Research libraryForecasting From a Cold Start
Chapter 11 of 14 · Part III · Draft web edition
A forecast becomes a planning instrument when its uncertainty is calibrated and its component forecasts add up.
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 weekly quantile trajectories (p16.5, p50.0, p83.5 and p97.5) for a cold-start apparel style across an eight-week launch window, structured after the M5 Uncertainty quantile grid (S16). Values are illustrative planning examples, not observed GFT factory sales.
Unit: illustrative weekly units. Status: illustrative. Sources: S16. Illustrative figures are teaching examples, not measured factory results.
A single point forecast is an operational trap for a greenfield apparel manufacturer because it conceals asymmetric loss and breaks apart across organisational boundaries.[S08] When a venture with no historical sales commits capital to yarn, greige fabric, dye lots and sewing-line capacity, the cost of ordering too much stock rarely equals the cost of running out during an initial retail test.[S06] Commercial teams plan revenue by channel and region, merchandisers allocate budgets by garment style and colourway, and cutting-room supervisors schedule markers by individual size variant.[S02] If the forecasting engine emits only isolated point predictions at each grain, the sum of the size forecasts will contradict the style total, and the sum of the channel forecasts will contradict the factory production plan.[S17] A forecast becomes a usable planning instrument only when it satisfies two structural conditions: it must express uncertainty through calibrated predictive distributions that expose both central tendency and tail risk, and it must reconcile every disaggregated variant and aggregated total into a single coherent hierarchy.[S14] [S17]
In GFT's Cameroon textile programme, where industrial policy seeks to shift domestic cotton fibre processing from an inherited baseline of about 5 percent toward a national target of 50 percent by 2030, capital discipline at launch is paramount.[R06] [R07] Policy ambition explains why local garment manufacturing matters, but it does not guarantee sell-through for any specific shirt, uniform or dress.[R06] Historical underutilisation in domestic textile plants, including July 2023 capacity utilisation figures of 33.7 percent at Garoua, 48.9 percent at CICAM 1 and 10.83 percent at CICAM 2, alongside an estimated underutilisation cost of about CFAF 2 billion in 2020, illustrates the financial penalty when production capacity, inventory commitments and market absorption drift apart.[R12] Moving from the candidate baselines and global models of Chapter 8, Chapter 9 and Chapter 10 to an operational release requires transforming raw model outputs into calibrated quantile fans and reconciled hierarchical tables before a single production order is signed.[S15] [S17]
11.1
A point forecast forces a planner to guess downside and upside exposure, whereas a predictive distribution states those risks explicitly across the range of plausible outcomes.[S14] In early apparel launches, demand uncertainty, seasonal swings and missing history coexist, meaning that weekly demand is frequently skewed rather than symmetric.[S01] A deterministic model predicting an illustrative 150 units a week tells the production controller nothing about whether the 10th percentile sits at an illustrative 120 units or an illustrative 30 units, nor whether a wholesale placement could push the 90th percentile to an illustrative 320 units.[S14] Because manufacturing commitments carry different lead times and reversal costs, a usable cold-start forecast must publish a fixed grid of quantiles rather than a solitary expectation.[S16]
nor whether a wholesale placement could push the 90th percentile to an illustrative 320 units
S14 · illustrative
The M5 Uncertainty competition established a practical benchmark for retail demand distributions by requiring nine quantiles (0.005, 0.025, 0.165, 0.250, 0.500, 0.750, 0.835, 0.975 and 0.995) to describe future unit sales across hierarchical levels.[S16] Paired around the median (0.500), these quantiles form four nested central prediction intervals covering 50 percent (0.250 to 0.750), 67 percent (0.165 to 0.835), 95 percent (0.025 to 0.975) and 99 percent (0.005 to 0.995) of the distribution.[S16] Adopting this nine-quantile structure inside a greenfield apparel venture prevents teams from inventing ad hoc buffers on top of an unexamined mean.[S14] [S16]
Each region of the quantile grid answers a distinct operational question along the production warp.[S08] Lower-tail quantiles (0.025 and 0.165) define the downside volume that the venture can expect to clear almost regardless of early promotional disappointments, anchoring irreversible raw-material or long-lead trim commitments in an illustrative launch plan.[S16] The central quantiles (0.250, 0.500 and 0.750) bracket the operating regime for sewing-line balancing and standard shift rosters.[S16] Upper-tail quantiles (0.835, 0.975 and 0.995) quantify the surge capacity, greige fabric reservation and conditional dyehouse slots required if early repeat orders accelerate toward the top of the analogue prior.[S16] [S28]
Uncertainty is not static across the life of a new style.[S03] Prior to launch, when the distribution relies on analogue launch curves, market covariates and calibrated choice research, the gap between the 16.5th and 83.5th percentiles is wide because transfer uncertainty dominates.[S28] [S31] As availability-verified pilot observations arrive during the first four to eight weeks, the posterior distribution updates and narrows over near-term horizons while remaining appropriately wide for later seasonal peaks.[S28] Exhibit 49 illustrates how a multi-quantile fan represents central and tail trajectories across an eight-week launch window without disguising uncertainty as a single line.[S16]
the gap between the 16.5th and 83.5th percentiles is wide because transfer uncertainty dominates.…
S28, S31, S28
As Exhibit 49 shows, presenting four illustrative quantile boundaries (p16.5, p50.0, p83.5 and p97.5) makes the skewness of launch demand legible to both commercial buyers and factory schedulers.[S16] In Week 4 of this illustrative fan, the distance from the median (145 units) down to the p16.5 boundary (85 units) is 60 units, whereas the distance up to the p83.5 boundary (220 units) is 75 units, and up to the p97.5 tail (310 units) is 165 units. What Exhibit 49 cannot establish by itself is whether those nominal probability labels match reality: an uncalibrated fan is merely a graphic until its empirical coverage is verified on held-out launches.[S15]
whereas the distance up to the p83.5 boundary (220 units) is 75 units
illustrative
11.4
Nom 50%
Nom 50%: Aggregate portfolio 49, Core basics 54, Seasonal fashion 36, Sparse tail sizes 32 (illustrative empirical coverage (%)) · illustrative
Emitting a probability or quantile from a machine-learning architecture does not guarantee that the stated uncertainty is trustworthy.
Nom 67%
Nom 67%: Aggregate portfolio 66, Core basics 71, Seasonal fashion 51, Sparse tail sizes 46 (illustrative empirical coverage (%)) · illustrative
Nom 80%
Nom 80%: Aggregate portfolio 79, Core basics 84, Seasonal fashion 64, Sparse tail sizes 59 (illustrative empirical coverage (%)) · illustrative
Nom 95%
Nom 95%: Aggregate portfolio 94, Core basics 97, Seasonal fashion 81, Sparse tail sizes 76 (illustrative empirical coverage (%)) · illustrative
Illustrative comparison of empirical coverage (%) across nominal 50%, 67%, 80% and 95% intervals for an aggregate portfolio, core basics, seasonal fashion and sparse tail sizes, demonstrating why group-conditional calibration checks are required alongside portfolio totals (S15, S1101). Values are illustrative.
Unit: illustrative empirical coverage (%). Status: illustrative. Sources: S15, S1101. Illustrative figures are teaching examples, not measured factory results.
11.4 · the bodyExhibit 50 above; its figures are illustrative where marked.
Emitting a probability or quantile from a machine-learning architecture does not guarantee that the stated uncertainty is trustworthy.[S15] Whether a candidate forecast comes from DeepAR, a Temporal Fusion Transformer, Chronos or a hierarchical Bayesian regression, its raw intervals reflect internal model assumptions and training residuals rather than guaranteed cold-start coverage.[S21] [S24] [S25] [S28] Under domain shift between historical analogues and a newly launched garment, raw neural and parametric intervals frequently suffer from overconfidence, producing narrow bands that exclude realised demand far more often than their nominal level promises.[S15] [S38]
Statistical learning offers two remedies for miscalibrated regression and quantile outputs, each carrying strict operational prerequisites.[S15] [S1101] First, calibrated regression procedures use a held-out calibration dataset to learn an isotonic mapping from predicted cumulative probabilities to empirical cumulative frequencies.[S15] When applied to Bayesian or probabilistic models, this mapping is guaranteed to produce calibrated uncertainty estimates given enough data, but that sufficient-data qualification is decisive in a cold start.[S15] Fitting an isotonic recalibration curve on only three or four pilot items will overfit noise and distort the tails.[S15] [S37]
Second, conformalized quantile regression combines classical quantile regression with split conformal prediction to construct prediction intervals that attain valid finite-sample coverage without distributional assumptions.[S1101] Unlike constant-width conformal methods, conformalized quantile regression adapts to heteroscedasticity by adjusting lower and upper quantile estimates with conformity scores evaluated on a calibration split, producing tighter intervals where signal is strong and wider intervals where volatility is high.[S1101] Even so, finite-sample conformal guarantees rely on exchangeability between calibration launches and the target launch.[S1101] If the target launch introduces a new retail channel, a holiday regime or an unrecorded stockout, exchangeability fails and empirical coverage must be re-verified.[S04] [S06] [S1101]
A portfolio-wide calibration score can mask severe miscalibration in the specific segments where capital is most at risk.[S15] If an apparel maker over-covers core institutional basics (achieving an illustrative 92 percent empirical coverage on a nominal 80 percent interval) while under-covering seasonal fashion styles and peripheral sizes (achieving an illustrative 58 percent coverage on the same nominal 80 percent interval), the pooled average may sit near an illustrative 79 percent and appear sound.[S15] [S16] To prevent this aggregation illusion, GFT requires calibration audits to report empirical coverage and interval width separately across product families, channels, size tiers and horizons, as shown in Exhibit 50.[S15]
In Exhibit 50, the illustrative aggregate portfolio tracks the nominal targets (50, 67, 80 and 95 percent) within one percentage point (49, 66, 79 and 94 percent), passing a naive global check.[S15] Disaggregating the audit reveals that core basics are conservatively wide (54, 71, 84 and 97 percent), whereas seasonal fashion (36, 51, 64 and 81 percent) and sparse tail sizes (32, 46, 59 and 76 percent) suffer severe under-coverage.[S15] [S1101] A planner relying on the nominal 80 percent interval for seasonal fashion would experience misses in 36 percent of weeks instead of 20 percent. Group-conditional coverage tables paired with interval width checks ensure that recalibration is stratified by demand regime.[S15] [S1101]
11.7
Evaluating a probabilistic forecast requires metrics that assess the entire predictive distribution honestly.[S14] If a model is judged only on the root mean squared error or mean absolute error of its central tendency, forecasters have no incentive to report accurate tail probabilities.[S14] Strictly proper scoring rules solve this problem by assigning a numerical score under which a forecaster optimises the expected score if and only if the quoted distribution equals the true underlying distribution.[S14]
Probabilistic forecasting aims to maximise sharpness (concentrating probability mass into narrow, informative intervals) subject to calibration (matching stated probabilities to empirical frequencies).[S14] For quantile forecasts, the canonical proper scoring rule is the pinball loss (quantile loss), which penalises an observation y falling above a predicted tau-quantile q_tau by tau * (y - q_tau) and penalises an observation falling below q_tau by (1 - tau) * (q_tau - y).[S14] [S16] At the median (tau = 0.500), pinball loss is proportional to absolute error.[S16] At the 0.975 quantile, an under-forecast error is penalised 39 times more heavily than an over-forecast error of the same size (0.975 / 0.025 = 39), pushing the upper quantile upward until only 2.5 percent of realisations exceed it.[S14] [S16]
Operational protocol for multi-quantile scoring and coverage verification
0.005, 0.025, 0.165, 0.250, 0.500, 0.750, 0.835, 0.975 and 0.995) for every active horizon and hierarchical node.[S16]Averaging pinball loss across the quantile grid links quantile evaluation to the Continuous Ranked Probability Score (CRPS), which integrates the squared difference between the predictive cumulative distribution and the empirical step function of the outcome.[S14] Unlike the logarithmic score, which can degenerate when a discrete count outcome falls in a zero-density region, CRPS and pinball loss remain well-behaved across continuous and discrete count distributions and measure error in the unit of demand.[S14] [S1102]
Proper distributional scores complement point accuracy metrics rather than replacing them.[S12] [S14] A candidate model could improve average CRPS by tightening its extreme 0.995 tail while introducing a persistent 12 percent illustrative upward bias at the median (0.500), inflating sewing-line schedules if planners schedule to the median.[S12] [S33] Consequently, GFT's evaluation contract places proper distributional scores, empirical calibration checks, point error (MASE and MAE) and coherence diagnostics side by side in Exhibit 51.[S12] [S14] [S17]
…CRPS by tightening its extreme 0.995 tail while introducing a persistent 12 percent illustrative…
S12, S14, S12, S33, S12, S14, S17 · illustrative
Read-out
| Pinball loss (quantile score) | Asymmetric piecewise linear loss at target tau | Miscalibrated tail location or poor sharpness | Compare against prior and baseline at each tau |
|---|---|---|---|
| Empirical interval coverage | Observed share inside nominal central interval | Overconfident narrow bands or excessive width | Recalibrate via isotonic or conformal adjustment |
| Continuous ranked probability score | Integrated squared difference of cumulative distributions | Global distributional shape and dispersion error | Reject candidate if worse than pooled analogue prior |
| Scaled point error and bias | MASE where defined, plus MAE and signed percentage bias | Systematic overproduction or underproduction at median | Block release if central bias breaches planning tolerance |
| Coherence discrepancy | Max absolute residual across hierarchy summing matrix | Component size or channel forecasts fail to sum to total | Apply MinT or sample-path projection before release |
Derived diagnostic contract synthesising strictly proper scoring rules (S14), calibrated regression checks (S15), M5 multi-quantile evaluation (S16), scaled error baselines (S12) and trace-minimising reconciliation (S17).
Unit: diagnostic specification. Status: derived. Sources: S12, S14, S15, S16, S17.
As Exhibit 51 establishes, each diagnostic guards against a specific failure mode.[S14] [S17] Pinball loss and CRPS test distributional skill against the analogue prior and baselines, empirical coverage verifies interval honesty, MASE and signed bias protect central capacity planning, and coherence discrepancy checks ensure that component forecasts add up.[S12] [S15] [S17] What Exhibit 51 cannot do automatically is trade off a small gain in tail pinball loss against a worse central bias; that decision requires human thresholds set under the governance rules of Chapter 12 and Chapter 13.[S33] [S36]
Interlude · the cloth half-woven
Woven: sections 11.0 to 11.7Still to weave: 11.10 to 11.19
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 ↗11.10
Reconciled style total (1,000) to Wholesale channel (600)
Reconciled style total (1,000) to Wholesale channel (600): 600 illustrative reconciled units · illustrative
Apparel demand spans multiple levels of aggregation.
Wholesale channel (600) to Small size (200)
Wholesale channel (600) to Small size (200): 120 illustrative reconciled units · illustrative
Wholesale channel (600) to Medium size (380)
Wholesale channel (600) to Medium size (380): 230 illustrative reconciled units · illustrative
Wholesale channel (600) to Extra-large size (140)
Wholesale channel (600) to Extra-large size (140): 80 illustrative reconciled units · illustrative
Direct retail and online (400) to Small size (200)
Direct retail and online (400) to Small size (200): 80 illustrative reconciled units · illustrative
Direct retail and online (400) to Large size (280)
Direct retail and online (400) to Large size (280): 110 illustrative reconciled units · illustrative
Direct retail and online (400) to Extra-large size (140)
Direct retail and online (400) to Extra-large size (140): 60 illustrative reconciled units · illustrative
The method behind it: Worked example: reconciling an incoherent launch style across channels and sizes.
Illustrative flow conservation across a grouped apparel hierarchy after MinT reconciliation (S17), mapping a 1,000-unit monthly style forecast through two sales channels (wholesale 600, direct retail 400) into four size groups (Small 200, Medium 380, Large 280, Extra-Large 140). Values are illustrative.
Unit: illustrative reconciled units. Status: illustrative. Sources: S11, S17. Illustrative figures are teaching examples, not measured factory results.
11.10 · the bodyExhibit 52 above; its figures are illustrative where marked.
Apparel demand spans multiple levels of aggregation.[S11] In the M5 competition, accuracy and uncertainty were evaluated across 42,840 hierarchical series spanning products, departments, categories, stores and states.[S11] [S16] Even for a greenfield manufacturer launching an initial collection of an illustrative 12 to 25 styles in Cameroon, the planning structure forms both a nested hierarchy (total demand split by region, channel and account) and a grouped structure (category crossed with style, colourway and size).[S02] [S11] [S17]
accuracy and uncertainty were evaluated across 42,840 hierarchical series spanning products
S11, S16, S02, S11, S17 · illustrative
When models produce base forecasts independently at the style level, channel level and bottom SKU-size level, those forecasts rarely add up.[S17] [S1102] A global attribute model might forecast 1,000 illustrative units for a woven cotton workshirt in Month 2.[S02] Simultaneously, channel regressions might sum to 1,080 illustrative units across wholesale and direct retail, while bottom-level intermittent size models sum to 920 illustrative units.[S04] [S18]
Published without reconciliation, these numbers cause every department to act on a different plan: procurement orders cloth for 1,000 shirts, sales allocates 1,080 shirts, and the cutting room lays markers for 920 shirts.[S17] Pure bottom-up aggregation sums noisy size series and ignores style-level attribute signals, whereas pure top-down disaggregation relies on static historical proportions that do not exist in a cold start and cannot capture how size mix differs between institutional wholesale and online retail.[S17] [S1102]
Optimal forecast reconciliation generates base forecasts at all levels and projects them onto the coherent linear subspace where aggregation constraints hold exactly.[S17] [S1102] Let y_t denote the vector of all n series at time t, and b_t the m-dimensional vector of bottom-level series, linked by an n x m summing matrix S so that y_t = S * b_t.[S17] [S1102] Base forecasts y_hat map into coherent forecasts y_tilde = S * G * y_hat via a matrix G satisfying S * G * S = S, preserving unbiasedness.[S17]
The Minimum Trace (MinT) estimator chooses G using the h-step-ahead base forecast error covariance matrix W_h, minimising the trace of the coherent error covariance matrix (and thus total mean squared error under unbiasedness).[S17] Because estimating an unconstrained matrix W_h is infeasible with short launch histories, cold-start implementations use structural scaling (weighting each node by the count of bottom series it sums) or shrinkage toward a diagonal variance matrix.[S17]
Point coherence does not guarantee quantile coherence because quantiles are not additive: the 90th percentile of a style total does not equal the sum of the 90th percentiles of its individual sizes unless those sizes are perfectly correlated.[S1102] [S1103] Summing the 95th percentile of every size variant to size a fabric buffer will substantially overstate the 95th percentile of style-level fabric demand.[S1102]
To reconcile full distributions, GFT uses a two-step sample-path procedure.[S1102] [S1103] First, the engine draws B joint future sample paths (for example, an illustrative 1,000 trajectories) from the calibrated base distributions at all nodes, using empirical copula reordering or Bayesian posterior draws to retain rank correlations across sizes and channels.[S28] [S1102] Second, the engine applies the MinT projection S * G to each sample vector, yielding B coherent sample paths in which every trajectory adds up from bottom size-channel cells to style, channel and market totals.[S1102] [S1103] Under Gaussian errors, MinT projection also minimises the multivariate logarithmic score and beats ordinary least squares projection on marginal logarithmic scores.[S1102] Empirical studies confirm that probabilistic reconciliation improves both point and distributional accuracy across hierarchies.[S1103] Exhibit 52 shows how a reconciled 1,000-unit style forecast conserves volume across channels and sizes.[S17]
Exhibit 52 shows how a reconciled 1,000-unit style forecast conserves volume across channels and sizes
S17
As Exhibit 52 demonstrates, the 1,000 illustrative units at the style node equal the sum of the wholesale (600 units) and direct (400 units) channels, and each channel's size allocations sum to the cross-channel size totals (200 Small, 380 Medium, 280 Large and 140 Extra-large).[S17] What Exhibit 52 cannot correct is systematic bias in the base inputs: because MinT assumes unbiased base forecasts, stockout censorship must be corrected before projection.[S06] [S17]
the 1,000 illustrative units at the style node equal the sum of the wholesale…
S17 · illustrative
11.14
While aggregated style and channel totals behave smoothly once weekly volume exceeds several dozen units, bottom-level apparel series (such as a specific colourway in size Double-Extra-Large) are almost always sparse and intermittent.[S18] Intermittent demand occurs sporadically, with many periods recording zero demand interspersed with occasional small purchases.[S18] Applying continuous Gaussian intervals or unadjusted linear projection to sparse series creates immediate distortions: lower intervals turn negative, central quantiles take fractional values such as 0.37 shirts, and standard estimators overstate average demand.[S18] [S19]
central quantiles take fractional values such as 0.37 shirts
S18, S19
Croston's method recognised that intermittent demand requires modelling two components: the probability of nonzero demand occurrence (or inter-demand interval) and the conditional demand size when an order occurs.[S18] As Syntetos and Boylan proved, combining separate exponentially smoothed estimates of demand size and inter-demand interval by simple division introduces a positive inversion bias that systematically over-forecasts demand per period.[S19] The Syntetos-Boylan Approximation (SBA) corrects this bias by multiplying the ratio by (1 - alpha / 2), where alpha is the interval smoothing parameter.[S19]
In a probabilistic hierarchy, this separation governs bottom-level quantile generation.[S18] [S21] Rather than fitting continuous Gaussian likelihoods to zero-heavy variants, global probabilistic models such as DeepAR or hierarchical Bayesian count models use discrete likelihoods (negative binomial, Poisson or zero-inflated hurdle distributions) that place positive mass at zero and restrict quantiles to non-negative integers.[S21] [S28] When sample paths from discrete bottom distributions are projected onto the coherent subspace, negative values are resolved through non-negative constrained reconciliation or integer rounding so planners never see negative size allocations.[S17] [S1102]
Sparsity also depends on the temporal grain of the decision.[S1103] A peripheral size variant recording zeros on four out of five days looks smoother when aggregated into four-week production buckets.[S18] Temporal hierarchies reconcile forecasts across time frequencies (weekly replenishment, four-week sewing cycles and quarterly fabric windows) using the same projection logic applied to cross-sectional product hierarchies.[S1103] Whenever sparse variants feed a style or channel total, GFT's release reports both the Syntetos-Boylan bias check at the bottom grain and the sensitivity of the reconciled total to pooling the sparsest cells.[S18] [S19] [S1103]
11.17
Even a calibrated, coherent predictive distribution describes uncertainty only within its conditioning assumptions and reference class.[S08] Statistical prediction intervals state what range of weekly demand to expect at a chosen probability level given the analogue pool, pilot transactions and current offer structure.[S08] [S14] They do not state what follows if a commercial or policy condition changes outside that regime.[S08]
Greenfield apparel makers regularly confront discrete contingencies: a ministry or corporate buyer may award an institutional uniform tender, a port delay may hold up imported trims for six weeks, or management may test a 15 percent illustrative promotional discount.[S01] [S08] Treating these discrete events as random variance inside a 95 percent statistical interval corrupts both the model and the planning review.[S08] [S36]
Artificially widening a prediction interval to cover a hypothetical uniform contract destroys its empirical calibration meaning and prevents honest scoring with pinball loss or CRPS.[S14] [S15] Conversely, relying solely on the statistical fan leaves the venture exposed to regime shifts.[S08] Following forecasting theory and the NIST AI Risk Management Framework, GFT separates model-based statistical distributions from human-authored scenario overlays, assigning every scenario a named owner, a conditioning trigger and an explicit review date.[S08] [S36]
11.19
Exhibit 536 paragraphs step it
Read-out
Derived release architecture for cold-start probabilistic forecasts, combining post-hoc calibration checks (S15), trace-minimising hierarchical reconciliation (S17), as-of leakage controls (S37) and MLOps monitoring and rollback metadata (S40).
Unit: system governance layers. Status: derived. Sources: S15, S17, S37, S40.
A probability distribution is reproducible and auditable only when the exact state of the data, priors, calibration mappings and reconciliation matrices that produced it travels with the numbers.[S37] [S40] When forecasts circulate as detached spreadsheets, nobody can reconstruct two weeks later whether the intervals used pre-pilot analogue weights, whether stockout days were masked, or which calibration window governed interval widths.[S06] [S37]
Data leakage is a spurious relationship between inputs and the target arising from collection, sampling or preprocessing artefacts, and it is as hazardous in calibration and reconciliation as in feature engineering.[S37] Estimating a reconciliation covariance matrix W_h or a conformal quantile adjustment using records from Weeks 5 to 8 and applying it to a launch replay at Week 2 contaminates coverage and coherence metrics with future information.[S15] [S17] [S37]
To prevent retrospective leakage and operational confusion, GFT wraps every published forecast distribution in an immutable provenance envelope, shown in Exhibit 53.[S37] [S40] Each release package records four metadata layers: the as-of provenance layer (training cut-off timestamp, feature snapshot hash and stockout censorship mask flags), the probabilistic base layer (model ID, analogue prior shrinkage weight and discrete versus continuous family), the coherence and calibration layer (calibration method, sample size, group-conditional coverage status and reconciliation matrix hash), and the planning layer (decision horizon, separate scenario IDs and a hard expiry timestamp).[S15] [S17] [S28] [S37] [S40]
As Exhibit 53 shows, a planner inspecting a forecast in cold-start-forecast or ramp-planner never receives uncontextualised model outputs.[S40] Every figure traces through the reconciliation and calibration layer to the base model and its as-of cut-off.[S15] [S17] [S37] When calibration sample support falls below the minimum independent launch count required by Chapter 12, the envelope explicitly flags the distribution as prior-dominated rather than implying certified empirical coverage.[S15] [S39]
Structuring every cold-start forecast as a calibrated nine-quantile distribution reconciled across sizes, styles and channels eliminates the two most common excuses for ignoring quantitative evidence at launch: that a point forecast was inevitably wrong, or that the model's totals contradicted the merchandiser's size breakdown and the commercial director's channel plan.[S16] [S17] The statistical distribution states what current evidence supports and how wide residual doubt remains; the reconciliation layer ensures every operational table adds up; the scenario layer records what human leaders wish to test beyond the historical regime; and the provenance envelope preserves the exact information set for retrospective scoring.[S08] [S14] [S17] [S37] Chapter 12 now specifies how these probabilistic, hierarchical forecasts are tested in chronological launch replays without hindsight leakage, and Chapter 13 shows how planners govern, challenge and roll back the resulting forecast service inside a monthly operating cadence.[S12] [S33] [S40]
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.
34 sources · 255 citation marks on this pageMost cited: S17, 29 times
Grade AGrade BGrade CKnot size: citations on this page
All 34 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.
R09 · Grade C · inherited record
Inherited evidence R09
Camtext SA integrated cotton-to-textile project, US$300 million (180 billion XAF); Panafritex and Arise with SODECOTON, CNPS and Marlo; 12 November 2025.
Publication date unrecorded. Inherited source-text record. The original publication has not been independently reopened for this web edition.
R10 · Grade C · inherited record
Inherited evidence R10
Camtext SA: 12,000 t a year processed; about 12,000 jobs at the Dibamba hub and 3,000 at Garoua; return on investment near 11 years; optimal capacity around 2033.
Publication date unrecorded. Inherited source-text record. The original publication has not been independently reopened for this web edition.
R12 · Grade C · inherited record
Inherited evidence R12
CICAM capacity utilisation, July 2023 report: Garoua 33.7 percent, CICAM 1 48.9 percent, CICAM 2 10.83 percent; underutilisation cost about CFAF 2 billion in 2020.
Publication date unrecorded. Inherited source-text record. The original publication has not been independently reopened for this web edition.
c11-S01 · Grade B
Demand Forecasting in the Fashion Industry: A Review
This is especially the case in the fashion industry, where demand uncertainty, lack of historical data and seasonal trends usually coexist.
2013. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-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 ↗c11-S03 · Grade B
Product age based demand forecast model for fashion retail
We have modeled sales of a fashion item in a store as a function of its age, attributes, selling price, temporal and store features.
2020-07-10. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S04 · Grade B
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.
Open the recorded source ↗c11-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 ↗c11-S07 · Grade B
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.
Open the recorded source ↗c11-S08 · Grade B
Forecasting: theory and practice
In fact, there may be no historical observations at all for the item of interest, when judgmental methods must be used (e.g., time taken to complete construction of a new airport).
2022. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S11 · Grade B
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.
Open the recorded source ↗c11-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 ↗c11-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 ↗c11-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 ↗c11-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 ↗c11-S17 · Grade B
Optimal Forecast Reconciliation for Hierarchical and Grouped Time Series Through Trace Minimization
Our approach minimizes the mean squared error of the coherent forecasts across the entire collection of time series under the assumption of unbiasedness.
2019. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S18 · Grade B
The accuracy of intermittent demand estimates
Intermittent demand appears sporadically, with some time periods showing no demand at all.
2005. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S19 · Grade B
On the bias of intermittent demand estimates
In addition, a modification to his method that gives approximately unbiased demand per period estimates is introduced.
2001. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S21 · Grade B
DeepAR: Probabilistic Forecasting with Autoregressive Recurrent Networks
In this paper we propose DeepAR, a methodology for producing accurate probabilistic forecasts, based on training an auto regressive recurrent network model on a large number of related time series.
2019-02-22. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S24 · Grade B
Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting
The TFT also uses specialized components for the judicious selection of relevant features and a series of gating layers to suppress unnecessary components, enabling high performance in a wide range of regimes.
2020-09-27. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S25 · Grade B
Chronos: Learning the Language of Time Series
Chronos tokenizes time series values using scaling and quantization into a fixed vocabulary and trains existing transformer-based language model architectures on these tokenized time series via the cross-entropy loss.
2024-11-04. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S28 · Grade B
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.
Open the recorded source ↗c11-S31 · Grade B
Preference Learning and Demand Forecast
This allows us to forecast demand in real purchase settings using inexpensive choice data with small to moderate realization probabilities.
2020. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S33 · Grade B
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 ↗c11-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 ↗c11-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 ↗c11-S38 · Grade B
A Survey on Concept Drift Adaptation
Assuming a general knowledge of supervised learning in this paper we characterize adaptive learning process, categorize existing strategies for handling concept drift, discuss the most representative, distinct and popular techniques and algorithms, discuss evaluation methodology of adaptive algorithms, and present a set of illustrative applications.
2014-04-30. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-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 ↗c11-S40 · Grade B
MLOps: Continuous delivery and automation pipelines in machine learning
Therefore, you need to track summary statistics of your data and monitor the online performance of your model to send notifications or roll back when values deviate from your expectations.
2024-08-28. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S1101 · Grade B
Conformalized Quantile Regression
Conformal prediction is a technique for constructing prediction intervals that attain valid coverage in finite samples, without making distributional assumptions.
2019-05-08. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S1102 · Grade B
Probabilistic forecast reconciliation under the Gaussian framework
Forecast reconciliation of multivariate time series is the process of mapping a set of incoherent forecasts into coherent forecasts to satisfy a given set of linear constraints.
2021-03-20. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S1103 · Grade B
Reconciliation of probabilistic forecasts with an application to wind power
New methods are proposed for adjusting probabilistic forecasts to ensure coherence with the aggregation constraints inherent in temporal hierarchies.
2018-08-08. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗S08 International Journal of Forecasting, Elsevier; full text through arXiv. 2022.
S06 Lancaster University Management School, Munich Personal RePEc Archive. 2018-11-19.
S02 arXiv. 2019-06-27.
S17 Journal of the American Statistical Association, Taylor & Francis; abstract through RePEc IDEAS. 2019.
S14 Journal of the American Statistical Association, American Statistical Association. 2007.
R06 Cameroon Trade Hub. Publication date unrecorded; inherited source-text check.
R07 Ecofin Agency. Publication date unrecorded; inherited source-text check.
R12 Business in Cameroon. Publication date unrecorded; inherited source-text check.
S15 Proceedings of the 35th International Conference on Machine Learning, PMLR. 2018-07-03.
S01 InTech, International Journal of Engineering Business Management. 2013.
S16 International Journal of Forecasting, Elsevier; abstract through RePEc IDEAS. 2022.
S28 Journal of Industrial Ecology, Wiley. 2022-10-17.
S03 arXiv. 2020-07-10.
S31 INFORMS, Marketing Science; open manuscript from MIT. 2020.
S21 arXiv; journal version in International Journal of Forecasting, Elsevier. 2019-02-22.
S24 arXiv. 2020-09-27.
S25 arXiv. 2024-11-04.
S38 ACM Computing Surveys; open record from Bournemouth University. 2014-04-30.
S1101 arXiv; Neural Information Processing Systems (NeurIPS 2019). 2019-05-08.
S37 Cell Press, Patterns; open full text from PubMed Central. 2023-08-04.
S04 MDPI, Processes. 2021-09-03.
S07 arXiv. 2026.
S12 International Journal of Forecasting, Elsevier; abstract through RePEc IDEAS. 2006.
S1102 arXiv; Journal of Business & Economic Statistics. 2021-03-20.
S33 Google for Developers. 2025-08-25.
S36 National Institute of Standards and Technology, U.S. Department of Commerce. 2023.
S11 International Journal of Forecasting, Elsevier; abstract through RePEc IDEAS. 2022.
S18 International Journal of Forecasting, Elsevier; abstract through Lancaster EPrints. 2005.
S1103 arXiv; European Journal of Operational Research, Elsevier. 2018-08-08.
S19 International Journal of Production Economics, Elsevier; abstract through Lancaster EPrints. 2001.
R09 Ecofin Agency. Publication date unrecorded; inherited source-text check.
R10 Ecofin Agency. Publication date unrecorded; inherited source-text check.
S40 Google Cloud Documentation. 2024-08-28.
S39 Neural Information Processing Systems Foundation, NIPS 2015. 2015.
Keep the thread · chapter 11 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.