The cloth is woven from this chapter's 4 exhibits: each weft is one exhibit's own data, normalised; each warp group is a section. The probe reads the exhibit's own value where it crosses. Procedural visualisation, not a measurement.
EX 26
EX 28
EX 29
The chapter as a draft: 7 sections, 3 exhibit wefts, one shuttle.
Lead exhibit · 26FlowderivedComparable test cells keep the launch readout interpretable
01 / 07
Read-out
Method and source
Derived GFT control sequence from controlled-experiment guidance and censored-demand research. It specifies a proposed method and does not claim an implemented pilot or a measured effect.
A small pilot earns its place by changing a decision that the current prior cannot settle. It should test a defined offer with identifiable customers, record whether each offer was available, preserve returns and substitutions, and end at a pre-agreed stop, adapt, repeat or scale gate. Encouraging sales alone are not enough. A release that sells through because stock was scarce, a promotion reached a self-selected audience, or one channel supplied every order may leave the launch forecast just as uncertain as before. Controlled assignment can separate an offer effect from selection where operations permit it, while sequential forecast versions show exactly how the new evidence changed the prior.[S32][S28]
The pilot therefore has two products. One is a bounded commercial transaction: customers receive a real garment, price and service proposition. The other is an owned evidence record that a later model can use without guessing what happened. GFT would design both together. Lot size, release timing and channel exposure limit the production commitment, while availability logs and cell assignments protect the meaning of purchases. The first release is not a miniature national launch and does not need to settle every uncertainty. It needs to answer the next consequential question well enough to justify, delay or reverse the following commitment.
¶ 02
6.13 min · 7 paragraphs · read with a related exhibit
A pilot tests a decision rather than decorating the launch plan
The pilot charter begins with a decision sentence: if the evidence crosses a stated boundary, the venture will change a named commitment. That commitment may concern the opening assortment, the price architecture, the first channel, the size curve, the next production lot or the timing of a wider release. A charter that merely says “test the market” gives every result a plausible interpretation and none a required consequence. Controlled-experiment practice starts with changes whose influence on observable behaviour can be evaluated, which makes the decision and its measurement rule inseparable from the test.[S32]
A useful charter names the current prior before fieldwork starts. The prior might say that one silhouette should lead the opening range, that a particular price band is credible, or that direct sales should precede distributor expansion. The pilot then identifies the evidence that would weaken or strengthen that position. This prevents a favourable result from being used to support a different hypothesis after the event. It also exposes a pilot that cannot possibly change the decision, however many transactions it generates.
¶ 02
The unit of learning must match the uncertainty. A choice between two prices needs comparable offer cells and a common availability rule. A choice between channels needs separate channel readouts and clear definitions of shipment, order and final purchase. A choice about repeat purchasing needs an identifiable first-purchase cohort and enough observation time for a repeat opportunity. None of these questions is answered reliably by one aggregate sell-through figure. The charter therefore links each hypothesis to its target event, grain, exposure denominator and decision owner.
¶ 03
Scope should remain deliberately narrow. The first pilot should not simultaneously change garment specification, price, promotion, delivery promise and channel unless the decision concerns that whole bundle. A bundle test can be legitimate, but its result applies to the bundle and does not identify which component produced the response. Where the team needs component-level learning, it should hold other conditions stable or use a design whose cells can support the intended comparison. The write-up must not claim more resolution than the design contains.
¶ 04
A decision threshold also needs an evidence-quality threshold. A strong purchase result can still be unusable if assignment failed, listing time differed between cells, stock records are incomplete, or returns have not had time to arrive. The charter distinguishes the commercial readout from the permission to learn from it. This avoids rewarding a cell for favourable conditions that the future launch cannot reproduce. It also makes a failed measurement exercise visible before its numbers enter the forecast.
¶ 05
The pre-launch register records the hypothesis, commitment at risk, offer cells, assignment unit, eligible population, primary outcome, guardrails, availability rules, observation window, analysis owner and gate actions. Any amendment receives a date and reason. Changes made after outcomes are visible remain in the record, but they are labelled exploratory rather than silently becoming the original plan. This is the practical boundary between a forecasting experiment and launch communication.
¶ 06
Note
6.23 min · 9 paragraphs · 2 exhibits
Randomisation separates offer effects from self-selected demand
Exhibit 279 paragraphs step it
Exhibit 27RegisterillustrativeA common denominator keeps the pilot comparison honest
Measure
Valid offer exposures
Cell A
500
Cell B
500
Readout role
Common denominator
01 / 05
Read-out
Open the full register (5 rows)
Valid offer exposures
500
500
Common denominator
Gross orders
60
45
Initial response
Cancellations
6
5
Outcome adjustment
Net purchases
54
40
Gross less cancellations
Net purchase rate
10.8
8
Percent of exposures
Method and source
Illustrative worked example. A net purchases = 60 - 6 = 54; B net purchases = 45 - 5 = 40. Rates = 54 / 500 = 10.8 percent and 40 / 500 = 8.0 percent. Difference = 2.8 percentage points. No confidence interval, causal effect estimate or scale forecast is claimed.
Unit: illustrative counts and percent of valid exposures. Status: illustrative. Sources: S32. Illustrative figures are teaching examples, not measured factory results.
Customers who choose a channel, location or launch date can differ before they see the offer. A pilot that lets the most enthusiastic buyers select the test cell can confuse that enthusiasm with the effect of the garment or price. Controlled experiments address this problem by assigning comparable units to different conditions, supporting a causal interpretation when the assignment and measurement are implemented correctly.[S32] For an apparel launch, the unit might be an eligible customer, a selling location, a distributor account or a defined time cell, depending on how the offer can actually be delivered.
The assignment unit must prevent the test from leaking across cells. A customer who can see two prices on consecutive visits may not provide independent responses to either. A shop whose staff promote the test garment more strongly is no longer delivering only the assigned product condition. Locations that share customers, stock or sales staff can contaminate one another. The operational design should choose a unit that keeps the assigned experience coherent, then record exceptions rather than assume that the assignment code solved every practical problem.
Randomisation does not replace comparability checks. Before outcomes are read, the team examines whether cells received the intended numbers and types of eligible units, whether listing and availability conditions were equivalent, and whether any external event affected one cell alone. These checks diagnose implementation; they do not justify deleting inconvenient units after seeing purchases. The primary readout follows assigned cells, while an exposed-only analysis can be secondary and clearly labelled. This keeps refusal, non-collection and missed exposure from quietly reshaping the comparison.
The measurement sequence in Exhibit 26 makes those dependencies visible. It treats availability and returns as part of the experiment rather than fields to be appended after a winner has been declared. The flow is a GFT method design derived from controlled-experiment and censored-demand principles, not evidence that a particular launch used it.[S32][S06][S07]
The primary outcome should be observable and consequential. Gross orders, fulfilled purchases, net purchases after cancellations, returns and repeat orders answer different questions. A pilot about initial purchase demand might use fulfilled net purchases per valid offer exposure, while retaining gross orders and returns beside it. A pilot about institutional interest should not call a non-binding expression of interest a sale. The denominator likewise needs care: visitors, eligible customers, account contacts and offer impressions represent different exposure opportunities.
Exhibit 27 preserves that worked calculation in a compact readout. The important comparison is not merely A against B. It is the movement from a common exposure denominator to gross orders, cancellations and net purchases, with the same definitions in both cells. Counts and rates remain illustrative and do not carry a significance claim.
Small cells require restraint even when the arithmetic is exact. A handful of large accounts, repeated orders from one buyer, or an event affecting one location can dominate the result. The readout should show the number of independent buying units beside transactions and identify concentration without turning every repeated garment into independent evidence. Where the design has too few assignment units for a stable comparison, the appropriate conclusion is inconclusive. The team can retain the observations as descriptive history and run another test, rather than promote an exploratory difference into a causal law.
Where randomisation is impossible, the charter records how units entered each condition. A distributor may have selected its own assortment, a location may have been chosen because of expected demand, or early buyers may have received a different price from later buyers. Those results can still inform a forecast, but the language changes. The report describes observed purchases under selected conditions, examines plausible selection mechanisms and avoids attributing the whole difference to the offer. A transparent non-randomised pilot is more useful than a supposedly controlled one whose assignment failed.
6.32 min · 6 paragraphs · 1 exhibit
Limited releases identify the ramp while containing exposure
Step 1 of 6Freeze charter and cells
Exhibit 28 · illustrativeA twelve-week pilot can sequence learning without rushing scale
Freeze charter and cells
Freeze charter and cells: from 0 to 1 (illustrative weeks from charter approval) · illustrative
A limited release allows the venture to buy information with a bounded production commitment.
Instrument and A/A checks
Instrument and A/A checks: from 1 to 2 (illustrative weeks from charter approval) · illustrative
Limited release
Limited release: from 2 to 5 (illustrative weeks from charter approval) · illustrative
Availability audit
Availability audit: from 2 to 6 (illustrative weeks from charter approval) · illustrative
Replenishment and repeat watch
Replenishment and repeat watch: from 5 to 10 (illustrative weeks from charter approval) · illustrative
Gate review
Gate review: from 10 to 12 (illustrative weeks from charter approval) · illustrative
Method and source · Exhibit 28
Illustrative planning calendar. Tasks and durations are assumed to show sequencing and overlap. They are not observed milestones, a critical-path calculation or a required duration for an apparel launch.
Unit: illustrative weeks from charter approval. Status: illustrative. Sources: S32, S06. Illustrative figures are teaching examples, not measured factory results.
6.3 · the bodyExhibit 28 above; its figures are illustrative where marked.
A limited release allows the venture to buy information with a bounded production commitment. The operative word is bounded. The lot, selling window, eligible channels and replenishment conditions are fixed before the result is known, so an enthusiastic first response cannot automatically trigger an unplanned expansion inside the same evidence window. This separates learning from the pressure to make the launch look successful. It also preserves a meaningful distinction between the initial release and the later ramp that the pilot is meant to inform.
The release should be large enough to exercise the real buying and fulfilment process, not merely to display samples. Customers need a genuine offer, a defined price, a delivery or collection mechanism and a route for returns. Production needs to execute the intended specification rather than a handmade substitute whose quality and cost cannot be repeated. At the same time, the pilot need not cover every size, region or channel. Uncovered conditions remain explicit transfer questions for a later release.
¶ 02
Release timing is part of the treatment. A lot released in several planned tranches can test whether early purchases persist after initial novelty, while a single lot can address first-window response when production flexibility is limited. The charter should not switch between those designs because early sales are disappointing. If commercial safety or customer harm requires interruption, the human stop rule takes precedence and the analysis records the deviation. The statistical plan never obliges the team to continue an unsafe or misleading offer.
¶ 03
The illustrative calendar in Exhibit 28 sequences one possible first release over twelve weeks. It allows instrument checks before sale, availability audit during sale, and a later window for replenishment and repeat observation. The durations are planning assumptions, not a universal minimum, a GFT client commitment or evidence that twelve weeks can resolve every apparel question.
The calendar protects different learning clocks. Initial conversion can be read before repeat purchase, but a gate that requires repeat evidence cannot be cleared early merely because the first lot sold through. Returns and cancellations also need their own maturity rules. A readout taken before customers have had a realistic opportunity to return or repeat should be labelled provisional. Later observations arrive as new evidence versions, not as silent edits to the first conclusion.
¶ 05
A small lot creates a further risk: scarcity can become the story. Selling out may indicate strong demand, insufficient stock, delayed replenishment or some combination of all three. The release plan therefore separates the demand test from the replenishment test. During the demand window, availability should be as comparable as the design requires. During a later replenishment window, the team can test response time and continuing purchases without pretending that the two phases share identical supply conditions.
¶ 06
6.43 min · 7 paragraphs · read with a related exhibit
Availability must be designed into the pilot
Low sales are interpretable only when the customer could have bought. Retail research shows that demand estimates based on sales alone can be negatively biased when inventory constrains observation, and that stockouts can shift purchases towards substitutes.[S06] A pilot built to create purchase history must therefore measure the opportunity to purchase with the same care as the purchase itself. Otherwise, the model learns the venture's stock and listing failures as if they were customer preferences.
Availability begins at the actionable product and location grain. The record identifies when each variant was listed, when it could be fulfilled, when available inventory reached zero, and when listing or fulfilment resumed. A style can appear available while the requested size is not. An online listing can remain visible while delivery is impossible. A distributor can report a shipment while the garment is not yet offered to the final customer. The pilot dictionary must choose the availability event that matches the demand target and preserve the others as context.
¶ 02
Stockout timestamps need verification, not just a system flag. Inventory records can be late or inaccurate, so the protocol includes physical or transactional checks proportionate to the pilot's size. Staff record suspected gaps, customer requests that could not be fulfilled, manual overrides and unexpected delisting. These observations do not automatically become demand quantities. They tell the analyst where the sales series is censored and where a zero cannot support the interpretation of no interest. Stockout-annotated retail research makes that distinction central to latent-demand recovery, although its fresh-retail setting and hourly granularity should not be imported uncritically into apparel.[S07]
Substitutes belong in the same observation window. When the requested variant is unavailable, a customer may buy another size, colour or product, delay the purchase or leave. The pilot should retain which alternatives were visible and available, together with their outcomes. It should not assume that all unmet demand transferred, vanished or waited. Those are different behaviours with different consequences for the focal product and the substitute. The analysis can test a correction only if the collected fields support its assumptions.
¶ 04
Returns and cancellations complete the availability record. A fulfilled order that is later returned is not the same evidence as a retained purchase, and a cancelled order may never have entered physical demand fulfilment. The readout keeps gross order, fulfilment, cancellation, return and net purchase events separate, with their dates and reasons. It also marks outcomes still immature at the information cut-off. Deleting the original transaction after a return would destroy the sequence needed to understand both product response and data quality.
¶ 05
The availability rule can require a minimum valid exposure condition without inventing a universal threshold. The charter sets that condition for each channel and cell before outcomes are read. Intervals failing it are excluded from the primary availability-valid analysis, retained in a separate contaminated record and included in sensitivity reporting. Exclusion is not permission to remove weak sales selectively. It applies to defined measurement failures under a frozen rule, with counts and consequences visible.
¶ 06
A further distinction separates intended scarcity from operational failure. A deliberately small lot can constrain exposure by design, but unplanned early exhaustion can still prevent the intended comparison. Likewise, a promotion that drives an unexpected stockout changes both demand and observation. The pilot report should describe the mechanism rather than label every sold-out release a success. Where censoring is material, the forecast may retain a broad range, present alternative correction scenarios or await another release. More arithmetic cannot replace missing demand observations.
¶ 07
Interlude · the cloth half-woven
4 exhibits carry this chapter's argument.
Woven: sections 6.0 to 6.4Still to weave: 6.5 to 6.7
Key figures from chapter 01 · Exhibit 04Key figuresillustrativeFour questions discipline every cold-start forecast
CommitmentGate 1: DecisionWhich lot, purchase or allocation changes if the range shifts?
Latent demandGate 2: TargetHow are stockouts, returns and channel timing separated?
Prior weightsGate 3: ProvenanceWhich analogues, surveys and judgements form the range?
Stop ruleGate 4: ReversalWhat pilot or replay outcome forces a fallback to baseline?
01 / 04
Read-out
Method and source
Illustrative governance tiles summarising the four questions required before a cold-start forecast may inform a production commitment. Tiles state governance gates rather than empirical performance metrics.
6.52 min · 7 paragraphs · read with a related exhibit
Channel and geography readouts remain separate
A pooled launch total can hide the only result that matters for the next commitment. Direct retail, online sales, distributor accounts and institutional buying create different purchase events, exposure denominators, inventory responsibilities and return paths. Multichannel fashion research evaluates physical and non-store channels separately and calls for region- and channel-specific testing of candidate predictors.[S04] The pilot should preserve that separation even when the venture ultimately wants one production forecast.
Channel separation starts with the event boundary. A distributor's opening shipment is not consumer sell-through, and an institutional award is not the same as wearer uptake. Direct customer purchases can be counted against valid offer exposures, while a distributor readout may focus on qualified accounts, orders, subsequent replenishment and documented downstream availability. These measures can coexist without being added as if they were identical demand events. The forecast contract assigns each one a role and an aggregation rule.
¶ 02
Geography adds a second boundary. Selling locations can differ in customer mix, calendar, access, substitutes, delivery reliability and staff execution. The pilot register records which of those conditions were observed and which are merely suspected. It does not invent a regional adjustment factor from a successful location. A result from one selling context becomes evidence for another only through an explicit transfer assumption, normally accompanied by wider uncertainty or another test.
¶ 03
The readout should be hierarchical rather than collapsed. Product variants roll into styles, selling locations roll into markets, and channels roll into the venture total only after their component evidence has been inspected. The M5 retail evaluation was explicitly organised around hierarchical unit sales and required forecasts at the lowest cross-sectional level to be summed to higher levels, which provides a useful evaluation principle rather than an apparel effect size.[S11] For the pilot, component and total views should reconcile while retaining their different sample support.
A small pilot will have sparse cells. The answer is not to pool them until every apparent problem disappears. Each readout carries its independent assignment units, valid exposures, availability coverage and outcome maturity. Cells below the charter's support condition are labelled preliminary or inconclusive. Partial pooling may later borrow strength across comparable groups, but that modelling choice belongs after the raw channel and geography evidence has been preserved. It must not erase a failed channel by hiding it inside a favourable total.
¶ 05
The practical output is a channel and geography transfer register. For every cell it records the offer, event definition, assignment unit, selling window, availability condition, net outcome, concentration, returns maturity and permitted next use. One cell may justify another direct-sale lot but no distributor expansion. Another may justify repeating the same test in a second location before any wider commitment. The pilot's value lies in these specific permissions and prohibitions, not in one launch-wide approval.
¶ 06
This separation also protects planning conversations. Production may need a total range, but commercial owners need to know which channels and places created it. Reviewing both levels reveals whether the total depends on one account, one city, one promotion or one size curve. The team can then direct the next limited release towards the uncertainty that blocks expansion rather than repeat the easiest selling condition and call the accumulated transactions new evidence.
¶ 07
6.63 min · 9 paragraphs · 1 exhibit
Sequential updates preserve the prior and every revision
Step 1 of 4Prior
Exhibit 29 · illustrativeForecast revisions should show both learning and residual doubt
Prior
Prior: low 40, central 80, high 150 (illustrative indexed demand per launch window) · illustrative
The pilot should update a forecast, not overwrite its history.
Post-choice
Post-choice: low 50, central 85, high 140 (illustrative indexed demand per launch window) · illustrative
Post-pilot
Post-pilot: low 55, central 105, high 180 (illustrative indexed demand per launch window) · illustrative
Post-repeat
Post-repeat: low 70, central 100, high 150 (illustrative indexed demand per launch window) · illustrative
Method and source · Exhibit 29
Illustrative bounds and midpoints. Prior = 40, 80, 150; post-choice = 50, 85, 140; post-pilot = 55, 105, 180 after an assumed availability and transfer exception; post-repeat = 70, 100, 150 after assumed valid repeat evidence. These are constructed ranges, not statistical intervals, observed demand or guaranteed uncertainty reductions.
Unit: illustrative indexed demand per launch window. Status: illustrative. Sources: S28, S06. Illustrative figures are teaching examples, not measured factory results.
6.6 · the bodyExhibit 29 above; its figures are illustrative where marked.
The pilot should update a forecast, not overwrite its history. Before launch, the venture holds a versioned prior built from analogues, market evidence, choice research and structured judgement. Each pilot readout then creates a new forecast version with an information cut-off, the evidence admitted, the update rule and the resulting range. Hierarchical Bayesian modelling offers one candidate for combining shared and local evidence, but its supporting demand application concerns copper rather than apparel, so the method and its assumptions still require launch-level validation.[S28]
An update register makes the transition inspectable. It stores the prior identifier, analogue corpus and research versions, pilot cell definitions, valid exposure counts, availability treatment, cancellations and returns cut-off, support by channel, model or weighting rule, posterior or revised quantiles, reviewer and approval status. A planner can then reconstruct why the central estimate moved and which new observations caused the change. Without that chain, each weekly spreadsheet becomes a new story with no way to test whether learning actually occurred.
¶ 02
Evidence weight should not equal transaction count. Repeated purchases from one account, several sizes bought in one order, and transactions sharing a promotion can carry dependent information. Conversely, a smaller number of well-assigned, availability-valid purchases may answer the named uncertainty more directly. The update rule therefore considers assignment quality, independence, censoring, offer stability and channel relevance. These are not decorative metadata fields. They determine whether the new history deserves to displace the prior.
¶ 03
Double counting is a related danger. If the same buyers contributed to choice research and the pilot, or if an expert revised the prior after seeing early sales, those inputs are no longer independent confirmations. The evidence register records such dependencies. A combined forecast should not present them as separate votes for the same outcome. Where dependence is unclear, the review can compare a conservative update with the proposed one and retain additional uncertainty rather than claim a precise solution.
¶ 04
More data can legitimately widen the forecast. The pilot may reveal that the target channel behaves unlike the analogue channel, that the size curve is unstable, or that availability failures are more consequential than expected. A revision that responds by widening the range is not a failure of learning. It is evidence that the previous confidence was not earned. The update protocol should permit restoration of an earlier prior when new records fail validation, rather than forcing every version to move forward with narrower bounds.
¶ 05
Exhibit 29 illustrates that non-monotonic path. The constructed post-pilot range widens when conditions depart from plan, then narrows only after repeat evidence becomes usable. Values are indexed illustrations, not garment sales, fitted posterior quantiles or calibrated coverage statements. Their purpose is to show the reporting form: prior, evidence stage and residual doubt remain visible together.
The midpoint in that illustration does not tell the whole story. A revised central value can rise while the upper and lower bounds both move outwards, signalling disagreement between the first purchases and the transfer assumptions. Reviewers should therefore inspect the full distribution and the evidence exceptions, not just the direction of the point forecast. They should also compare the update with the frozen pre-launch rule. If the rule is changed after seeing the result, the new version is a new modelling proposal that requires separate evaluation.
¶ 07
Sequential updating needs disciplined timing. The charter states when initial purchase, return and repeat windows close, and which cut-off applies to each forecast release. Late-arriving events enter the next version with their original event times and their arrival times. This preserves both the commercial sequence and the information sequence. A later backtest can then reproduce what the venture actually knew at each decision instead of using a cleaned final history that was unavailable at the time.
¶ 08
Finally, the update pack should be readable by commercial and production owners without reference to model code. It states what was expected, what was observed, which observations were excluded and why, how the range changed, and which commitment the revision supports. Technical reviewers retain the detailed model diagnostics. The planning version keeps the decision chain visible so that a forecast revision can be challenged before it becomes a material or capacity commitment.
¶ 09
6.73 min · 8 paragraphs · read with a related exhibit
Stop, adapt and scale rules are fixed before results arrive
Related · chapter 148 paragraphs step it
Related exhibit 64 · chapter 14FlowillustrativeEvery release requires evidence and a reversal path
01 / 08
Read-out
Method and source
Illustrative GFT process design, informed by controlled experiments, decision value and monitoring guidance. Nodes are requirements, not measured readiness scores. Failure at any gate holds the current approved prior or baseline; the recorded issue must be repaired before re-entry.
Unit: illustrative release sequence. Status: illustrative. Sources: S32, S33, S40. Illustrative figures are teaching examples, not measured factory results.
A pilot needs more than a success threshold. It needs a complete set of authorised outcomes before anyone knows whether the preferred offer is winning. The charter should distinguish stop, adapt, repeat and scale actions, together with the evidence-quality and commercial conditions attached to each. Human owners set the metrics and thresholds in the context of the actual decision rather than importing a universal benchmark. NIST's risk framework specifically places human judgement behind the choice of trustworthiness metrics and their precise thresholds.[S36]
Stop rules protect customers, operations and evidence. A misleading offer, an unresolved fulfilment fault, compromised assignment or a material safety concern can justify immediate interruption even when purchases are strong. The stop owner does not need to wait for statistical maturity. The event and its timing are recorded so that later analysis can distinguish commercial response from the conditions under which the test ended. Continuing a faulty experiment merely to complete a sample would subordinate the operation to the measurement plan.
Adapt rules allow a defined correction without turning every disappointing result into a new hypothesis. An adaptation may repair an availability failure, correct a listing error, or repeat a cell whose assignment mechanism failed. It should not quietly change the price, audience or product after weak sales and then count the later response as the original test. Material changes create a new pilot version and a fresh evidence window, while the original outcome remains in the register.
Repeat rules address inconclusive evidence. The trigger may be insufficient independent buying units, excessive censoring, immature returns, conflicting channel readouts or a difference too unstable to support the intended commitment. Repetition should have a purpose: more comparable locations, cleaner availability, a longer repeat-purchase window or a revised lot condition. Simply repeating the easiest cell until it produces an encouraging number would make the gate a selection device rather than a learning rule.
Scale rules join statistical evidence to operating consequences. The review asks whether the observed offer can be produced and fulfilled repeatedly, whether availability remained valid, whether returns and repeat behaviour are sufficiently mature, and whether the revised forecast changes the plan for the better. Where a forecast candidate is also under review, the gate retains its pre-specified accuracy and calibration conditions beside the operational and commercial tests. Practical machine-learning guidance warns that an improvement in model loss is not acceptable if system performance deteriorates.[S33] For a pilot, the equivalent rule is that an attractive purchase metric cannot compensate for a production or fulfilment condition that would fail at the proposed scale.
Scale itself should remain staged. Clearing the pilot gate authorises only the commitment named in the charter, such as another lot, an additional comparable location or a controlled channel extension. It does not automatically authorise every product, geography and production horizon. The next release carries its own monitoring and reversal conditions. If the transfer assumptions weaken, the venture can return to the approved prior, repeat a narrower test or stop the expansion without rewriting the meaning of the earlier evidence.
A complete pilot handover therefore contains the charter, assignment and exposure logs, availability audit, cell readouts, update versions, gate decision and next-release conditions. It distinguishes observed purchase evidence from illustrative scenarios and from the owner's judgement. It also states what the pilot did not test. Those limits keep the first owned history useful after the launch team changes and prevent a small, context-specific success from becoming an unsupported general rule.
Check the record
The chapter keeps its sources within reach.
Every knot is a source this chapter cites; select one to read its record, or filter the register by grade. Publication dates and inherited-source limits stay explicit. Read the paper's full evidence summary.
8 sources · 22 citation marks on this pageMost cited: S32, 5 times
Grade AGrade BGrade CKnot size: citations on this page
All 8 records shown
c06-S04 · Grade B
MDPI, Processes
Demand Forecasting for Multichannel Fashion Retailers by Integrating Clustering and Machine Learning Algorithms
Therefore, it is recommended that future researchers consider other external predictor variables such as holidays (school and business), commercially relevant events (Easter, Christmas), extraordinary events (e.g., a pandemic), and economic conditions when predicting demand for a more specific region or different sales channels.
2021-09-03. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
FreshRetailNet-LT: A Stockout-Annotated Censored Demand Dataset for Latent Demand Recovery and Forecasting in Fresh Retail
Consequently, the sales-equals-demand assumption collapses in the event of product stockouts, as observed sales volumes systematically under-represent the latent demand.
2026. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
International Journal of Forecasting, Elsevier; abstract through RePEc IDEAS
M5 accuracy competition: Results, findings, and conclusions
The main objective in the M5 “Accuracy” competition was to accurately predict 42,840 time series representing the hierarchical unit sales for the largest retail company in the world by revenue, Walmart.
2022. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
A hierarchical Bayesian regression model that reduces uncertainty in material demand predictions
Bayesian regression models permit the use of smaller datasets through hierarchical modeling (Vehtari et al., 2017) which allows sharing of information across heterogeneous sites.
2022-10-17. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Springer Nature, Data Mining and Knowledge Discovery
Controlled experiments on the web: survey and practical guide
Controlled experiments embody the best scientific design for establishing a causal relationship between changes and their influence on user-observable behavior.
2008-07-30. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Human judgment should be employed when deciding on the specific metrics related to AI trustworthiness characteristics and the precise threshold values for those metrics.
2023. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Move the shuttle along the rack to preview any chapter: its sections, exhibits and reading time. Draft web edition. This research is not investment, legal or tax advice; illustrative models are not measured factory results.
The paper's rack · 14 boltsBolt length: reading time
1.1
1.3
1.5
1.7
1.9
1.11
1.13
01Part I: Define the decision and the prior
No history is a forecasting condition, not a forecasting ban
A greenfield maker can forecast responsibly when it states what evidence transfers, what remains uncertain and what would change the plan.
1.1Greenfield apparel begins with transfer uncertainty rather than an empty evidence set
1.3GFT positions forecasting as a phased decision capability
1.5A new firm's cold start differs from a new SKU's cold start
1.7The forecast unit follows the commitment it must inform