Skip to the evidence
AI Trading SignalsDefinition / Population / Proof

AI claims / Consumer evidence

Walk-Forward AI Models: Identity and Information Cutoffs

Reconstruct which model version existed at each prediction instead of attributing an entire moving-window history to one final artefact.

A walk-forward label needs a schedule

A sequence of training and prediction windows must specify what was fitted, which information was available and when each artefact became effective. Preserve model, features, data and transform versions for every segment. A final model file does not by itself reconstruct earlier predictions. The scikit-learn chronological split reference helps describe sample ordering, while the vendor's actual training records must establish its chosen windows and runtime transitions.

A split helper's gap is not automatically an exchange-session purge or a solution to every variable-horizon target. Identify the actual endpoint of each relevant training label and information release. If a daily dataset uses final revised observations, earlier row dates do not prove those values were available then. The existing leakage protocol retains that broader enquiry; this chapter adds the version chain needed to review a repeated retraining claim.

Training completion and deployment differ

An artefact can finish training before it becomes the version used for predictions. Keep the completion time and effective-use time separate. A research LIVE label, a filename, code presence or a merge does not prove the deployed version or subscriber route. Request the actual version binding for the specific output. If the record does not contain it, mark the output's producing model unknown rather than selecting the newest available file.

Preserve the original transitions

Do not overwrite segment A's outputs with predictions recomputed by segment B after more data arrive. A correction can be retained as a new version without erasing the original source. Distinguish a retrospective demonstration of the schedule from a prospective record that was captured at each step. Both can be useful, but only the latter supplies its own original chronology when the required receipts and sources are actually available.

A completed consumer review

List training cutoff, target endpoint, artefact identity, transformation identity, effective use and output population for each segment. Missing data or unsupported calibration remain gaps, not evidence that a full strategy failed. Tests of this publication's calendar checker can establish its own comparisons only. A provider's complete research process, independent validation and production certification cannot be inferred from that small date exercise or from a platform's ability to run model training.

Completed fictional evidence case

In an invented schedule, artefact A is trained using records available by day 10 and begins predictions on day 12. Artefact B uses information through day 20 and becomes effective on day 22. Predictions from days 12-21 must retain A's identity unless actual version bindings show otherwise. Recomputing them with B does not establish the original history. The reviewer records both segments and leaves missing runtime receipts unknown. The days are synthetic sequence labels, not trading-session dates, and the example trains no market model or deploys an automated service.

Evidence objectRecorded valueClaim boundary
Artefact ACutoff 10 / effective 12Earlier segment
Artefact BCutoff 20 / effective 22Later segment
Recomputed old outputsNew calculationNot original chronology
An original claim, its available references and the bounded review
Illustrative record flow; no vendor experiment or customer account is pictured.

Direct sources and reviewed scope

Public documents, not private vendor experiments. The worked cases are invented; source dates do not establish a model-validation engagement.

Related evidence questions