Analogue products provide the first outside view

A similar garment sold in another retail system is a hypothesis about transfer, not an established comparator. Multichannel fashion research evaluates physical and non-store channels separately and recommends testing regional and channel-specific covariates rather than assuming a universal relationship.[S04] The analogue market screen should therefore begin with the mechanics through which a customer encounters, evaluates and pays for the offer.

Channel comparability covers assortment control, inventory ownership, customer discovery, payment, delivery, returns and replenishment. A distributor's initial order is not equivalent to repeated consumer sell-through, even if both are recorded as sales in their respective systems. An online launch with immediate delivery is not directly comparable with a made-to-order offer that has a long waiting period. These differences can alter both the level and timing of recorded demand. Where they cannot be reconciled, the launch should move to a scenario pool or leave the central reference class.

Geography adds more than a location label. Climate, calendar, sizing, language, retail concentration and the availability of substitutes can all affect the buying condition. The analogue record should identify which of these are observed and which remain assumptions. It should not convert a broad national similarity statement into an exact adjustment multiplier. A documented mismatch can justify exclusion or a wider prior; it does not justify an arbitrary percentage correction.

Macroeconomic indicators are useful for screening scale and context at this stage, not for predicting a style's weekly sales. The World Bank defines total population using a de facto count of residents and reports midyear estimates.[S44] That provides a clear denominator for a market-scale discussion. It does not identify the subset of residents who share the target product's occasion, price access or channel reach.

Income requires the same restraint. The Bank's GNI per capita series divides an economy-wide income measure by population and expresses it in constant prices using a stated reference year.[S45] An average income measure can help rule out obviously different market contexts and support slow-moving scenarios. It cannot establish clothing expenditure, willingness to pay for a particular garment, or the income distribution among reachable customers.

Urbanisation can describe locational context, but the underlying definitions are national. The World Bank explicitly advises caution in cross-country comparisons because countries classify urban and rural populations differently.[S46] The market register should preserve those definitions and the observation period. A similar urban share does not prove similar retail access, and a national urban measure cannot substitute for evidence about the actual launch cities.

Exhibit 13 separates what each layer can contribute from what it cannot establish. The table is a control on scope. It keeps macro context above the product forecast and prevents channel resemblance from being inferred from a product photograph.

Market and product matching should remain separate until both pass review. A strong product match in a weak market may belong in a downside or upside scenario, depending on the documented difference, but not automatically in the central prior. A strong market match with a weak product match is similarly limited. This two-stage discipline makes it possible to explain whether forecast uncertainty comes from the garment, the selling system or the wider market.

The prior should preserve the launch distribution, not collapse it into a central estimate with an arbitrary band attached. Each admitted curve contributes its level, timing, censoring and outcome status. Slow sellers and withdrawals remain in the readout even when a planner considers them unlikely for the target. Their presence is what allows the outside view to correct the natural attraction of a persuasive launch story.

Three uncertainty sources should remain distinct. Observation uncertainty concerns errors, missing events and demand censoring within an analogue. Between-launch dispersion concerns the variation among eligible outcomes. Transfer uncertainty concerns the possibility that the target will not behave like the reference class even after matching. Combining them into one unexplained interval hides the lever that later evidence can realistically reduce. Early target purchases may reduce the first two under suitable conditions, while a changed channel or market can increase the third.

Launch curves should be aligned by elapsed selling time before their distribution is summarised. The alignment uses the date on which each offer became available in the relevant channel and keeps calendar conditions as separate fields. Availability-censored intervals remain marked, and a withdrawn product is not given an imaginary continuation. Where a correction is not defensible, the report can show the observed curve and a sensitivity case rather than select a single reconstructed history.

Weights should change influence without quietly deleting tails. A weakly comparable strong performer may receive little weight, but its outcome can still inform the upper scenario boundary. A close but disappointing analogue should not be removed because it pulls the median down. The prior register therefore stores both the weighted empirical readout and the unweighted sensitivity readout, together with the independent launch support behind each.

Exhibit 15 illustrates three launch paths that should remain visible together. The central path does not replace the slow or strong case. These indexed curves are deliberately illustrative and do not represent an adoption law, an observed apparel series or a recommended production ramp.

A range formed this way is an evidence summary whose calibration still needs testing. Proper scoring rules provide a principled way to evaluate probabilistic forecasts and encourage honest assessments; they do not certify an analogue distribution before held-out outcomes exist.[S14] The launch review should eventually compare stated quantiles with realised outcomes across comparable cold starts. Until that evidence exists, the report calls the range a prior, states its construction and avoids presenting its bounds as a measured coverage guarantee.

Scenarios sit outside that statistical statement. Forecasting practice distinguishes probabilistic forecasts from conditional views formed under specified assumptions.[S08] A price change, a new distributor or a supply interruption can be explored as a named scenario with an owner and rationale. It should not be smuggled into the prior by widening every quantile until the preferred production plan looks safe.

The analogue prior is the beginning of an evidence sequence. As the target earns its own launch history, the forecast should rely less on borrowed outcomes and more on observed purchases, availability and repeat behaviour. Global models supply another source of shared evidence by fitting one function across series rather than an independent function for each product.[S22] The update record should therefore distinguish analogue weight, shared-model weight and target-history weight instead of describing the whole system as machine learning.

Decay should follow effective evidence, not elapsed time or row count. A week with well-recorded availability, stable offer conditions and clean customer identities may deserve more influence than a longer period with stockouts and changing promotions. Multiple transactions from one institutional order may also carry less independent information than the transaction total implies. The updating rule needs fields for censoring, offer changes, channel mix and dependence before it can assign target weight defensibly.

Hierarchical updating provides a coherent candidate for this transition because shared and local parameters can be estimated together.[S28] Yet the operational rule should remain visible to planners. Each forecast version records the prior used, the new evidence admitted, the resulting range and the contribution still attributable to analogues. A reviewer can then reconstruct why the forecast moved and challenge a transfer assumption without inspecting model code.

Confidence need not increase monotonically. Early sales can reveal that the target occupies a different channel, price or size context from the analogues. A stockout can make the new series less informative than expected. A sudden departure from the reference distribution can indicate that the prior was wrong, that the target definition changed, or that the observation process failed. In each case, widening the range or restoring an earlier prior may be more defensible than forcing rapid adaptation.

The update rule should also specify what happens when target and shared evidence disagree. One option is to retain both forecasts side by side until availability and definitions pass review. Another is to publish a combined range whose additional transfer uncertainty reflects the disagreement. What the team should not do is select whichever output best supports the existing production commitment. Combination and reversion rules are fixed before the launch readout and changed only through a documented review.

A small team can operate this discipline through a compact version register. It contains the forecast identifier, information cut-off, analogue corpus version, taxonomy version, availability correction, pooling specification, target evidence summary and approval status. It also records the reason for every material revision. These artefacts make the prior reproducible and allow later backtests to deny the model information that was unavailable at the original decision.

The final control is a stopping rule for weak evidence. If independent analogue launches are too few, their attributes are poorly observed, or their channels cannot be reconciled with the target, the prior remains wide. The next expenditure goes to better launch records, calibrated research or a limited pilot rather than a more elaborate transfer model. This is a successful application of the method: it identifies what the current evidence cannot support.

Before the first production commitment, the planning owner should receive the reference-class register, attribute scorecards, visual-review exceptions, market comparison, pooled and unpooled sensitivity cases, the full prior distribution and the updating rule. Commercial and production reviewers approve different parts of that pack, but neither can silently rewrite the other's evidence. The launch then has an outside view that can be inspected, challenged and gradually replaced by owned history.