Research libraryForecasting From a Cold Start
Chapter 12 of 14 · Part IV · Draft web edition
The only fair test is one that denies the model every piece of information it would not have had at the decision moment.
The cloth is woven from this chapter's 5 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.
Read-out
Illustrative chronology derived from unseen-item and expanding-window principles. The target launch is excluded from every fitting stage in this replay. Positions are planning assumptions, not observed dates or recommended intervals. Actual origins must follow the venture's decision and data-arrival clocks.
Unit: illustrative weeks relative to target launch. Status: illustrative. Sources: S02, S13, S1201. Illustrative figures are teaching examples, not measured factory results.
A cold-start backtest is a rehearsal of ignorance. It must remove the target launch from training, reconstruct the evidence that existed at each decision origin and score only outcomes that were still unknown at that origin. This rule applies to product attributes, sales, stock records, external series, analogue selection, demand correction, preprocessing, calibration and forecast combination. A later row split is not enough when the claim concerns an unseen product. The valid test holds out whole launches, repeats several launch origins and compares every candidate with the same frozen baselines. Release follows only when accuracy, distribution quality, planning consequences and operating fitness survive that rehearsal with adequate launch support.
This design measures the capability the venture actually intends to use. Before launch, a model may know approved product attributes and earlier evidence, but no target sales. After the first selling interval, it may know only the transactions and availability records that had arrived by then. An early-repeat forecast receives another bounded information set. Each condition needs its own replay and its own artefact versions. In Cameroon, where the inherited sector baseline records about 95 percent of cotton fibre exported raw against about 5 percent processed locally and national policy restates a 50 percent local-processing target by 2030, policy ambition explains why new apparel capacity matters, not how many garments a launch will sell.[R06][R07] The resulting backtest score is not a timeless property of an algorithm or investment, legal or tax advice. It is evidence about a specified forecasting service, evaluated on a stated launch population, under stated information constraints.
12.1
Related · chapter 017 paragraphs step it
Read-out
Illustrative sequence of evidence stages for a cold-start apparel launch, drawing on hierarchical pooling, preference calibration and controlled pilot design. Node order shows information maturity, not a fixed calendar duration.
Unit: launch evidence stages. Status: illustrative. Sources: S28, S31, S32. Illustrative figures are teaching examples, not measured factory results.
Read this exhibit in its chapter ↗A forecast origin is the instant at which the information set closes and the forecast becomes fixed. Everything used to produce that forecast must have been available by then. Everything scored must belong to its future. Expanding-window evaluation formalises this direction by training on successive historical supersets and testing later observations. Scikit-learn's time-series documentation contrasts these successive training sets with ordinary cross-validation and warns that time-dependent samples need a time-aware scheme.[S13] Empirical comparison of time-series performance estimation methods likewise indicates that when non-stationary variation is present in real-world series, out-of-sample methods that preserve the temporal order of observations produce the most accurate loss estimates.[S1201]
The cold-start harness applies that principle at two levels. Across the reference corpus, later origins may add completed launches that had become available. Within each target launch, later origins may add only the target observations that had arrived by the new cut-off. The two expansions are recorded separately. Otherwise, a backtest can appear chronological within one sales series while silently using a reference launch that did not yet exist when the original decision was made.
Several origins are necessary because one cut-off can be unusually kind to a method. A pre-launch origin tests transfer from attributes and analogues. An origin immediately after the first selling interval tests the first update from owned evidence. A later origin tests whether the model uses a short ramp without allowing that ramp to dominate its earlier prior. These are different forecasting tasks, not repeated measurements of one task. Their scores remain separate before any aggregate is considered.
The origin register therefore stores the decision deadline, data cut-off, forecast horizon, eligible reference launches, target-history length and expected publication lag for every input. Known future fields receive special treatment. A planned price or approved campaign date may be legitimate when genuinely known at the origin. The realised sales response to that campaign is not. A revised market statistic is represented by the vintage available at the cut-off, not by the cleanest version downloaded after the launch.
Exhibit 54 gives an illustrative sequence for one wholly held-out launch. Its numeric positions are weeks relative to launch, not a recommended planning calendar. The important feature is the repeated closure and freezing of evidence. The target launch never enters fitting for that replay, while each later forecast is rebuilt only from the information then permitted.
The sequence also prevents retrospective model shopping. A team that tries several origins and reports only the best has created another selection bias. The origin set belongs in the evaluation protocol before candidates are compared. Any additional origin is appended under a documented rule, with all prior results retained. Where an origin lacks the evidence required for a legitimate operational forecast, the result is an ineligible case with a reason, not a quiet deletion.
12.2
An unseen-item forecast must be tested on an item whose identity and outcomes were absent from training. Attribute-based fashion research builds generalised models from other products' attributes and merchandising factors, so its relevant generalisation question is whether those relationships transfer to a new item.[S02] A random row split cannot answer it. Later rows for a product already represented in training test temporal continuation for a known item, which is a different and usually easier deployment condition.
Product exclusion must cover every route through which target demand could return. Obvious routes include sales rows, returns and target-derived aggregates. Less obvious routes include normalisation statistics, selected features, demand corrections, analogue weights and hyperparameter choices. Even a target identifier can allow a flexible model to memorise historical behaviour rather than learn transferable attribute relationships. The split therefore occurs before any target-dependent transformation, not after a convenient feature table has been built for the whole corpus.
Visual transfer needs an additional exclusion boundary. In the opened cold-start fashion architecture, new products are matched to clusters and receive corrected demand signals from similar historical products.[S05] In a faithful replay, the target product cannot influence those corrected signals, cluster centroids, neighbour rankings or the choice of cluster count using its own outcomes. Its image and approved static attributes may enter at inference. Its demand history may not shape the reference structure that supposedly predicts it.
There is also a difference between leaving out a product and leaving out an independent launch group. Colour and size variants can share photography, a campaign, inventory and customers. Several launches can share one institutional order or channel event. Treating all related variants as independent holdouts can make transfer look stronger than the deployment evidence warrants. The group definition should follow these dependencies and remain fixed before modelling. Where independence is doubtful, the conservative choice is to hold out the wider group and report the reduced support.
Chronology still applies inside the held-out group. A product launched after the forecast origin cannot become an analogue merely because its full history exists in the evaluation file. The eligible training pool is the intersection of two rules: the product group is permitted, and its information was available by the origin. This joint rule blocks both entity leakage and temporal leakage without reducing the test to a single convenient split.
Finally, product holdouts need a role distinction. Validation launches may guide model and hyperparameter selection under a frozen procedure. Final test launches remain untouched until the selected pipeline is run once for the release readout. Reusing the final products to choose the winner turns them into validation evidence and removes the independent check. If the corpus is too small for separate roles, the honest result is preliminary evidence with wider uncertainty, not a claim of confirmed generalisation.
12.3
A product holdout becomes a launch replay only when the whole launch history is removed together. The test then reproduces the pre-launch prior, the first-sale update and the early-repeat update in sequence. It does not select a convenient later window from an otherwise known product. This extends unseen-product evaluation and chronological splitting into a single deployment rehearsal.[S02][S05][S13]
The evaluation corpus is frozen before candidates are compared. Its register identifies every launch, its source and access right, extraction time, product and channel attributes, complete observation window, availability rule, correction version and exclusion reason. Failed, withdrawn and slow launches remain visible when they belong to the deployment population. Removing them because their outcomes are awkward would change the population after seeing the answer and would overstate the quality of launch transfer.
Pilot structure is part of that provenance. Where a pilot assigned customers, locations or time cells to different offers, those cells remain intact in the replay. The evaluation cannot move a favourable treatment cell into training while leaving its control counterpart in the test set and still claim to assess the original experiment. Controlled-experiment guidance supports randomised designs for causal interpretation; preserving assignment is necessary to retain that meaning in a later forecast evaluation.[S32]
The launch register also records what was decided rather than only what was sold. Release timing, initial lot, price, promotion, listing intervals, stockouts and replenishment can all alter the observed ramp. These fields let the reviewer distinguish a forecasting failure from a selling condition that the model could not have known. They also prevent a later, tidier reconstruction of the launch from replacing the conditions under which evidence was actually earned.
Support is reported by independent launch group, not only by rows. Many weekly observations from one launch provide limited independent proof that a method transfers to another launch. Related variants can reduce effective support further. Every product, channel and horizon cell therefore carries its launch count, observation count, dependency grouping and eligibility count for each metric. No universal minimum is asserted here. The release owner sets a defensible requirement before seeing candidate results, using the decision risk and the structure of the available corpus.
Sparse cells remain preliminary. They are not pooled into a broader cell merely because the broader result looks more settled, and they are not assigned a precise confidence statement that their support cannot sustain. The readout can show their observed errors with a clear limitation, then specify what additional launches would change the state. If independent support is too weak for the intended release, the approved outcome is to retain the prior or baseline and collect more evidence.
12.4
Freeze origin and as-of joins
Freeze origin and as-of joins. Starting point. Leads to Exclude target launch group. · derived
Leakage is not limited to placing future sales in a feature column.
Exclude target launch group
Exclude target launch group. From Freeze origin and as-of joins. Leads to Fit target rules on training only. · derived
Fit target rules on training only
Fit target rules on training only. From Exclude target launch group. Leads to Fit imputation scaling and selection. · derived
Build eligible analogue artefacts
Build eligible analogue artefacts. From Fit imputation scaling and selection. Leads to Fit and freeze models and calibration. · derived
Fit and freeze models and calibration
Fit and freeze models and calibration. From Build eligible analogue artefacts. Leads to Infer from permitted target evidence. · derived
Infer from permitted target evidence
Infer from permitted target evidence. From Fit and freeze models and calibration. Leads to Score against later actuals. · derived
Score against later actuals
Score against later actuals. From Infer from permitted target evidence. End point. · derived
The method behind it: GFT note.
Derived GFT audit sequence from chronological evaluation and leakage taxonomy. The flow covers target derivation, preprocessing, analogue construction, calibration and combination. It is a control design, not evidence of an implemented pipeline or measured performance.
Unit: validation control sequence. Status: derived. Sources: S13, S37.
12.4 · the bodyExhibit 55 above; its figures are illustrative where marked.
Leakage is not limited to placing future sales in a feature column. Kapoor and Narayanan define it as a spurious relationship between predictors and target created by collection, sampling or preprocessing, and their taxonomy includes contaminated train-test separation, illegitimate features and mismatched evaluation distributions.[S37] A cold-start audit must therefore inspect the entire evidence path, from event capture to the scored forecast.
The first check concerns the target itself. If observed sales are corrected for stockouts, the correction must use only information permitted at the relevant stage. A correction fitted on the full launch can encode later demand, substitutions or availability into the training target. The same applies to return allocation and cancellation rules. Target derivation is versioned and fitted inside the training fold, then applied under a frozen rule. Where no defensible correction is available, the affected observations are marked censored or excluded under the protocol rather than filled with invented demand.
The second check concerns every learnt artefact around the model. Imputation values, scaling parameters, feature selection, outlier rules, visual embeddings adapted to the reference catalogue, analogue clusters, combination weights and calibration maps are all fitted on training evidence only. A pipeline is not clean merely because the final estimator receives no target rows. Earlier stages can already have transferred the answer.
The third check concerns time. Each feature carries an event time, an available time and, where relevant, a revision or ingestion time. Event time alone is insufficient for externally published data and late-arriving operational records. A trade value for an earlier month can still be unavailable at the forecast origin if publication occurred later. An updated attribute can replace the description customers actually saw. The as-of join selects the version whose available time precedes the cut-off, not the record whose nominal period merely looks earlier.
Exhibit 55 places these controls in one derived GFT sequence. It is a proposed validation architecture, not an implemented client system. The central rule is that fitting finishes before the held-out launch is transformed or scored. Later actuals enter only at the final scoring node.
Two practical probes make the audit testable. First, a deliberate canary feature stamped after the origin should cause the information gate to reject the fold. Second, rerunning a frozen fold should reproduce its forecasts and transformations from stored artefacts rather than from the current database. These probes test the harness, not forecast skill. They are still necessary because an elegant score produced by an unverifiable pipeline cannot support release.
Interlude · the cloth half-woven
Woven: sections 12.0 to 12.4Still to weave: 12.5 to 12.8
Read-out
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.
Unit: mandatory pre-launch contract gates. Status: illustrative. Sources: S08, S14, S33, S36. Illustrative figures are teaching examples, not measured factory results.
Read this exhibit in its chapter ↗12.5
Launch forecasts cover products with different volumes and error units. MASE addresses comparability by scaling absolute forecast error with a naive benchmark, and Hyndman and Koehler propose it for accuracy comparisons across multiple time series.[S12] In launch replay, the scaling denominator must be computed only from information available at the forecast origin. It must also retain the same seasonal and availability definitions as the forecast contract.
A pre-launch target often has no valid naive denominator. A very short history can make it unstable, and an all-zero eligible history can make it undefined. The harness must not replace an absent denominator with one, drop the inconvenient launches or calculate the denominator using future target sales. MASE is reported only for eligible cases, with the eligible-launch count and denominator diagnostics beside the result.
Where MASE is unavailable, the readout retains MAE and signed bias in demand units. It also reports aggregate relative absolute error against the frozen analogue baseline: sum candidate absolute error divided by sum baseline absolute error over the same eligible forecast cases. This is GFT arithmetic on paired outputs, not a replacement definition of MASE. A zero baseline denominator is disclosed and prevents the ratio. Unscaled errors remain visible so that a relative gain cannot hide operationally large misses.
Exhibit 56 supplies an illustrative paired readout for three pre-launch forecast cases. Actuals and forecasts are constructed garments in one forecast period, not observations or a Cameroon apparel result. Signed error is actual minus forecast, so a negative value denotes overforecasting. The example shows why absolute error, relative error and bias must coexist.
Read-out
| A | 120 | 100 | 110 | 20 | 10 |
|---|---|---|---|---|---|
| B | 80 | 110 | 95 | -30 | -15 |
| C | 60 | 50 | 75 | 10 | -15 |
Illustrative constructed inputs for the worked calculation, not observed launches. Signed error equals actual minus forecast. Pre-launch MASE is undefined because no eligible target-series naive denominator exists. The analogue baseline is frozen before candidate comparison.
Unit: illustrative garments per forecast period. Status: illustrative. Sources: S12. Illustrative figures are teaching examples, not measured factory results.
Aggregation needs the same care as the formula. An average of per-launch ratios gives each eligible launch equal weight; a ratio of summed errors gives greater influence to high-volume cases and large misses. Neither is universally correct. The protocol names the estimand before the run and shows both where they answer different planning questions. Error distributions and signed bias by horizon then reveal whether a central summary conceals repeated underproduction at the commitment horizon or overproduction during the early ramp.
12.6
A useful central forecast can still carry a damaging distribution. It may place too little probability on a large order, issue intervals that rarely contain realised demand or produce broad ranges that cover outcomes without helping a planner choose. Proper scoring rules assess the predictive distribution against the event that materialises and encourage honest probabilistic assessments.[S14] The release harness therefore evaluates the complete published distribution, not only its median or mean.
Quantile loss is retained for every published quantile and horizon. A distributional score such as the continuous ranked probability score can summarise the full cumulative forecast where the output contract supports it. Empirical coverage records how often realised demand falls inside a stated central interval, while interval width records the price of that coverage in demand units. These measures answer different questions and cannot be substituted for one another.
Retail uncertainty evaluation supports this multi-quantile view. The M5 uncertainty competition required nine quantiles to describe complete future-sales distributions, including central and tail outcomes.[S16] The cold-start harness need not copy those exact quantiles. It should evaluate the quantiles the planning service actually publishes, with tail definitions fixed before the replay rather than chosen after the errors are visible.
Coverage is reported by product family, channel, horizon and launch age. Aggregate coverage can conceal a service that is conservative for institutional orders and overconfident for size-level consumer demand. Width belongs beside every coverage figure, and observed support belongs beside both. A narrow interval that misses often is not precise skill. A very wide interval that covers almost everything may still be unusable for a material commitment.
Calibration itself requires validation. Kuleshov and colleagues provide a regression-calibration procedure and explicitly qualify its guarantee by the availability of enough data.[S15] The calibration map is fitted on training or dedicated calibration launches, never on the final test outcomes. Where the launch corpus is small, the readout shows empirical coverage descriptively and avoids claiming that a calibrated population-level frequency has been established.
The audit also checks quantile coherence and output validity. Retail sales research on count data notes that point forecasts alone are insufficient for supply-chain and inventory decisions and that full discrete predictive distributions should be evaluated with proper scoring rules rather than relying only on continuous approximations or point metrics.[S1202] Quantiles must not cross, demand support must match the contract and any count transformation must be identical in replay and operation. Negative demand or impossible order quantities cannot be silently clipped after scoring. If a repair is required, it belongs in the candidate specification and is evaluated as part of that candidate. This keeps distribution quality attached to the forecast planners would actually receive.
12.7
The same numerical error can have different consequences across launch conditions. An overforecast before a material commitment can create excess output, while an underforecast during a constrained replenishment interval can leave demand unobserved. Sales-only estimates can be negatively biased when stockouts censor the focal product, so availability must shape both eligibility and interpretation of forecast error.[S06] The readout therefore separates fully observable, partly censored and unobservable periods under rules fixed before comparison.
Results are cut by product family, channel, horizon and availability state. The purpose is not to multiply tables until a favourable cell appears. Each segment corresponds to a decision the service will actually support, and the complete set is reported together. A candidate that improves the style total while worsening the size-level commitment has not passed an undifferentiated release gate. A candidate that performs well only when stockout-affected cases are removed has evidence for a narrower deployment condition.
The planning replay then maps each frozen forecast distribution into the decision rule under test. That rule may select a first lot, reserve capacity or release a replenishment quantity. Demand quantiles remain separate from capacity, material, cash and commercial constraints. The protocol records the rule and its cost or loss weights before revealing actuals. It then compares the outcome of using the candidate with the outcome of using each baseline on the same information and constraints.
This step is necessary because a lower statistical loss does not guarantee a better system result. Google's engineering guidance explicitly rejects a change that improves log loss while degrading system performance.[S33] In an apparel launch, a small central-error gain can be irrelevant if it increases severe overproduction, misses a decision deadline or creates forecasts that planners must repeatedly override. The final readout asks whether the model improved the approved planning result after those consequences are included.
Exhibit 57 illustrates that reconciliation with constructed decision-value points. The opening gross planning benefit is 100 points. Deductions are 18 for stockout distortion, 12 for forecast-override effort, 9 for compute, 7 for integration and 11 for monitoring, leaving 43 points because 100 minus 18 minus 12 minus 9 minus 7 minus 11 equals 43. These illustrative values show the required arithmetic only. They are not money, a return estimate or a benchmark for a real venture.
Read-out
Illustrative constructed case, not observed value or a return forecast. Inputs are gross planning benefit 100, stockout distortion 18, override effort 12, compute 9, integration 7 and monitoring 11. Net equals 100 - 18 - 12 - 9 - 7 - 11 = 43. Actual inputs must be measured for the pilot and stated decision.
Unit: illustrative decision-value points. Status: illustrative. Sources: S06, S33, S39. Illustrative figures are teaching examples, not measured factory results.
Offline decision replay has a limit. If a larger recommended lot would have changed availability, sell-through or the later training record, the observed counterfactual is not present in the file. The report must not pretend that the complete outcome of a different production history is known. It can score the fixed decision rule under documented assumptions, mark the feedback limitation and require a controlled pilot for claims that depend on changed operations.
12.8
MASE where eligible
MASE where eligible: Product 2, Channel 2, Horizon 2, Availability 1 (GFT requirement code: 2 mandatory, 1 diagnostic) · derived
Release is a comparative decision, not a demonstration that a candidate can produce a plausible forecast.
MAE and signed bias
MAE and signed bias: Product 2, Channel 2, Horizon 2, Availability 2 (GFT requirement code: 2 mandatory, 1 diagnostic) · derived
Quantile loss
Quantile loss: Product 2, Channel 2, Horizon 2, Availability 1 (GFT requirement code: 2 mandatory, 1 diagnostic) · derived
Coverage and width
Coverage and width: Product 2, Channel 2, Horizon 2, Availability 1 (GFT requirement code: 2 mandatory, 1 diagnostic) · derived
Planning outcome
Planning outcome: Product 2, Channel 2, Horizon 2, Availability 2 (GFT requirement code: 2 mandatory, 1 diagnostic) · derived
Runtime and failure behaviour
Runtime and failure behaviour: Product 1, Channel 1, Horizon 2, Availability 1 (GFT requirement code: 2 mandatory, 1 diagnostic) · derived
Independent launch support
Independent launch support: Product 2, Channel 2, Horizon 2, Availability 2 (GFT requirement code: 2 mandatory, 1 diagnostic) · derived
Derived release-readout matrix from availability, scaled-error, proper-scoring and system-value principles. Codes are control requirements, not measured performance. MASE appears only where its denominator is eligible. Every cell also carries independent launch support and, where relevant, interval width.
Unit: GFT requirement code: 2 mandatory, 1 diagnostic. Status: derived. Sources: S06, S12, S14, S33.
12.8 · the bodyExhibit 58 above; its figures are illustrative where marked.
Release is a comparative decision, not a demonstration that a candidate can produce a plausible forecast. The comparison set includes the frozen analogue prior, eligible naive methods, an attribute baseline and a small combination baseline. M4 results caution against assuming that pure machine-learning methods automatically beat naive and combination benchmarks.[S10] A broad review of forecast combinations likewise finds that equally weighted averages can perform well against more elaborate weighting schemes.[S20] These rivals stay in the harness throughout the model ladder.
The first gate is evidence integrity. Every origin manifest, launch exclusion, as-of join and fold artefact must pass. The second is support: independent launch counts and effective cell sizes must meet the pre-registered release condition, or the result remains preliminary. The third is paired point performance, including MAE, eligible MASE, bias and the agreed baseline-relative measure. The fourth is distribution quality across the quantiles and intervals intended for planners.
The fifth gate is decision performance under the frozen planning rule, including availability-sensitive segments and operating consequences. The sixth is operating fitness: runtime, failure behaviour, data dependencies, monitoring effort, reproducibility and the tested baseline fallback. Technical-debt research warns that a small accuracy benefit purchased through a large increase in system complexity is rarely wise.[S39] A candidate therefore earns deployment as a complete service component, not as an isolated score.
Exhibit 58 summarises why no single metric can carry that decision. In this illustrative GFT control coding, a value of 2 means the cut is mandatory for the release readout, 1 means it is a required diagnostic or sensitivity cut, and 0 would mean not applicable, although no zero is needed in this design.[S06][S12][S14][S33] The matrix is a derived GFT control requirement, not an empirical score or a claim that one candidate has passed.
a value of 2 means the cut is mandatory for the release readout
S06, S12, S14, S33 · illustrative
The release protocol names its owners before the final readout. The evidence owner certifies corpus and as-of controls. The modelling owner certifies reproducible fitting and scoring. The planning owner confirms that metrics and thresholds represent the decision. NIST guidance calls for human judgement in choosing trustworthiness metrics and their precise thresholds, which supports explicit approval rather than an automatic promotion rule.[S36] The release and reversal conditions are signed before the preferred candidate's result is seen.
Four outcomes should remain available. A candidate can graduate for a defined product, channel and horizon. It can remain a shadow challenger while more launches accumulate. It can be rejected because the evidence or operating burden is unfavourable. Or the whole comparison can be declared insufficient, leaving the approved prior or baseline in place. The last outcome is a successful control decision when support is weak, not a failed modelling exercise.
Re-entry follows the same discipline. A candidate that failed calibration, arrived too late or depended on unstable features does not return after a cosmetic repair. Its changed pipeline receives a new version and repeats the relevant launch replay on evidence permitted for that decision. Any final holdout consumed by an earlier release decision is no longer described as untouched. Fresh independent launches provide the cleanest confirmation.
The handover pack contains the corpus and origin manifests, fold artefacts, paired forecasts, metric eligibility, segment support, distribution diagnostics, planning outcomes, runtime evidence, approvals and fallback test. It tells the service owner exactly which claims were tested and which remain open. Later monitoring can then compare live behaviour with the validated conditions rather than treating release as permanent proof. The backtest's product is not a winning algorithm. It is a bounded, reversible decision with an inspectable evidence trail.
Check the record
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.
19 sources · 45 citation marks on this pageMost cited: S02, 3 times
Grade AGrade BGrade CKnot size: citations on this page
All 19 records shown
R06 · Grade C · inherited record
Inherited evidence R06
About 95 percent of cotton fibre exported raw against 5 percent processed locally; industrialisation target of 50 percent local processing.
Publication date unrecorded. Inherited source-text record. The original publication has not been independently reopened for this web edition.
R07 · Grade C · inherited record
Inherited evidence R07
SND30 restates the 50 percent local-processing target by 2030.
Publication date unrecorded. Inherited source-text record. The original publication has not been independently reopened for this web edition.
c12-S02 · Grade B
Fashion Retail: Forecasting Demand for New Items
To address this problem, we present a study of large scale fashion sales data and directly infer which clothing/footwear attributes and merchandising factors drove demand for those items.
2019-06-27. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c12-S05 · Grade B
Deep learning for new fashion product demand prediction: integrating visual similarity and demand correction in cold-start scenarios
By matching new products to these clusters, the model transfers corrected demand signals from similar historical products to estimate future demand.
2026-01-08. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c12-S06 · Grade B
Retail forecasting: research and practice
Demand estimates using only sales data would result in a negative bias in demand estimates of the focal product.
2018-11-19. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c12-S10 · Grade B
The M4 Competition: Results, findings, conclusion and way forward
The six pure ML methods performed poorly, with none of them being more accurate than the combination benchmark and only one being more accurate than Naïve2.
2018. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c12-S12 · Grade B
Another look at measures of forecast accuracy
Instead, we propose that the mean absolute scaled error become the standard measure for comparing forecast accuracy across multiple time series.
2006. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c12-S13 · Grade B
3.1. Cross-validation: evaluating estimator performance
Note that unlike standard cross-validation methods, successive training sets are supersets of those that come before them.
2026-10-30. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c12-S14 · Grade B
Strictly Proper Scoring Rules, Prediction, and Estimation
In prediction problems, proper scoring rules encourage the forecaster to make careful assessments and to be honest.
2007. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c12-S15 · Grade B
Accurate Uncertainties for Deep Learning Using Calibrated Regression
Here, we propose a simple procedure for calibrating any regression algorithm; when applied to Bayesian and probabilistic models, it is guaranteed to produce calibrated uncertainty estimates given enough data.
2018-07-03. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c12-S16 · Grade B
The M5 uncertainty competition: Results, findings and conclusions
To do so, the competition required the prediction of nine different quantiles (0.005, 0.025, 0.165, 0.250, 0.500, 0.750, 0.835, 0.975, and 0.995), that can sufficiently describe the complete distributions of future sales.
2022. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c12-S20 · Grade B
Forecast combinations: an over 50-year review
That is, equally weighted averages, which ignore past information regarding the precision of individual forecasts and correlations between forecast errors, work reasonably well compared to more sophisticated combination schemes.
2022-09-23. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c12-S32 · Grade B
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.
Open the recorded source ↗c12-S33 · Grade C
Rules of Machine Learning: Best Practices for ML Engineering
If there is some change that improves log loss but degrades the performance of the system, look for another feature.
2025-08-25. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c12-S36 · Grade A
Artificial Intelligence Risk Management Framework (AI RMF 1.0)
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.
Open the recorded source ↗c12-S37 · Grade B
Leakage and the reproducibility crisis in machine-learning-based science
Data leakage is a spurious relationship between the independent variables and the target variable that arises as an artifact of the data collection, sampling, or pre-processing strategy.
2023-08-04. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c12-S39 · Grade B
Hidden Technical Debt in Machine Learning Systems
Research solutions that provide a tiny accuracy benefit at the cost of massive increases in system complexity are rarely wise practice.
2015. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c12-S1201 · Grade B
Evaluating time series forecasting models: An empirical study on performance estimation methods
However, in real-world scenarios, when different sources of non-stationary variation are at play, the most accurate estimates are produced by out-of-sample methods that preserve the temporal order of observations.
2019-05-28. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c12-S1202 · Grade B
Evaluating predictive count data distributions in retail sales forecasting
In addition, it is not sufficient to calculate point forecasts alone: we need to forecast entire (discrete) predictive distributions, particularly for supply chain forecasting and inventory control, but also for other planning processes.
2016. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗R06 Cameroon Trade Hub. Publication date unrecorded; inherited source-text check.
R07 Ecofin Agency. Publication date unrecorded; inherited source-text check.
S13 scikit-learn. 2026-10-30.
S1201 arXiv; journal version in Machine Learning, Springer. 2019-05-28.
S02 arXiv. 2019-06-27.
S05 Springer Nature, International Journal of Data Science and Analytics. 2026-01-08.
S32 Springer Nature, Data Mining and Knowledge Discovery. 2008-07-30.
S37 Cell Press, Patterns; open full text from PubMed Central. 2023-08-04.
S12 International Journal of Forecasting, Elsevier; abstract through RePEc IDEAS. 2006.
S14 Journal of the American Statistical Association, American Statistical Association. 2007.
S16 International Journal of Forecasting, Elsevier; abstract through RePEc IDEAS. 2022.
S15 Proceedings of the 35th International Conference on Machine Learning, PMLR. 2018-07-03.
S1202 International Journal of Forecasting, Elsevier; abstract through RePEc IDEAS. 2016.
S06 Lancaster University Management School, Munich Personal RePEc Archive. 2018-11-19.
S33 Google for Developers. 2025-08-25.
S39 Neural Information Processing Systems Foundation, NIPS 2015. 2015.
S10 International Journal of Forecasting, Elsevier; abstract through RePEc IDEAS. 2018.
S20 arXiv; journal version in International Journal of Forecasting, Elsevier. 2022-09-23.
S36 National Institute of Standards and Technology, U.S. Department of Commerce. 2023.
Keep the thread · chapter 12 of 14
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
01Part I: Define the decision and the prior
A greenfield maker can forecast responsibly when it states what evidence transfers, what remains uncertain and what would change the plan.
02Part I: Define the decision and the prior
A model cannot repair a target that confuses orders, sales, availability and consumption.
03Part I: Define the decision and the prior
A transparent set of comparable launches is more useful than a precise forecast whose provenance cannot be inspected.
04Part I: Define the decision and the prior
Trade, demographic and digital signals are most credible when each is assigned a limited and testable forecasting role.
05Part II: Earn evidence before scale
Stated interest becomes useful only when sampling, incentives and calibration connect it to a future purchase decision.
06Part II: Earn evidence before scale
The first pilot should be designed to reduce a named uncertainty, not merely to produce encouraging sales.
07Part II: Earn evidence before scale
The first sales history becomes valuable when availability, product context and time are captured with the transaction.
08Part III: Model with scarcity and uncertainty
A modest model with a visible assumption is the correct starting rival for every more elaborate candidate.
09Part III: Model with scarcity and uncertainty
Cross-series learning can reduce reliance on one short history, but only when the shared evidence is relevant and the transfer is tested.
10Part III: Model with scarcity and uncertainty
Zero-shot benchmark performance is a reason to test a pretrained forecaster, not a reason to skip the evidence contract.
11Part III: Model with scarcity and uncertainty
A forecast becomes a planning instrument when its uncertainty is calibrated and its component forecasts add up.
12Part IV: Prove value and graduate with evidence
The only fair test is one that denies the model every piece of information it would not have had at the decision moment.
13Part IV: Prove value and graduate with evidence
Forecasting creates value only when a named planning rhythm can understand, challenge and reverse the model's output.
14Part IV: Prove value and graduate with evidence
A new apparel maker should buy evidence in stages and let each forecasting capability earn its place through measured launch performance.