Empirical Bayes

A fund runs 50 small strategies, each with only a few months of live returns — far too little data to trust any single one's Sharpe ratio estimate on its own. A pure Bayesian approach would want a prior for each strategy's true Sharpe, but where would 50 separate, carefully elicited priors come from? Empirical Bayes offers a shortcut: instead of specifying the prior from outside knowledge, estimate it directly from the data itself — from the other 49 strategies — and let each strategy's estimate borrow strength from the whole group.

An analogy: guessing one batter's true skill using the whole league

A scout wants to estimate one rookie batter's true long-run batting average after only 20 at-bats — far too few to trust the raw number, which might be an inflated 0.400 or a deflated 0.150 purely from small-sample luck. But the scout also has data on hundreds of other players: batting averages cluster in a narrow, well-known range, mostly between 0.220 and 0.300. That league-wide pattern is a prior — not one the scout made up from theory, but one estimated directly from everyone else's data. Combining the rookie's raw 20-at-bat number with this data-derived prior pulls an extreme early number back toward the league-typical range, exactly as much as the small sample warrants. This pulling-back effect is called shrinkage, the signature move of empirical Bayes.

The maths, one symbol at a time

Suppose you have $K$ related units (strategies, batters, stores) each with its own true parameter $\theta_k$ and a noisy observed estimate $\hat\theta_k$. Empirical Bayes assumes the true $\theta_k$'s themselves are drawn from a common population distribution, say $\theta_k \sim N(\mu, \tau^2)$, and estimates $\mu$ and $\tau^2$ directly from the observed $\hat\theta_1,\dots,\hat\theta_K$ across all $K$ units — rather than specifying them by hand. The resulting shrinkage estimator for each individual unit is approximately:

$$\hat\theta_k^{\text{EB}} = \hat\mu + \left(1 - \frac{\sigma_k^2}{\sigma_k^2 + \tau^2}\right)(\hat\theta_k - \hat\mu)$$

In words: start from the group average $\hat\mu$, then move toward the unit's own raw estimate $\hat\theta_k$, but only by a fraction that depends on how noisy that unit's own estimate is ($\sigma_k^2$, its estimation variance) relative to how much true variation exists across units ($\tau^2$). A unit with a very noisy individual estimate (small sample) gets shrunk hard toward the group average; a unit with a precise individual estimate (large sample) barely moves, because its own data is already trustworthy enough to stand largely on its own.

Worked example 1: shrinking Sharpe ratios across strategies

Suppose the 50 strategies' raw Sharpe estimates have group mean $\hat\mu = 0.6$ and between-strategy variance $\tau^2 = 0.25$ (Sharpes genuinely range fairly widely). One strategy, with only 3 months of data, has raw Sharpe $\hat\theta_k = 2.4$ (unusually high, likely partly luck) and a large sampling variance $\sigma_k^2 = 1.0$ from the short history. The shrinkage factor is $1 - 1.0/1.25 = 0.2$. The empirical Bayes estimate is $0.6 + 0.2 \times (2.4 - 0.6) = 0.96$ — far more sober than the raw 2.4, pulled heavily toward the group average because a 3-month sample carries little weight against the group's known spread.

Worked example 2: a strategy with a longer track record shrinks less

A different strategy in the same group of 50 has 5 years of data, giving a much smaller sampling variance, $\sigma_k^2 = 0.05$, with the same raw Sharpe of 2.4. The shrinkage factor is now $1 - 0.05/0.30 = 0.833$. The estimate is $0.6 + 0.833 \times (2.4-0.6) = 2.1$ — much closer to the raw 2.4, because 5 years of data is precise enough that the group prior has far less to add. Same raw number, same group prior, very different shrinkage — driven entirely by how much each estimate can be trusted on its own.

The explorer above shows a single prior-to-posterior update; empirical Bayes runs this same mechanic across many units simultaneously, except the "prior" itself is fitted from the group rather than chosen ahead of time.

What this means in practice

Empirical Bayes is the standard fix for the "many small samples" problem across quant research: shrinking noisy per-stock earnings surprise estimates toward a sector average, shrinking per-strategy Sharpe ratios toward a fund-wide average before capital allocation, or smoothing sparse per-category default rates in credit models (see Bayesian smoothing for rare categories). It's cheaper to implement than eliciting a real prior for every unit, and it directly attacks the regression-to-the-mean problem of chasing whichever noisy estimate happened to look best this quarter.