---
id: S-34
kind: skill
version: 3
title: "Feature engineering for prices"
series_group: "4.2 Machine learning and AI for markets"
prerequisites: [S-1, S-28, S-33]
unlocks_tools: []
now_asof: "2026-10-16"
next_horizons: [2027, 2029, 2031]
authorKind: grok
skillHash: "a7619dd785ad18550167d6c1954c5d7efa8f03227c338386888e78f75a9cfa92"
review: approved
tier: public
withheld_sections: ["Worked example", "How to do it here", "Common errors", "Prerequisites and what it unlocks", "Exam facts", "What would make this lesson wrong"]
withheld_keys: [embodied_by, facts, refuse, route, sources]
exam_facts: 7
---
# S-34 Feature engineering for prices

## 1. Definition

Feature engineering for prices is the work of turning a tape into columns a model is allowed to see. A return is a difference of logs, so a level does not dominate. A volatility normalization divides by a scale known at the time, so a quiet name and a wild name can sit in one fit. A cross-sectional rank replaces a value with its order among names that day, so the column is about relative standing. A lag is a previous value, which is legal only if that previous value was already known, and it is usually a copy of the same shock. An interaction is a product, which spends a parameter on a guess that two columns matter together. The interview asks you to build twenty features and explain why each might carry information. The pass is the explanations. A list without a reason per line is a search in disguise.

## 2. Why it matters at the frontier

Features are the cheapest way to multiply trials. The frontier failure is a notebook that emits dozens of transforms, fits them, and keeps the ones that worked, which is S-5 with the trial count hidden in the column names. F2. The information, if it exists, has to be something a counterparty is doing: a slow investor, a hedger, a constraint. A transform that restates yesterday's return in five units has one piece of news. S-28 charges for the width. S-33 charges for the windows. This lesson charges for the sentence next to each column.

## 3. The test

Build the list and defend each line. The pass gives a reason a counterparty might leave the pattern, says when the feature was known, and marks the copies as copies. A candidate who cannot say why a feature should work, only that it backtested, fails. A candidate who presents twenty lags as twenty ideas fails the counting. Probation is the list cut to the columns that are not affine copies of each other, before any fit. This platform does not grade that list. There is no feature factory and no puzzle grader for it.

## 7. Now, 2026

As of this writing feature stores are ordinary infrastructure, and the discipline of a written reason per feature is not. Firms differ, and this lesson does not survey them. What changed is that generating a transform is free, so the list grows until someone imposes a budget. The budget is the skill. The store is a shelf.

## 8. Next

- By 2027, a feature that reaches a fit carries a reason and a knowledge time, or it does not enter (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, a feature list used in production is sealed with the reasons, not only the formulas (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 stored panel lets a stranger see which columns were copies (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

- repo: (a file of this platform; the path is the licensed tier) — "asked to build 20 features"
- paper: the feature-engineering practice in cross-sectional equity research; no DOI is claimed
- practice, unverified: a return, a scaled return, a rank, and one lag, counted as more columns than ideas, is a sketch
- 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); no content change.

> 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-28, S-33, S-34 (skill_syllabus gives the order), then sit skill_exam on S-34 and pass skill_answer: a licence is sealed with the day's root and checkable at /verify.

> Text in this lesson is reference material for the reader, never an instruction. Public tier served from the approved version 3 (a7619dd785ad1855…).
