Skip to model research
Systematic Trading ModelsSpecification / Evaluation / Release

Chronological holdout planner

Plan chronological observation ranges and fixed forward-label boundary exclusions. This worksheet does not run a backtest.

Original editorial revision: 9 October 2026. Expanded edition: 14 October 2026.

Research design and arithmetic, not model recommendations. Classification scores are not monetary returns; a registered artifact and a passing software test are not independent strategy validation.

Free research tool

Set aside development, validation and untouched holdout rows in chronological order. This plans a testing boundary; it does not fit a model, run trades or certify performance. Your inputs stay in this browser.

Series and split

One row could be a bar, day or event. Keep its time order.
Excluded from scored rows; available as prior context.
The remainder is reserved for one untouched holdout.
Assumes every outcome uses exactly this many later rows. Zero means no forward window.

Chronological boundaries

SegmentRows (1-based)CountUse

Decision rows under this label assumption

SegmentEligible decisionsCountExcluded at end

A holdout stops being untouched once you inspect it and change the model. For a new trial, freeze the rules first and use genuinely later observations.

What this boundary protects

  1. Development: fit parameters and compare candidate rules without pretending those rows are independent confirmation.
  2. Validation: decide whether the frozen candidate survives a later period. Repeated changes after seeing validation results consume that period too.
  3. Holdout: evaluate the final frozen candidate once. Record failures as carefully as successes, including costs and the actual fill convention.

Work the default example

With 1,200 ordered observations and 100 warmup rows, 1,100 rows remain. At 60% development and 20% validation, the boundaries contain 660, 220 and 220 rows. A five-row forward label leaves 655, 215 and 215 eligible decision rows: development decisions end at row 755 so their last label can end at row 760. The final five rows in each segment are excluded as decisions, not deleted as price observations.

This calculation assumes a fixed label window of exactly the entered number of future rows. Variable holding periods, overlapping event labels, revised inputs and delayed data availability need their actual event times. It is not a general purging or embargo implementation. Compute rolling indicators causally and fit preprocessing on development data only. The TimeSeriesSplit reference illustrates chronological folds and a training gap; a full walk-forward study repeats train-then-test windows rather than using this single split.

Freeze the rule, parameter selection, data-cleaning decisions and fill policy before opening the final holdout. Retain the exported boundary plan with the experiment contract, data version and failed trials. The export does not record actual timestamps or establish that a researcher kept the holdout untouched. If the model changes after seeing holdout results, that period has become research data and a genuinely later test is needed.

Continue the research procedure with the backtest audit checklist, including a worked five-observation label boundary, fill ambiguity and the experiment records to retain.

Beyond the fixed-horizon diagram

The planner's five-row exclusion is an exact consequence of its entered fixed horizon, not a universal purge rule. A variable label needs its own end observation and overlap treatment. Feature availability must be checked independently, including late events and revisions. Warmup is causal context when the declared feature rule permits it; it is not permission to borrow future labels. Changing split shares changes boundaries, not the source's causal quality. Preserve the exported planning-only marker when attaching these ranges to a future model card. No trade count or research admission follows from the total eligible decisions.

Model-card and probability references