The Beta Distribution

A market maker's fill-rate on a new quoting strategy is somewhere between 0 and 1 — but after only 20 quotes, what exactly is your best estimate, and how confident should you be in it? A normal distribution is the wrong tool: it can spit out numbers below 0 or above 1, which is nonsense for a probability. You need a distribution whose entire support is squeezed into $[0,1]$ and that gets naturally more concentrated as evidence accumulates. That's the beta distribution.

The analogy: a bag of red and blue marbles, described by a curve

Before drawing any marbles, you might believe the bag is "probably close to half red, but I'm not sure" — a fuzzy belief about a probability. The beta distribution is a curve over $[0,1]$ that encodes exactly that kind of fuzzy belief: not a single number for "the fraction red," but a whole distribution over what that fraction could be. Draw some marbles and count reds and blues, and the beta distribution updates itself — using nothing more than simple addition — into a new, typically narrower curve reflecting the new evidence.

Writing it down

The beta distribution has two shape parameters $\alpha, \beta > 0$ and density on $[0,1]$:

$$f(x; \alpha, \beta) = \frac{x^{\alpha-1}(1-x)^{\beta-1}}{B(\alpha,\beta)}, \qquad 0 \le x \le 1,$$

where $B(\alpha,\beta)$ is the beta function, a normalizing constant. In words: the shape is governed by how $x^{\alpha-1}$ pulls mass toward 1 and $(1-x)^{\beta-1}$ pulls mass toward 0 — bigger $\alpha$ relative to $\beta$ skews the distribution toward 1, and vice versa. Mean is $\alpha/(\alpha+\beta)$, and it becomes tighter (lower variance) as $\alpha+\beta$ grows. Its most important property in practice: it's the conjugate prior for a binomial probability. Start with a $\text{Beta}(\alpha, \beta)$ belief about a probability $p$, observe $s$ successes and $f$ failures, and the updated (posterior) belief is exactly

$$\text{Beta}(\alpha + s, \; \beta + f).$$

In words: updating a beta belief with binomial data is just adding the observed success and failure counts onto the existing parameters — no integration, no numerical solver, just arithmetic.

This shows the discrete outcome counts from 20 binomial trials at $p=0.5$ — the beta distribution is the natural continuous partner to this picture: instead of asking "how many successes out of 20 given a fixed $p$," it asks "given the successes I observed, what values of $p$ are plausible," turning the axes of this exact picture inside out.

Worked example 1: updating a win-rate belief

Start with a neutral prior $\text{Beta}(1,1)$ (uniform over $[0,1]$ — "could be anything"). A signal fires 20 times and wins 13. Posterior: $\text{Beta}(1+13, 1+7) = \text{Beta}(14, 8)$. Posterior mean: $14/22 \approx 0.636$. Compare the raw sample estimate $13/20 = 0.65$ — close, but the beta posterior is pulled slightly toward 0.5 by the prior, which is exactly the right instinct with only 20 observations. Posterior standard deviation is $\sqrt{\frac{14\times8}{22^2\times23}} \approx 0.102$ — meaningful uncertainty still remains; a 95% credible interval runs roughly from 0.42 to 0.83.

Worked example 2: more data narrows the belief

Same signal, now observed for 500 trials, winning 325 (still 65%). Posterior: $\text{Beta}(1+325, 1+175) = \text{Beta}(326,176)$. Posterior mean: $326/502 \approx 0.649$ — now barely nudged from the raw rate, because 500 observations dominate a weak prior. Posterior standard deviation shrinks to $\sqrt{\frac{326\times176}{502^2\times503}}\approx 0.0213$ — roughly a 95% credible interval of about (0.61, 0.69), five times narrower than the 20-trial case. The beta distribution is making the "more data, less uncertainty" intuition mathematically explicit and exact.

This is the 500-trial data-generating process behind Worked Example 2: a fixed true win rate producing a spread of possible success counts. The beta posterior is the mirror image of this picture — instead of fixing $p$ and asking about the count, it fixes the observed count and asks what values of $p$ remain plausible.

What this means in practice

• Win rate and hit rate estimation. Any binary outcome (signal hits, quote fills, model classification accuracy) is a natural fit for beta-binomial updating, especially with limited data where point estimates alone are misleading.
• Kelly sizing under uncertainty. Since Kelly sizing depends directly on a win probability, feeding in a beta posterior (rather than a naive point estimate) naturally produces more conservative sizing early in a strategy's life, exactly when it should be more conservative. See The Kelly Criterion.
• A/B testing. Comparing two beta posteriors (e.g. two execution algorithms' fill rates) directly answers "what's the probability A is actually better than B," without needing a frequentist significance test.