What is a conjugate prior and why is it useful?

Question

AcadiFi · Accepted Answer

A conjugate prior is one where the posterior is in the same family as the prior. This gives closed-form posterior updating — no MCMC needed.

Classic conjugate pairs:
- Normal likelihood, normal prior for mean → Normal posterior
- Normal likelihood (known mean), inverse-gamma prior for variance → Inverse-gamma posterior
- Bernoulli likelihood, Beta prior → Beta posterior
- Poisson likelihood, Gamma prior → Gamma posterior
- Multinomial likelihood, Dirichlet prior → Dirichlet posterior

Kehinde Adebayo at Ibadan Credit Analytics estimates default probabilities for small business loans. Each loan defaults with probability p. He observes x defaults in n loans. Bernoulli likelihood with Beta(alpha_0, beta_0) prior gives Beta(alpha_0 + x, beta_0 + n - x) posterior.

Example: prior Beta(2, 18) captures belief that default rate is around 10%. Observing 4 defaults in 60 loans: posterior = Beta(6, 74), mean = 6/80 = 7.5%, with 95% credible interval (2.9%, 13.4%).

```mermaid
flowchart LR
  A[Prior: Beta a,b] --> C[Bayes Rule]
  B[Data: x defaults of n] --> C
  C --> D[Posterior: Beta a+x, b+n-x]
  D --> E[Closed Form]
```

Why conjugacy matters:
1. Analytic posterior — no numerical integration or simulation
2. Posterior updating is just parameter arithmetic — can update in real time as data streams
3. Interpretable: prior parameters act like 'pseudo-observations' (Beta(2,18) = 2 pseudo-defaults, 18 pseudo-non-defaults)
4. Teaching clarity — students see mechanism clearly

Limits: Real problems often have non-conjugate structure (hierarchical models, mixtures). Modern Bayesian computation uses MCMC to handle non-conjugate posteriors. But when conjugate forms apply (credit scoring, AB testing, simple reliability), use them — they're fast and exact.

What is a conjugate prior and why is it useful?

Master Level II with our CFA Course

Related Questions

Practice Questions