Clean launch records protect the history the model will learn

List price is a commercial intention, not necessarily the price condition that produced a purchase. The event line should preserve list price, paid price, explicit discount, bundle membership, campaign reference and delivery terms as separate fields. Fashion forecasting research includes selling price alongside age, attributes, temporal and store features, which makes a reliable price history a candidate model input rather than administrative detail.[S03] The evidence comes from an established retailer, so no coefficient or effect should be imported into the new venture.

Paid price belongs at the order-line grain where possible. A basket-level discount needs a documented allocation method if analysis requires variant prices. The rule should state how rounding and returned quantities are treated, and the allocated components should reconcile to the amount paid. A bundle should retain its constituent products and the bundle price condition. Otherwise, a lower apparent garment price can be confused with a change in demand for that garment alone.

Promotion capture extends beyond the discount percentage. The offer register records eligibility, audience, channel, placement, start and end times, and any coupon or membership requirement. A campaign that only some customers saw cannot be represented by a flag applied to every sale in the same week. Exposure and redemption remain distinct. A transaction can carry the promotion it used, while the offer ledger carries the conditions under which it was available.[S03]

Delivery and collection terms form part of the effective proposition. A customer may respond to the combined garment, fee, timing and collection condition rather than to the ticket price alone. The record should therefore retain the terms attached to the offer and the terms actually applied at fulfilment. Changes after purchase are later events, not reasons to rewrite the original selling condition.

Calendar and event variables need the same provenance. Multichannel fashion research recommends testing holidays, commercially relevant events, extraordinary events and economic conditions for specific regions and channels.[S04] That is a research agenda, not permission to assume that every named day affects every launch. The capture register identifies the calendar source, geography and event classification, then leaves incremental value to chronological validation.

Price and promotion validation should focus on commercial meaning. Negative paid prices, discounts without a defined basis, promotion dates that exclude the transaction, and basket allocations that fail to reconcile all require review. The team should also test the opposite error: a supposedly full-price sale carrying an implicit bundle or delivery concession. These checks keep a demand model from learning administrative artefacts as customer response.[S37]

A return does not erase the history that led to it. The original order, payment and dispatch remain events, while the return adds its own time, quantity, condition and reason. Cancellation similarly needs a position in the process. Cancelling before dispatch and returning after delivery have different fulfilment and working-capital consequences. The forecast contract can later define net demand at a stated cut-off, but the ledger should not precompute that target by deleting the original sale.[S06][S37]

Return reasons serve product learning as well as target construction. Fit, perceived quality, damage, late delivery and an unwanted substitute describe different failure mechanisms. They should use a controlled taxonomy with an unknown state and room for a verified note. A reason recorded by the customer, inferred by staff and assigned after inspection should not carry the same status. The forecast target can remain net fulfilled purchases while the reason fields inform separate product and service reviews.

Outcome maturity is essential at every evaluation cut-off. At any forecast origin, some orders have not had a realistic opportunity to be delivered, cancelled or returned. The record should mark those outcomes as open rather than treating them as retained sales. Later arrivals create a new target version with both original event times and arrival times. This prevents a replay from benefiting from a final net-sales value that was unknowable when the earlier decision was made.[S37]

Substitution needs a chain of custody. The requested variant, suggested alternatives, customer acceptance, delivered variant and subsequent return should remain linked where the operation supports that detail. Availability and substitution are linked in retail demand analysis, so the alternative set at the moment of purchase belongs beside the transaction.[S06] A delivered substitute should count as demand for the delivered product under the physical-sales contract, while any inferred transfer to the requested product remains a modelling assumption.

The aggregation rule must be consistent across reporting views. Gross orders, fulfilled units, retained units and net units after cancellation answer different planning questions. They can reconcile without being collapsed into one field. A weekly report should identify the cut-off and maturity rule applied to each. The raw ledger remains event based, so another defensible target can be constructed later without requesting a new extraction whose historical definition has changed.[S0701]

This approach also exposes operational feedback across the garment range. A poor size offer can create stockouts in some variants and returns in others. A substitution rule can lift recorded sales while weakening satisfaction. Those are hypotheses to investigate with linked evidence, not conclusions to assert from a return rate alone. The value of the ledger is that it preserves the sequence needed to test them.

An external feature needs a value time and a knowledge time. The value time identifies the period the observation describes. The knowledge time records when that value became available to the venture. Publication and ingestion timestamps, source version and revision history complete the chain. Without them, a backtest can join a launch to a later revised series and call the result foresight. Leakage research specifically covers spurious relationships created through data collection and preprocessing, so as-of capture belongs in the raw evidence layer rather than in a final modelling disclaimer.[S37]

Search data illustrates the required restraint. Research on consumer behaviour found that search queries can guide near-future prediction in some settings, while results varied by application and baseline.[S34] A launch register should therefore preserve query definition, geography, collection method, observation interval and retrieval time. A later change in query grouping or geographic mapping creates a new feature version. The model must test incremental value against the existing prior; the capture system must not assume that a larger search number represents additional purchases.

Weather, macroeconomic and trade inputs follow the same contract even when their publication rhythms differ. Empirical work on retail sales found that unusual weather has a modest role in explaining monthly fluctuations, which justifies recording weather observations and forecast vintages separately rather than substituting realised weather for what was predicted before the selling week.[S35] Trade series require similar care: country-reported flows are submitted to international compilers on varying schedules and often show bilateral mirror asymmetries between reporting partners.[S41][S42] Each record therefore states the source, reporter, partner, original period, unit, transformation and revision status. Missing means not yet published or not successfully collected, never zero by default. Where a source issues a correction, the earlier release remains attached to any forecast origin at which it was the available value.

Exhibit 33 shows the intended replay rule with a constructed publication sequence. The first forecast can use only the initial release available at its origin. A later replay may use the revision if its own origin permits it. The dates are illustrative planning positions, not a claim about a particular publisher's lag.[S37][S34]

The exhibit also clarifies what a feature snapshot must contain. A file of current values is insufficient. The snapshot pins every source record and transformation used at the forecast origin, including missingness and any imputation fitted on the training evidence. Preprocessing fitted across the full history can leak later information even when each individual value carries a plausible date. The capture layer makes the necessary separation possible; the evaluation protocol must still enforce it.[S37]

A practical registry gives every external feature an owner and a permitted forecasting role. The owner monitors failed retrieval, changed definitions and unexpected revisions. The permitted role states whether the feature may inform a near-term covariate, an analogue screen or only a contextual note. Features that cannot be reproduced as of their claimed origin stay out of launch replay. This is stricter than retaining every interesting signal, and it protects the small launch corpus from hindsight.