Skip to model research
Systematic Trading ModelsSpecification / Evaluation / Release

Causal features and chronological label boundaries

Audit availability times, train-only transformations and forward targets before trusting a trading-model split.

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.

Availability is stricter than a row date

A row timestamp does not necessarily say when every field was observable. An announcement can arrive after the session; a fundamental series may be revised months later; an aggregated bar becomes final only after its closing boundary. Record event time, availability time and revision policy for each input. The feature vector must be constructible at the declared decision clock using the allowed source version. Sorting a table by date does not remove future revisions or join leakage. Trace one decision end to end before treating a large chronological table as a causal dataset.

Fit transformations inside the training scope

The scikit-learn pitfalls guidance distinguishes fitting from application and explains why preprocessing can leak evaluation information. Scaling, imputation, feature selection and learned encodings need the declared training scope. A pipeline can preserve that scope when it is used correctly; a separately fitted transformation using the full dataset can still break it. Save fitted parameters and code identity. Apply the frozen transform to later observations rather than silently recomputing its historical outputs with a later fit. A deterministic indicator may have no learned parameters but still use future observations through its window definition.

A target can cross an otherwise clean split

Suppose a label uses the next five observations. A decision near the end of training may depend on outcomes inside validation. Our fixed-horizon planner excludes such decision rows from their region's eligible label population. With 1,200 observations, 100 warmup rows, a 60% development allocation and 20% validation allocation, the post-warmup counts are 660, 220 and 220. A five-observation horizon leaves 655, 215 and 215 eligible decisions under that stated convention. The excluded tails are not unexplained lost trades; they are part of the experiment accounting.

A generic gap is not a universal cure

TimeSeriesSplit provides chronological splitting and an optional gap. That is a useful mechanism, not a full audit of variable horizons, event overlap or external source clocks. If a label ends when a position exits, its information interval can differ by observation. If features use a long causal history, earlier observations may be legitimate warmup while their future targets are not. Declare the rule that determines eligibility at each boundary and test concrete examples. Do not advertise a fixed row gap as proof that every possible dataset and label policy has been purged of leakage.

Separate selection from reserved evaluation

Development and validation can be used within the stated selection process. The reserved region must retain its declared purpose and cannot become a repeated tuning surface. Save every candidate, not just the best result, and record when the evaluation was first opened. A later revision can be researched, but the previously inspected region is no longer untouched evidence for it. If an operational refit is part of the policy, its allowed inputs and schedule must be frozen in advance. Refitting retrospectively on the complete history cannot be disguised as an unchanged chronological experiment.

Keep a boundary receipt readers can inspect

Export the planned ranges, warmup count, fixed label horizon and eligibility counts. Add source versions, feature availability examples, transformation fit scope and a software fixture showing a label that should be excluded. These receipts demonstrate how the implementation is intended to behave; they are not independent economic outcomes. The retained backtest checklist goes further into instrument, fills, exits and costs. A model can pass causal split arithmetic and still fail an execution study. Treat that possibility as part of the research contract rather than replacing the study with a neat train-test diagram.

Trace one excluded label

Take the planner default and inspect development decision index 760. With a five-observation forward target, its label reaches beyond the development endpoint, so it cannot belong to the eligible development label set under this contract. Index 755 can resolve its five following observations within that endpoint. Draw those two intervals and retain them as a fixture. Next describe a variable-exit target where the end must be checked per observation. That second problem cannot be solved by pretending every label has the same fixed horizon.

Source documents

Documentation establishes the stated mechanism, not tested trading performance. Checked 9 October 2026; a later version can differ.

Continue the investigation

Binary prediction count audit

Anonymous rows are not genuine chronological evidence

The probability specimen orders authored positives then negatives within its three groups purely for reproducible arithmetic. It is intentionally not a time-ordered source and cannot be used as a planner validation sample. A real chronological evaluation needs actual decision and label-end identities as well as source availability. Do not infer those clocks from row numbers, CSV order or a file modification date. Software boundaries can be checked on invented indices while independent financial validity remains untested. Document both scopes plainly so a later researcher does not reuse a teaching packet as authentic source history.