Bind forecast validation to time order and the sample gap

Start with the prediction origin, not disjoint identifiers

An AI-assisted forecasting report can use different training and validation record IDs while still learning from observations that happened after the prediction it claims to evaluate. Identifier separation does not establish that the training information existed at the relevant forecast origin.

This worksheet selects a proposed development-validation plan for one-step forecasting over an ordered series. It supplies no estimator, prediction or score. The existing holdout guide still governs exposure and related-case lineage; the new question is which past observations can support each later validation window.

Keep execution order and data order distinct. Randomizing the order in which candidate configurations are evaluated can serve a comparison design. Randomly distributing timestamped observations into training and validation asks a different question about the data available to a forecaster.

Use the two documented roles at their actual scope

The scikit-learn 1.9.1 TimeSeriesSplit reference describes ordered splits that avoid training on future data and evaluating on past data. It requires equally spaced samples for comparable fold metrics. Training sets accumulate earlier observations; test_size controls test-set size, max_train_size bounds training size, and gap excludes a number of samples at the end of training before testing.

The same-version lagged-feature forecasting example contrasts random splitting with temporal evaluation for its bike-demand setup. It reports overly optimistic shuffled-split error in that example. That motivates inspecting the forecast-origin contract; it is not a current performance result for our application or a universal magnitude of bias.

Both actual bodies are documentation, not an installed library or inspected dataset. No linked dataset, notebook, code download or additional modeling guide was opened. This page quotes none of the example’s measured scores and recommends no production estimator or split size.

Declare two small windows on paper

Invent twelve ordinal observation references, 1 through12, in trusted chronological order with equal spacing assumed only for this teaching plan. Choose two validation windows of two observations each, one excluded observation immediately before each window, and no maximum training-size restriction.

This paper arrangement follows the documented ordering and count roles with explicitly chosen two splits, test size two and gap one. It is not captured splitter output. The reference’s default-size formula is not used for these custom test sizes, and the small counts are not a recommended evaluation design.

A different proposal trains using observation12 while claiming to evaluate a forecast for9. Even with unique disjoint IDs, that proposal fails this past-to-future purpose. Do not repair its label by calling the split random, balanced or reproducible.

Observation9 can become training evidence in the later V-B window under this declared development progression. That is recorded exposure, not an untouched final test. A future final-selection claim needs its own separately controlled evidence and decision history.

Keep omitted samples separate from elapsed time and availability

Our original receipt records three different boundaries: the last permitted training observation, excluded observation identities, and the first prediction origin in the next window. A row count is not a clock. For V-A, one observation is omitted, while the ordinal move from7 to9 spans two steps. No hours or calendar duration are established by these labels.

Actual spacing must be supported from trustworthy timestamps and ordering evidence. If observations arrive irregularly, an assumed equal-duration comparison is unestablished. Do not silently sort, fill missing periods or resample records until a check passes; each changes the proposed data contract.

The gap also does not certify feature or target availability. A lagged field may require an observation that was not yet known at its intended prediction origin. Record when each required input became available, the declared target horizon and the applicable training boundary. Those actual associations are independent prerequisites here, not properties this worksheet’s indices implement.

Likewise, excluding a row from fitting does not mean that every possible use of its information is forbidden or permitted. The application must define the allowed information at each forecast origin. This page supplies no feature-generation, delayed-label or multi-step forecasting recipe.

Admit a bounded development receipt, not a quality result

The ordinary positive conditionally admits only this development-validation plan when every actual prerequisite holds. It reports no fitted model, useful forecast, correct metric or generalization result. A later numerical output still requires supported capture and interpretation.

Keep failed and missing scheduled outcomes visible. Do not change the forecast horizon, drop an inconvenient window or relabel development feedback as final evidence after seeing results. The paired-resampling worksheet remains a separate analysis; no bootstrap or uncertainty procedure is supplied here.

Walk eight inert temporal decisions

Actual timestamps, availability, lineage, supported adapter, training and scoring procedures, resources and rights are unknown or unimplemented. They block a real positive. No split proves universal leak prevention, causal validity, future performance or a suitable sample size.

No splitter, array, datetime, dataset, estimator, fitting, prediction, score, metric, plot, model call or runtime fixture was constructed or executed. Boundary is prelaunch with planned request checks; its synthetic policy evidence does not implement this forecasting evaluator.