Beta Priors & Robust Kelly Sizing 2026

Beta posterior credible bound for robust Kelly sizing — Bayesian probability distribution with shaded confidence interval — Photo via Picsum

Most people who use the Kelly Criterion plug their probability estimate straight into the formula and bet whatever it spits out. That works perfectly — assuming your probability estimate is exactly correct. In real life, it isn't. Robust Kelly sizing takes your hunch, treats it as the center of a distribution rather than ground truth, and sizes off a conservative bound. This post explains how, using Beta priors and credible intervals.

Why your point estimate isn't enough

Suppose you say "I think this Polymarket market is 60% YES." The textbook Kelly formula treats that 60% as truth. But how confident are you? Could the true number be 55%? Could it be 65%? If you have no answer, you have no business sizing off 60%.

The Kelly formula is steeply convex around the edge. A 5-percentage-point overestimate of your win probability translates into a roughly 40% overbet at the textbook formula. Compounding that over a season of bets and you're not just leaving growth on the table — you're systematically blowing up.

What a Beta posterior actually is

A Beta distribution is the canonical way to represent uncertainty about a probability. It has two parameters, α and β, that you can think of as "successes" and "failures" in your evidence so far.

If you think p = 0.60 and you have evidence equivalent to about 30 observations, your Beta posterior is roughly:

α = p × n + 1 = 0.60 × 30 + 1 = 19
β = (1 − p) × n + 1 = 0.40 × 30 + 1 = 13

This Beta(19, 13) is your belief distribution. Its mean is roughly 0.60, but it has a width — a 95% credible interval might be roughly 0.42 to 0.76. That width is what your point estimate was hiding.

Sizing off a credible bound, not the mean

Here's the robust trick: instead of using p = 0.60, use the lower bound of your credible interval. At 75% credible level (a common middle-ground choice), the lower bound of Beta(19, 13) is around p_cons ≈ 0.49.

That's a much more conservative number. If the displayed Polymarket ask is 50¢, your raw point gives an edge of 10pp; your credible bound gives an edge of −1pp. Robust Kelly says no bet. Your hunch wasn't sharp enough to justify the size.

This isn't pessimism for its own sake. It's the right answer to the question "what's the bet I'd be confident in even if I was a little wrong about the underlying probability?" Kelly is exquisitely sensitive to overconfidence, and the credible-bound layer immunizes you against it.

Picking the confidence level (n)

The strength of your prior is parameterized by n, the effective number of observations. Higher n means a tighter posterior and less shrinkage. Reasonable defaults:

n = 10 (Low confidence): a hunch backed by general reasoning, no specific data
n = 30 (Medium confidence): research-backed view with at least one piece of decisive evidence
n = 100 (High confidence): thoroughly investigated with multiple supporting data points and a track record of similar calls

It's almost impossible to be at n = 100 honestly. Most retail bettors are at 10-30. Be honest with yourself — overstating confidence is the dominant way Kelly sizing goes wrong.

Picking the credible level

The credible level (e.g. 75%, 90%) controls how aggressive your bound is. Higher level = more conservative = smaller bet.

50%: barely any shrinkage — basically the median, close to your point estimate
75%: standard robust default — meaningful shrinkage
90%+: hardcore conservative — only bets you're very confident in survive the cut

Pick higher when your stakes per bet are large or you're new to a market. Pick lower when you have well-calibrated edges and want to capture more of them.

Robust shrinkage vs Bayesian Kelly

One subtlety worth flagging: this isn't "Bayesian Kelly" in the strict sense. The classical Bayesian Kelly under log utility actually says to size off the posterior mean, not a credible bound. The math is exact and the result might surprise you: it tells you to bet at your point estimate, not below it.

The credible-bound approach is a robust optimization instead — it picks the worst-case probability inside your credible region and sizes for that. It's a different philosophical choice, more conservative than Bayesian Kelly, and in practice it produces drawdowns operators can actually live with.

How to actually compute the bound

For a Beta(α, β) distribution, the lower bound at credible level CL is the inverse CDF evaluated at the (1 − CL)/2 percentile. For CL = 75%, that's the 12.5th percentile. There's no closed form — the standard approach is the regularized incomplete beta function and a bisection root-finder.

You don't have to do this by hand. Our Kelly calculator implements Numerical Recipes-style lgamma + betacf for the Beta CDF, then a 60-step bisection for the inverse. Plug in your point estimate, n, and CL — out pops p_cons.

Pros and cons of robust Kelly sizing

Pros

Immunizes against estimate overconfidence — the dominant failure mode
Visible, interpretable safety margin (you can see how much shrinkage you're applying)
Composes cleanly with fractional Kelly — robust shrinkage first, then a Kelly fraction on top

Cons

Conservative by design — sometimes the right bet is bigger than this method allows
Requires you to honestly pick n and CL, which is its own judgment call
Not the textbook Bayesian Kelly answer (it's robust optimization instead)

Frequently asked questions

What's a sensible default credible level?

75% is the practical sweet spot. It applies meaningful shrinkage without being so conservative that legitimate edges fail to size up. Bump to 90% if you're sizing large or new to a market.

Does this make Kelly too cautious?

By design, yes. The whole point is to absorb estimate noise. If you're confident your point estimate is right, you can drop the credible level toward 50% (no shrinkage), but be aware that's the failure mode that wrecks bankrolls.

How does the Beta posterior get its parameters?

The standard parameterization is α = p̂·n + 1 and β = (1 − p̂)·n + 1, where n is the strength of your prior in equivalent observations. This is a Beta(1,1) uniform prior updated with n imagined trials having p̂ success rate.

Bottom line

The single most important upgrade you can make to your Kelly sizing is to stop trusting your point estimate. Treat your hunch as the center of a Beta posterior, take a conservative lower bound, and size off that. Our calculator handles the math — you just need to be honest about how confident you really are.

Size Your Polymarket Bets Like a Pro

Try our free Kelly Criterion calculator — Bayesian shrinkage, walk-the-book sizing, 10k-trial Monte Carlo. Built for serious Polymarket traders. Auto-fills from live ProbBrain signals.

🎯 Open the Kelly Calculator →

18+ only. Gamble responsibly.

Search This Blog

Crypto Bonus Hub