Skip to system research
Build A Trading SystemRules / Evidence / Replay

Overlapping labels and evaluation boundaries

Inspect event spans before treating separated rows as independent financial observations.

A row boundary can miss a future label

An entry observation can have an outcome that depends on prices several bars later. If its label span crosses a reserved boundary, assigning its start row to development does not prevent later evaluation information from entering selection. Keep label start, label end and producing exit convention, not only a timestamp used for splitting. A historical transformation can also use future observations or revised values. These are different defects. A calendar division is useful only after the actual information and exposure spans are understood; a neat percentage split cannot by itself establish independence or an untouched economic evaluation.

Use an interval example before a generic gap

Imagine an invented development event beginning at bar 18 whose outcome is known only at 24, while a reserved segment begins at 22. Its label overlaps that segment. A predeclared treatment may exclude that event from selection or choose another design appropriate to the authorised method. The exact treatment and any embargo need the actual horizon and dependence, not an arbitrary gap selected for better results. The scikit-learn time-series split gap is a documented software feature, but it does not automatically inspect every financial label interval, open position or data-release revision for the user.

Warmup and open exposure need separate treatment

Earlier observations may be used causally to build state without scoring their outcomes, depending on the declared method. A position already open at a boundary is another object that can affect later cash and decisions. Record whether it is carried, excluded or separately measured under the original policy. Do not force an unrecorded liquidation just to make the split look clean. The small teaching replay has only invented observations and no economic holdout. Its warmup, eligible intent range and buffer tail demonstrate software timing, not a financial purging or embargo certificate.

Preserve the exposed status after a change

Once a reserved result is inspected, changing label horizon, exit rule or split policy using it creates exposure that a new filename cannot remove. Retain the original decision log, source, effective treatment and reason for revision. A correctness fixture can verify that code excludes the specified intervals while remaining irrelevant as an independent market sample. Actual instrument research must follow its full authorised method, resource bounds and separate master requirements. This guide supplies an interval question and explicit missing evidence, not a universal split ratio, sample-size guarantee or permission to optimise a new exit on the same viewed period.

Filled invented system case

The fictional development event spans bars 18 through 24; a reserved boundary is 22. Its start is earlier, but its producing label reaches into the reserved interval. The review marks overlap instead of calling the row independent. A fixed row gap would need its own justified contract. No actual label population, market dependence or economic validation is supplied. The BATS prefix checks concern invented software behaviour and do not resolve this financial evaluation design.

ObjectRecordBounded finding
Event start18Insufficient alone
Outcome known24Label end matters
Reserved boundary22Overlap present
Design treatmentPredeclare for actual methodNo universal gap inferred
Invented closes and moving averages above simulated equity
Invented library replay, not historical data or economic validation.

Primary references and scope

Expanded source review: 13 October 2026. Documentation supports the stated mechanism, not a tested account, hypothetical fill or provider outcome.

Related research questions