Research libraryForecasting From a Cold Start
Chapter 10 of 14 · Part III · Draft web edition
Zero-shot benchmark performance is a reason to test a pretrained forecaster, not a reason to skip the evidence contract.
The cloth is woven from this chapter's 4 exhibits: each weft is one exhibit's own data, normalised; each warp group is a section. The probe reads the exhibit's own value where it crosses. Procedural visualisation, not a measurement.
Read-out
Derived qualitative mapping from the cited abstracts. Horizontal: 0 means probabilistic output is not established by the opened abstract and 10 means it is explicit. TimesFM's low position is an evidence-register limitation, not a claim that every version lacks distribution output. Vertical: 0 means no context integration is needed and 10 means a separate path is required for all launch-contract context. Positions are not accuracy scores.
Unit: GFT rule-based evidence coordinates from 0 to 10. Status: derived. Sources: S25, S26, S27.
Overture · 1 min
A pretrained forecaster belongs in a cold-start apparel programme only as a challenger with a reversible route to production. Its external training can supply useful evidence when the target product has no sales history, but it cannot establish that the evidence transfers to a particular garment, channel and market. Chronos, TimesFM and Lag-Llama therefore enter beside the analogue prior, simple baselines and global attribute models, not above them. Each must reproduce the same launch conditions, preserve the demand contract, improve planning-relevant results and operate within the team's means. A public benchmark can justify an experiment. It cannot approve a production commitment.
This position separates two propositions that are often compressed into one claim. The first is that a model can produce a forecast without fitting parameters on the target series. The second is that the resulting forecast is useful for the decision at hand. The first proposition is a plausible reason for inclusion in a test harness. The second requires target-domain evidence, calibrated uncertainty where distributions are used, acceptable decision consequences and a working fallback. The remainder of this chapter specifies that test without dismissing the genuine convenience of pretrained inference.
10.1
Related · chapter 088 paragraphs step it
Read-out
Derived conceptual mapping from the cited method requirements. The horizontal coordinate rises from no target history at 0 to repeated target cycles at 10. The vertical coordinate rises from no transferable analogue or attribute evidence at 0 to strong documented evidence at 10. Positions are eligibility guidance, not accuracy measurements.
Unit: GFT rule-based eligibility coordinates. Status: derived. Sources: S02, S08, S18, S19, S30.
Read this exhibit in its chapter ↗Zero-shot describes the relationship between a model and an evaluation task. It does not describe an absence of training evidence. Chronos was pretrained on a large collection of public datasets and a synthetic dataset generated through Gaussian processes. Its reported benchmark covers 42 datasets, with comparable and occasionally superior zero-shot performance on new datasets relative to methods trained specifically on them. That is evidence of transfer across the study's tasks, not evidence learned without data and not a result on apparel launches.[S25]
TimesFM likewise uses pretraining on a large time-series corpus. Its authors report out-of-the-box zero-shot performance on public datasets that comes close to supervised forecasting models fitted for each dataset. The same abstract describes operation across different history lengths, prediction lengths and temporal granularities. These are useful candidate properties. They do not change the population on which the reported performance was observed.[S26]
Every result should carry three distinct labels. Pretraining evidence identifies the external series and synthetic processes used to fit the model, to the extent that the model provider discloses them. Evaluation evidence identifies the datasets and forecast origins used to report skill. Target evidence identifies what the apparel venture itself supplies at the decision moment, including attributes, availability and any valid launch observations. A zero target-history condition concerns only the third label. It cannot erase the other two.
This separation also prevents a misleading comparison with a local baseline. The pretrained candidate may have learned from extensive external histories before seeing the target, while a naive target-series baseline has access only to the launch record. The comparison remains useful because it represents the actual deployment options, but the report must not describe the former as evidence-free or the latter as the only model with assumptions. Both have evidence lineages that need inspection.
The evaluation label needs particular care when a model is used without target fitting. An unseen dataset in a public benchmark can still come from a domain, frequency or measurement process represented in pretraining. Unless the provider's dataset inventory and chronology establish otherwise, the prudent description is that the evaluation dataset was not used for fitting in the reported experiment, not that its domain was wholly unfamiliar. Where that inventory is incomplete, the limitation stays open rather than being resolved by assumption.
Exhibit 45 converts the opened model descriptions into a GFT evidence register. The horizontal coordinate records whether the cited abstract explicitly establishes probabilistic output. It does not rank model quality. The vertical coordinate records the integration work required by this paper's launch contract, whose product, price, availability and channel fields sit outside a univariate series. Positions are rule-based interpretations of the sources, not benchmark measurements.
The matrix makes two reviews unavoidable. A model whose probabilistic status is unclear needs an output-contract review before it can be compared on distribution scores. Every model needs a context review before its output can represent the launch rather than merely its sales sequence. Neither review is settled by calling the model a foundation model.
A model register should therefore store the exact model and checkpoint identifiers, provider revision, pretraining documentation, benchmark claim, input form, output form and unresolved evidence gaps. It should also state whether the candidate will be used zero-shot, adapted on reference launches or fine-tuned on target history. Those are different interventions with different validation requirements. Combining their results under one product name would hide what the evidence actually supports.
10.4
Related · chapter 119 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.
Read this exhibit in its chapter ↗Chronos converts scaled and quantised series values into tokens from a fixed vocabulary, then trains transformer-based language-model architectures with cross-entropy loss. The reported models use the T5 family and range from 20M to 710M parameters. This mechanism matters more than the linguistic analogy: the model learns a distribution over token sequences, not an understanding of garments, prices or customer intentions.[S25]
Tokenisation creates specific checks for the evaluation harness. The team should record the scaling rule, treatment of zeros and missing intervals, context length, quantisation behaviour and conversion from generated tokens back into demand units. A coarse representation may be acceptable for a noisy launch, but that is an empirical question. Very small counts, abrupt order spikes and long zero runs should be inspected in their original units so that a smooth token sequence cannot conceal an implausible operating forecast.
Sampling from a probabilistic model supplies a predictive distribution, but it does not certify calibration. The forecast service must retain enough sampled paths or quantiles to reproduce the published output, then assess interval coverage, width and quantile loss on held-out launches. A narrow distribution can be confidently wrong. A wide one can be unhelpful even when its coverage looks acceptable. Proper scoring rules provide a principled way to compare the full forecasts rather than selecting whichever quantile happens to flatter the candidate.[S14]
The Chronos results distinguish datasets included in training from new zero-shot datasets. The authors report stronger performance in the former setting and comparable, occasionally superior performance in the latter. That pattern is exactly why the apparel evaluation should preserve separate readouts for reference launches used in any adaptation and wholly unseen target launches. It also prevents the strongest reported setting from becoming the default headline for a different use case.[S25]
Model size is a deployment variable, not a proxy for cold-start relevance. The parameter range reported for Chronos gives the evaluator several possible operating points, but the smallest adequate candidate is the one that belongs in a small team's complexity budget. The selection record should show accuracy and distribution quality beside runtime and failure behaviour for each tested variant. A larger checkpoint earns its additional burden only through the same launch replay.[S25]
Lag-Llama is presented as a general-purpose model for univariate probabilistic forecasting, using a decoder-only transformer with lag covariates. Its authors describe pretraining on a diverse corpus and zero-shot generalisation on downstream datasets across domains. They also report gains from fine-tuning on relatively small fractions of previously unseen datasets. These findings support two distinct candidate modes: direct inference and documented adaptation.[S27]
The distinction controls leakage. Direct inference can be evaluated on a held-out launch with no parameter update from that launch. Fine-tuning requires the target launch, or a defined portion of its early history, to enter training. The evaluation must then replay exactly when those observations became available and keep later outcomes out of every fitting and preprocessing step. A fine-tuned result cannot be presented as a pre-launch forecast merely because the adaptation was small.
Lag covariates also have a precise operational meaning. They encode earlier values at selected offsets; they do not automatically encode product attributes, stock availability or future promotions. A forecaster that performs well after seeing a stable weekly rhythm may still have no valid representation of a first production lot whose sizes are only partly available. The input register should list the lags actually used and confirm that each exists at the forecast origin.
For both probabilistic candidates, output validation should test count plausibility as well as statistical loss. Negative demand, non-monotone quantiles, unstable samples under a fixed seed and excessive mass on impossible order sizes require explicit handling. The team should not silently clip every implausible result and call the repaired distribution a model success. Any transformation belongs in the candidate specification and is applied identically during replay and operation.
10.7
TimesFM uses a patched-decoder attention model pretrained on a large time-series corpus. Its design addresses variation in history length, prediction length and temporal granularity, while the reported zero-shot results concern a variety of public datasets. This makes it a practical challenger where launch records do not arrive in one neat shape. It does not make every shape statistically meaningful.[S26]
A flexible input interface should not redefine the forecast contract. A size-level weekly series, a style-level monthly series and an institutional order series may all fit the same technical input format while representing different events. The evaluation still fixes the target event, product grain, channel and horizon before inference. Padding, resampling and missing-value rules must preserve that meaning. Technical compatibility is not permission to aggregate away a decision.
A new product can have a very short valid history even when the model accepts a much longer one. The service should pass only observations that existed at the forecast origin and met the availability rule. It should not borrow future rows from another channel to lengthen the sequence unless that transfer is a declared feature with its own validation. Convenience in assembling a tensor can otherwise become a route for hindsight.
Prediction-length flexibility needs the same discipline. The planning team may require several horizons, but each has its own information set and decision consequence. An assortment commitment and a short-cycle replenishment decision should not share one undifferentiated accuracy score. The replay records each requested horizon separately and tests whether performance persists across origins rather than relying on one favourable launch window.
Temporal granularity is also an analytical choice. Weekly apparel demand can contain operational events that a monthly total hides, including stockouts, listing gaps and a single institutional order. Aggregating the series may improve a score while removing the variation that the production decision must manage. The chosen grain therefore follows the commitment, with any alternative grain treated as a separate candidate rather than a preprocessing default.
The public result supports a zero-shot trial on the venture's reference corpus. If the team later adapts the model, that new candidate receives a new identifier, training record and evaluation. The report should not carry over the original zero-shot claim, because the adapted model has consumed additional evidence and acquired new failure modes. Any fine-tuning, calibration or context layer is fitted inside the appropriate replay fold.[S26][S37]
This is especially important for institutional uniforms and size-level launches. Neither setting is validated by a general statement about temporal granularities. An institutional order can reflect a contract event rather than consumer sell-through, while a size variant can be censored by uneven availability. TimesFM may still produce a useful challenger for either target, but only a target-specific replay can establish that result. The public benchmark supplies the hypothesis, not the verdict.
10.10
Read-out
Derived conceptual mapping from cited method inputs. Horizontal coordinate rises from no target history at 0 to substantial valid target history at 10. Vertical coordinate rises from univariate input at 0 to mixed static, known-future and observed context at 10. Positions are not accuracy, calibration or implementation-cost measurements.
Unit: GFT rule-based eligibility coordinates. Status: derived. Sources: S02, S21, S23, S24, S28.
Read this exhibit in its chapter ↗10.10 · Introduction
The three reviewed approaches primarily consume series values, patches, tokens or lags.
Exhibit 42, chapter 09: Attribute regression
10.11 · Four integration routes carry different burdens
The first route is univariate challenger only .
Exhibit 42, chapter 09: Attribute regression
10.12 · Availability correction precedes model selection
A univariate model cannot tell whether a zero means no buyer or no stock unless that distinction is encoded upstream.
Exhibit 42, chapter 09: Hierarchical Bayes
10.10 · the bodyThe film above follows these sub-steps.
The three reviewed approaches primarily consume series values, patches, tokens or lags. Chronos tokenises scaled values, TimesFM learns from time-series patches and Lag-Llama is explicitly univariate with lag covariates. None of those input descriptions, by itself, carries style, size, price, stock availability, promotion or channel into the model. Those fields require an explicit integration path rather than an assumption that pretraining has inferred them.[S25][S26][S27]
The launch contract should not be impoverished to suit the challenger. Availability remains necessary for interpreting observed sales, product attributes remain necessary for comparing unseen items, and channel remains necessary for explaining the selling mechanism. If the pretrained model receives only a cleaned sales vector, the discarded fields still belong in the evaluation record, the analogue prior and the competing global model. A narrow model input is not a narrow decision context.
The first route is univariate challenger only. The model forecasts the valid target series and is compared with every eligible baseline on the same launches. This is the cleanest initial test because it adds no correction cascade. Context still informs eligibility, censoring and result segmentation. Failure returns directly to the approved baseline.
The second route is combination with a context model. A pretrained series forecast is combined with an analogue or global attribute forecast under weights fixed before evaluation. This can retain both temporal adaptation and product context, but it introduces dependence between components. The report must show each component, the combination rule and the result when either input fails. Combination weights are learned only inside training folds if they are learned at all.
The third route is residual or correction modelling. A context model learns to adjust the pretrained output using attributes, price or availability. This can target systematic errors, but it creates another model whose target, training population and information cut-off need validation. Research on machine-learning systems warns that correction cascades create dependencies and can make later improvements harder to assess. A small team should accept that burden only when the combined system beats both components in replay.[S39]
The fourth route is native context-capable adaptation where a documented model version supports additional inputs. The team then verifies the actual input schema, temporal role and implementation rather than inferring capability from the model family name. Static attributes, observed historical variables and known future inputs are not interchangeable. A promotion planned for next week may be legitimate for a forecast made today; the realised promotion result is not.
A univariate model cannot tell whether a zero means no buyer or no stock unless that distinction is encoded upstream. The target pipeline therefore preserves listing and availability intervals, marks censored observations and applies only an approved demand-correction method. Where the evidence cannot support a correction, the evaluation can exclude the affected interval under a frozen rule or retain it as a separate sensitivity case. It should not insert an invented demand value and reward the model for fitting it.
This order matters for every candidate, not only pretrained ones. The difference is that a convenient external model can make the cleaned vector look complete and authoritative. The forecast register should therefore link each inference input to the raw events, availability version and correction version from which it was built. A reviewer must be able to reconstruct why the model saw a particular zero, spike or missing interval.
Context also belongs in the error analysis. Results should be segmented by product family, channel, launch age, availability condition and horizon. An aggregate gain may come entirely from high-volume style totals while size-level forecasts deteriorate. Conversely, a model may improve sparse variants but fail on institutional orders. Those are different deployment decisions, and neither should be concealed by a single average across incompatible cells.
Interlude · the cloth half-woven
Woven: sections 10.0 to 10.10Still to weave: 10.13 to 10.18
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 ↗10.13
A leaderboard compresses performance across datasets into a convenient ordering. A cold-start launch asks a narrower and harder question: does this model improve the forecast for the next decision, under this measurement process, with only the information available at this origin? Fashion research describes the coexistence of demand uncertainty, missing history and seasonal effects, together with short selling seasons and fragmented product-level demand. Those conditions make a general rank an imperfect guide to a particular apparel cell.[S01]
Domain shift can enter through the product, the channel, the target and the observation process. A model trained partly on regular utility or traffic series may recognise temporal structure without ever encountering a size curve shaped by stock availability. A distributor's order pattern may differ from consumer purchases even when both are recorded weekly. A new firm may also change its prices, listings and replenishment rules during the launch, altering the process that generated the series. The transfer claim should name these differences rather than refer to domain shift as one undifferentiated risk.
The evaluator should request the model's pretraining dataset list, benchmark list, forecast origins, comparison methods and any overlap analysis that the provider can supply. This is not an accusation of contamination. It is a normal attempt to understand what zero-shot means for the cited result. If the necessary information is unavailable, the review records that limit and places more weight on the venture's own held-out launches.
The target corpus needs equal provenance. Whole launches remain held out when the claim concerns unseen products. Scaling, imputation, feature selection, residual correction and calibration are fitted only on the training side of each replay. Leakage research defines the problem as a spurious relationship created by data collection, sampling or preprocessing, and shows how it can inflate apparent model performance. Pretraining does not exempt the surrounding pipeline from that failure mode.[S37]
A fair readout also preserves launch independence. Many weekly rows from one product do not constitute many independent proofs of transfer. Closely related colour and size launches may share a campaign, stock pool or institutional order. The report should therefore show support by independent launch group beside the row count, with sparse cells marked preliminary. A model that wins after difficult launches are removed has not passed on the deployment population.
Concept-drift research concerns changes in the relationship between inputs and targets over time and reviews strategies for adaptation and evaluation. For a pretrained forecaster, drift includes changes between the external pretraining population and the launch population as well as changes after release. The initial model review is therefore the first drift analysis, not a separate administrative exercise.[S38]
The selection pack records the reference population, expected deployment population and known mismatches. After release, monitoring tracks input validity, forecast levels and dispersion, errors by launch group, availability conditions and the frequency of fallback. A shift in assortment or channel can invalidate transfer before average sales move enough to trigger a conventional alarm. The response may be renewed evaluation, a wider range or reversion to the approved baseline, depending on the evidence.
This approach preserves the value of public benchmarks without granting them the final vote. They help identify candidates whose learned representations may transfer. The local replay then tests whether that potential survives the actual product, channel, measurement and decision conditions. A poor local result is not repaired by citing a higher public rank.
10.16
A pretrained model is a service component with an operating footprint, not merely a checkpoint that produces a score. The review must measure inference time on the intended hardware, memory use, batch behaviour, failure rate and the effort required to reproduce output. It must also record hosting, licence terms, support arrangements, model revision policy and an exit path. These are procurement and engineering requirements whose values come from the candidate under test, not universal assumptions about foundation models.
The decision clock determines acceptable latency. A weekly batch forecast can tolerate a different runtime from an interactive planning tool, but both need a defined timeout and fallback. Precomputing forecasts can reduce interactive delay while creating staleness if new availability or order evidence arrives. The operating design should state when a cached forecast expires and who can request a refresh. Faster inference is not useful if it publishes an obsolete information set.
Vendor dependence requires particular attention when weights, code or service terms are outside the team's control. A provider revision can change outputs even when the input record is unchanged. The forecast register therefore pins the deployed version, retains the validation artefacts and tests any replacement before release. Where the service is hosted externally, the review also records the data sent, retention terms and the route for removing it. Those checks concern operational control; they do not replace legal review of the applicable terms.
Exhibit 46 is the minimum evidence sheet for a release review. Its entries are deliberately unpopulated because no candidate has been measured in this chapter. The values must come from tests on the intended deployment route and from documents actually supplied by the provider. An empty field is a blocker to approval, not a zero cost.
Read-out
| Model identity | Checkpoint, code, configuration and provider revision | Can the validated output be reproduced? |
|---|---|---|
| Runtime | Batch and interactive latency on intended hardware | Does the forecast meet its decision clock? |
| Resources | Memory, accelerator need and batch capacity | Can the team operate it reliably? |
| Cost | Measured inference, hosting, integration and support basis | Does planning value exceed full operating cost? |
| Rights and support | Licence, permitted use, support and revision terms | Can the service remain compliant and maintainable? |
| Data route | Fields sent, hosting location, retention and deletion path | Is information exposure understood and approved? |
| Exit | Version retention, baseline fallback and migration test | Can planners continue without this candidate? |
Illustrative GFT register derived from the cited technical-debt and monitoring guidance. No runtime, price, memory or vendor performance is claimed. Every value must be measured or documented for the exact candidate and deployment route before release.
Unit: required release-review entries. Status: illustrative. Sources: S39, S40. Illustrative figures are teaching examples, not measured factory results.
Machine-learning technical-debt research warns against tiny accuracy benefits purchased through large increases in system complexity. The relevant comparison is therefore not candidate inference cost alone. It includes data preparation, correction layers, monitoring, version management, planner training and the work needed to recover from failure. If a baseline or smaller global model produces equivalent planning value, the additional burden has not earned approval.[S39]
Monitoring guidance likewise treats data statistics, online performance, notification and rollback as connected controls. A candidate without those controls is an experiment, regardless of its offline score. The release pack should name the observed signal, the person or role responsible for review, the action threshold and the forecast that planners receive when the threshold is crossed.[S40]
10.18
Validate data and version
Validate data and version. From Review cause and requalify (New evidence). Leads to Run pretrained candidate (Pass), Return approved baseline (Fail). · derived
The acceptance test is a paired comparison, not a demonstration that the pretrained service can return plausible numbers.
Run pretrained candidate
Run pretrained candidate. From Validate data and version (Pass). Leads to Compare quality and planning value. · derived
Run approved baselines
Run approved baselines. Starting point. Leads to Compare quality and planning value. · derived
Compare quality and planning value
Compare quality and planning value. From Run pretrained candidate, Run approved baselines. Leads to Release monitored forecast (Clear all gates), Return approved baseline (Mixed or insufficient). · derived
Release monitored forecast
Release monitored forecast. From Compare quality and planning value (Clear all gates). Leads to Return approved baseline (Monitoring breach). · derived
Return approved baseline
Return approved baseline. From Validate data and version (Fail), Compare quality and planning value (Mixed or insufficient), Release monitored forecast (Monitoring breach). Leads to Review cause and requalify. · derived
Review cause and requalify
Review cause and requalify. From Return approved baseline. Leads to Validate data and version (New evidence). · derived
The method behind it: Worked example: beating a scaling reference is not beating the baseline.
Derived GFT control sequence from cited system-value, technical-debt and monitoring guidance. This is a proposed operating design, not an implemented GFT client service or a measured result.
Unit: forecast release and reversion sequence. Status: derived. Sources: S33, S39, S40.
10.18 · the bodyExhibit 48 above; its figures are illustrative where marked.
The acceptance test is a paired comparison, not a demonstration that the pretrained service can return plausible numbers. Every candidate enters the same rolling-origin, held-out-launch replay as the analogue prior, eligible simple models and the global attribute baseline. The target, availability correction, horizons and information cut-offs remain fixed. Only the declared model capability changes. This makes any improvement attributable to the candidate rather than to a more favourable dataset or preprocessing route.
Point accuracy begins with baseline-relative measurement. MASE scales absolute error by a naive reference and supports comparison across heterogeneous series. In this paper, its denominator must exist at the forecast origin and cannot be invented for a no-history or all-zero training series. Where MASE is undefined, the report retains MAE, bias and aggregate relative absolute error against the frozen analogue prior, together with the number of eligible launches.[S12]
Distribution quality is assessed separately. Proper scoring rules compare probabilistic forecasts without reducing them to a favourable mean or median. Coverage and interval width remain visible beside those scores because a probabilistic architecture does not guarantee calibrated tails. The release comparison should include exactly the output form that planners would receive, including any count transformation, context correction or combination layer.[S14]
The illustrative readout in Exhibit 47 extends that lesson to three comparisons. Indexed loss below 100 favours the challenger. The constructed candidate beats the analogue prior on central error and the best baseline on planning loss, but fails against the global model on central error. It is not released. Selecting only the favourable panels would change the governance conclusion without changing the evidence.
A planning-relevant improvement must then survive the decision review. Practical machine-learning guidance treats realised system performance as more important than an isolated loss improvement. For apparel, the review asks whether the forecast changes a viable lot, replenishment or inventory decision for the better after availability and override effects are included. A small statistical gain that leaves the plan unchanged cannot carry additional operating burden.[S33]
The release charter names the primary metrics, eligible launch cells, minimum support, acceptable planning consequence and operating limits before results are computed. It also states what happens when evidence is thin or mixed. The default is to retain the approved baseline and gather more launches, not to search additional metrics until the challenger wins. Later model variants repeat the complete comparison rather than inheriting approval from an earlier checkpoint.
Exhibit 48 places the fallback inside the serving sequence. The quality gate protects the input and model version, while the comparison gate protects the planning output. Reversion is a normal tested path, not an admission that the forecasting service has failed. It preserves a usable decision when the challenger cannot support one.
The resulting service treats pretrained forecasting as an option whose value can be measured and withdrawn. Public research supplies capable candidates. Owned launch evidence decides whether they deserve influence. The baseline remains available at every cycle, and each release remains reversible when transfer, operations or planning value deteriorates.
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.
11 sources · 33 citation marks on this pageMost cited: S25, 6 times
Grade AGrade BGrade CKnot size: citations on this page
All 11 records shown
c10-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 ↗c10-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 ↗c10-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.
2007-03. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c10-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 ↗c10-S26 · Grade B
A decoder-only foundation model for time-series forecasting
Our model is based on pretraining a patched-decoder style attention model on a large time-series corpus, and can work well across different forecasting history lengths, prediction lengths and temporal granularities.
2024-04-17. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c10-S27 · Grade B
Lag-Llama: Towards Foundation Models for Probabilistic Time Series Forecasting
We present Lag-Llama, a general-purpose foundation model for univariate probabilistic time series forecasting based on a decoder-only transformer architecture that uses lags as covariates.
2024-02-08. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c10-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 ↗c10-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 ↗c10-S38 · Grade B
A Survey on Concept Drift Adaptation
Concept drift primarily refers to an online supervised learning scenario when the relation between the in- put data and the target variable changes over time.
2014-04-30. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c10-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 ↗c10-S40 · Grade C
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 ↗S25 arXiv. 2024-11-04.
S26 arXiv. 2024-04-17.
S27 arXiv. 2024-02-08.
S14 Journal of the American Statistical Association, American Statistical Association. 2007-03.
S37 Cell Press, Patterns; open full text from PubMed Central. 2023-08-04.
S39 Neural Information Processing Systems Foundation, NIPS 2015. 2015.
S01 InTech, International Journal of Engineering Business Management. 2013.
S38 ACM Computing Surveys; open record from Bournemouth University. 2014-04-30.
S40 Google Cloud Documentation. 2024-08-28.
S12 International Journal of Forecasting, Elsevier; abstract through RePEc IDEAS. 2006.
S33 Google for Developers. 2025-08-25.
Keep the thread · chapter 10 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.