Define what was available at the decision instant
A rule can use only information that the defined process had available when it made a choice. A later price path is not such an input. Preserve a timestamped rule version and its input boundary before evaluating any supplied outcomes.
The chronological-evaluation principle is also visible in time-series validation documentation: later evaluation data should not become earlier training information. The cited software describes a particular splitting method, including an equal-spacing condition for comparable folds. This article does not claim that every irregular trade log can be passed to that method unchanged.
Inspect a deliberately invented outcome table
Imagine two rule labels, A and B, fixed before four separate teaching cases. The numbers below are arbitrary outcome units on one stipulated comparison scale. They are not price paths, normalised risk multiples or evidence that different stop distances have equivalent risk.
| Teaching case | Fixed rule A | Fixed rule B | Choose best after result |
|---|---|---|---|
| 1 | -2 | +1 | +1 |
| 2 | +3 | +1 | +3 |
| 3 | -2 | +1 | +1 |
| 4 | +3 | -1 | +3 |
| Total | +2 | +2 | +8 |
The larger after-the-event total uses a different information set
Both fixed columns total 2 fictional units. The row-by-row best column totals 8 because it uses knowledge of each outcome to decide which rule to count. No supplied pre-event selection process generated those choices. Calling the last column an achieved strategy result would misstate the experiment.
The arithmetic does not prove that either fixed rule works. Four invented cases have no empirical performance content. Their purpose is to expose how changing the selection procedure changes the meaning of the reported total.
Keep exploration and later evaluation in different records
It is legitimate to explore a historical table to form a new hypothesis. Label that activity as development, record which cases were inspected and freeze the resulting procedure before evaluating later observations. Do not relabel the development cases as an untouched test set.
Record revisions to inputs, filtering and scoring as well as the stop rule itself. Excluding a case after seeing its loss can introduce the same information leak through another column. A complete report describes the selection process and missing observations, not only the most attractive subtotal.
Chronology needs full dates and a declared time zone
For a US-local review, retain the full date and time-zone reference of each decision and outcome. A trading-session label does not prove their order across overnight activity or different source clocks. Keep the original timestamps and declare any conversion used for display.
A sound process boundary improves the honesty of a comparison; it does not establish a profitable rule, adequate sample size, realistic fills or suitability for live use. These fictional examples contain no live instruction and no tested result for the linked offer.
The optional link below leads to this site's separate crypto signals and Telegram updates offer. The record method is usable without that offer. It does not independently verify the offer, establish country eligibility or turn the teaching example into an actual transaction.
Primary sources and scope
- scikit-learn: TimeSeriesSplit documentation
Chronological evaluation separates earlier training observations from later evaluation. This article does not run or validate a trading strategy. Checked 2026-10-21.