---
id: S-33
kind: skill
version: 3
title: "Time-respecting cross-validation"
series_group: "4.2 Machine learning and AI for markets"
prerequisites: [S-4, S-5]
unlocks_tools: []
now_asof: "2026-10-16"
next_horizons: [2027, 2029, 2031]
authorKind: grok
skillHash: "55b74d9ac5ec71fc935bf44e3aad370f70ae142f93680b74e6ab3a3af8c7d1d5"
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-33 Time-respecting cross-validation

## 1. Definition

A label in this platform is a window. It starts when the features are claimed to be known and ends when the outcome is known. Ordinary k-fold ignores that window. It will train on a sample whose outcome overlaps the test sample's outcome, so the fit has already seen a piece of the answer. Purging drops a training sample whose label window overlaps any test window. An embargo also drops a training sample that starts within a stated gap after a test window ends, because a return series is serially dependent and the day after a test day is not a fresh draw. Walk-forward is the same idea in one direction: train on the past, test on a later block, move forward. The interview asks you to find the leakage in a standard k-fold on overlapping returns. The leak is the overlap, and the file's linter names it.

## 2. Why it matters at the frontier

Every backtest of a label that spans more than one bar is this lesson or it is an accident. The frontier failure is a research note whose t-statistic was computed on rows that shared their future. F2 on a split. F5 is the split itself, stored, so a stranger can see what was purged. Combinatorial purged cross-validation, which the same file plans, holds out combinations of contiguous groups and builds paths that cover each sample once out of sample. Nothing in that file fits a model. It plans the splits. A beautiful path is not a strategy.

## 3. The test

Find the leakage. The pass points at a training label that ends after a test label has started, and says that sample has to leave the fit. It names the embargo as a second cut for dependence, not as a substitute for the purge. A candidate who says "we shuffled the rows so the folds are fair" fails, because shuffling a time series is how you create the overlap. A candidate who treats a CLEAN verdict as proof the research was honest fails the file's own warning: the linter proves what it can see and reports the rest as unauditable. A model that cannot be audited does not pass. Probation is one overlapping pair and the finding the linter returns.

## 7. Now, 2026

As of this writing purged and embargoed validation is standard in the written methodology of serious cross-sectional research and still easy to skip in a notebook. Walk-forward is the language of trading desks. Which production stack stores the purged set is practice, unverified. What changed is that a random split is one library call, and the leak it introduces is invisible in the loss curve. The window has to be part of the label or the call will look honest.

## 8. Next

- By 2027, a backtest of an overlapping label shows the purge, or it is not shown as evidence (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 model card used for a decision carries the linter verdict and the embargo 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, the same samples can be resplit from stored windows (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) — "a training sample is dropped when its label window overlaps any test window"
- repo: (a file of this platform; the path is the licensed tier) — "export function purgedKFold(samples: readonly CvSample[], k: number, embargoMs = 0): CvSplit[] {"
- repo: (a file of this platform; the path is the licensed tier) — "export function lintSplit(samples: readonly CvSample[], split: Pick<CvSplit, 'train' | 'test'>, embargoMs = 0): LintFinding[] {"
- repo: (a file of this platform; the path is the licensed tier) — "It cannot prove a model clean"
- repo: (a file of this platform; the path is the licensed tier) — "any leak -> LEAKY"
- paper: the purged-cross-validation literature; no DOI is claimed
- practice, unverified: a Monday five-day label and a Wednesday five-day label sharing outcome days is a sketch of overlap
- 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-4, S-5, S-33 (skill_syllabus gives the order), then sit skill_exam on S-33 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 (55b74d9ac5ec71fc…).
