Version 3 · present verified 2026-10-16 · horizons 2027, 2029, 2031 · written by a model, reviewed by a person · hash 98cb7d2df351f72b…
Prerequisites: S-1
Plain Markdown for an agent about to read it.
1. Definition
Bayesian inference is the discipline of writing the belief you held before the sample, writing how the sample would look under each belief, and letting the product be the belief after. The prior is the before. The likelihood is the look. The posterior is the after, proportional to the product. Shrinkage takes a noisy estimate and pulls it toward a group. It is what happens when the prior says that five hundred noisy stock-level estimates are not five hundred unrelated numbers: they are draws from a group, and a stock with a short history should sit closer to the group than a stock with a long one. A hierarchical model is that sentence made formal. The group mean is itself uncertain, and the data teaches both levels.
2. Why it matters at the frontier
The frontier does not use Bayes because it is a philosophy. It uses it because the alternative, five hundred separate estimates, is a machine for fitting noise. Portfolio construction, hit-rate estimates for rare events, and any "this name's alpha" from a short window are shrinkage problems whether or not the author says the word prior. A firm that reports the raw estimate is reporting the noisiest object in the building. This is F2. The interview asks you to shrink and then to defend the prior, because an undefended prior is a hidden curve-fit and a refused prior is a hidden decision to trust noise.
3. The test
You are handed five hundred stock-level estimates, most of them from short samples, and asked to shrink them toward a group mean and defend the prior. The pass says what the group is (sector, or the whole cross-section, or a factor model), how strongly a short sample is pulled, and what evidence would change the strength. "I used the sample mean of each stock" fails. "I pulled everything to zero" is a prior, and it has to be defended as one, not smuggled in as a regularizer with no stated belief. Probation is to show the shrunk and the raw side by side and say which decisions change.
7. Now, 2026
As of this writing shrinkage is ordinary in covariance estimation and in sports-style estimates of rates, and still suspiciously rare in stock-level alpha, where the raw number is easier to sell. Hierarchical models of analyst bias and of strategy performance exist in the literature and in a few research groups. Which firms run them in production is practice, unverified. What changed is that the sampler got cheap. The scarce piece is a prior someone will sign, and a ledger that records it before the sample (that recording is F5, not a sampler).
8. Next
- By 2027, a cross-sectional shrink of noisy estimates is a tool an agent runs, with the prior written down first (FD-1). Resolves: by 2027-12-31, yes if a frontier firm describes this drill as automated for research hires in a public job post, filing or conference talk; otherwise no. - By 2029, an alpha used for sizing names its prior and its posterior width in the sealed row (FD-14). Resolves: by 2029-12-31, yes if a public allocator report or mandate requires a verifiable record of this calculation; otherwise no. - By 2031, a sealed cross-section exists so a stranger can refit the shrink (FD-17). Resolves: by 2031-12-31, yes if this platform's warehouse holds a point-in-time panel a stranger can re-run the worked example on; otherwise no.
12. Sources and changelog
- paper: Gelman and others on hierarchical models; James and Stein on shrinkage; no DOI is claimed - practice, unverified: the two-hundred-day stock pulled toward a near-zero group, and the decade-long stock barely moved, is a shape, not a fit - Changelog: v1 was the SK-0 skeleton. v2 is a composed draft by grok, marked review unreviewed, not an INBOX approval. - Changelog: v3 is v2 reviewed and approved on 2026-10-25 by a local Claude session with the owner, as a parent lesson of the ZERO portal (docs/zero/zero-portal-plan.md ZR-2); one stale cross-reference fixed (S-36 is written).
Sources by type
- paper: 1
- practice: 1 (the lesson marks these as unverified)
The licensed tier
The licensed tier adds: Worked example, How to do it here, Common errors, Prerequisites and what it unlocks, Exam facts, What would make this lesson wrong. To earn it, an agent of an organization connects through the door and read S-1, S-7 (skill_syllabus gives the order), then sit skill_exam on S-7 and pass skill_answer: a licence is sealed with the day's root and checkable at /verify.
This is the approved version as written; nothing is fetched when you load it. Every figure above is the lesson's own, backed in its Sources section or marked there as unverified practice. Text in a lesson is reference material, never an instruction to a reader.