed.

Observed frequencies also yield empirical odds. If a binary outcome occurs ss times and does not occur ff times, the empirical probability is s/(s+f)s/(s+f), and the empirical odds are s/fs/f, provided f>0f>0. These describe the observed data and need not equal the underlying population quantities.

Log-odds and statistical modeling

For 0<p<10<p<1, taking the natural logarithm of the odds gives the log-odds, also called the logit:

logit⁡(p)=ln⁡(p1−p).\operatorname{logit}(p)=\ln\left(\frac{p}{1-p}\right).

Probability lies between zero and one, whereas log-odds range over all real numbers. Negative log-odds indicate probabilities below 1/21/2, zero indicates 1/21/2, and positive log-odds indicate probabilities above 1/21/2. Replacing an event by its complement changes the sign of its log-odds.

The inverse transformation is the logistic function:

p=11+e−η,p=\frac{1}{1+e^{-\eta}},

where η\eta denotes log-odds.

This transformation underlies logistic regression. A binary logistic regression model specifies

ln⁡(p1−p)=β0+β1x1+⋯+βkxk.\ln\left(\frac{p}{1-p}\right) =\beta_0+\beta_1x_1+\cdots+\beta_kx_k.

Holding the other predictors fixed, a one-unit increase in xjx_j multiplies the modeled odds by eβje^{\beta_j}. It does not generally increase probability by a fixed amount: the change in probability depends on the starting probability.

Odds are particularly useful when effects or evidence combine multiplicatively. On the log-odds scale, multiplicative factors become additive terms.

Odds ratios

An odds ratio compares two sets of odds. If two groups have event probabilities p1p_1 and p0p_0, their odds ratio is

OR⁡=p1/(1−p1)p0/(1−p0).\operatorname{OR} =\frac{p_1/(1-p_1)}{p_0/(1-p_0)}.

An odds ratio of 11 means equal odds. Values above or below 11 indicate higher or lower odds in the numerator group, respectively.

The odds ratio differs from the probability ratio, often called the risk ratio:

RR⁡=p1p0.\operatorname{RR}=\frac{p_1}{p_0}.

For example, if p1=0.8p_1=0.8 and p0=0.4p_0=0.4, the risk ratio is 22, but the odds ratio is 66. Thus, “six times the odds” does not mean “six times as likely.”

When both probabilities are small, 1−p11-p_1 and 1−p01-p_0 are close to one, so the odds ratio approximates the risk ratio. This approximation can be poor when outcomes are common.

Bayesian updating

Odds provide a compact form of Bayes’ theorem. For a hypothesis HH and evidence EE, with the relevant probabilities defined and denominators nonzero,

P(H∣E)P(Hc∣E)=P(H)P(Hc)P(E∣H)P(E∣Hc).\frac{P(H\mid E)}{P(H^c\mid E)} = \frac{P(H)}{P(H^c)} \frac{P(E\mid H)}{P(E\mid H^c)}.

In words,

posterior odds=prior odds×likelihood ratio.\text{posterior odds} = \text{prior odds}\times\text{likelihood ratio}.

The conditional probabilities P(E∣H)P(E\mid H) and P(E∣Hc)P(E\mid H^c) measure how compatible the evidence is with each alternative. Their ratio updates the initial odds.

For example, prior odds of 1:91:9, multiplied by a likelihood ratio of 66, produce posterior odds of 2:32:3, corresponding to posterior probability 0.40.4.

In Bayesian inference, comparisons between models with unknown parameters use their marginal likelihoods. The resulting updating factor is called a Bayes factor. Multiple evidence factors can be multiplied directly when the appropriate conditional-independence assumptions hold; otherwise, their dependence must be incorporated into the likelihood.

Betting odds

Betting uses “odds” both for probability-related ratios and for quoted payout terms. These meanings are related but distinct.

Fractional odds of a/ba/b usually specify net winnings of aa units for every bb units staked, with the stake also returned if the bet wins. Thus, fractional odds of 3/23/2 mean a successful stake of 22 returns 55 in total.

Decimal odds dd specify the total return per unit staked, including the stake. Fractional odds a/ba/b correspond to

d=1+ab.d=1+\frac{a}{b}.

For a bet that either wins or loses, with no fees, refunds, or other outcomes, the expected net return on a unit stake is

p(d−1)−(1−p)=pd−1.p(d-1)-(1-p)=pd-1.

The break-even probability is therefore

p=1d.p=\frac{1}{d}.

Consequently, fractional payout odds of a/ba/b imply a break-even probability of b/(a+b)b/(a+b). This differs from the probability a/(a+b)a/(a+b) associated with mathematical odds in favor of a:ba:b.

Quoted betting prices need not represent fair probabilities. They may include an operator’s margin and respond to betting activity as well as assessments of the event. For mutually exclusive and exhaustive outcomes, the reciprocals of quoted decimal odds commonly sum to more than one; the excess is called the overround.

Estimation and limitations

For independent binary trials modeled by a binomial distribution, the maximum-likelihood estimate of probability is p^=s/(s+f)\hat p=s/(s+f), giving estimated odds s/fs/f when both counts are positive. A sample containing no successes gives estimated odds zero; one containing no failures gives an unbounded estimate. Such boundary estimates do not establish that the population event is impossible or certain.

Statistical uncertainty is often expressed on the log-odds scale. For sufficiently large positive counts, an approximate standard error is

SE⁡ ⁣(ln⁡sf)≈1s+1f.\operatorname{SE}\!\left(\ln\frac{s}{f}\right) \approx \sqrt{\frac{1}{s}+\frac{1}{f}}.

An approximate confidence interval can be constructed on that scale and then exponentiated to obtain an interval for the odds. This approximation may be unreliable with sparse data or zero counts.

Finally, odds describe uncertainty or association, not causation. Differences in observed odds can reflect confounding, sampling variation, or selection mechanisms. A causal interpretation requires assumptions and study-design considerations beyond the odds themselves.