Global models borrow strength across products and attributes

N-BEATS was proposed for univariate point forecasting through backward and forward residual links and a deep stack of fully connected layers. Its generic configuration avoids time-series-specific components, and the authors evaluated it across varied target domains.[S23] This makes it a valuable architectural comparator: it tests how much forecast skill can come from sequence modelling alone.

That same input restriction defines its limitation at launch. A univariate model receives a target series, not the full attribute, availability and channel contract described earlier. Before any target observations exist, it has no target sequence from which to forecast. An N-BEATS candidate for a genuinely unseen item would therefore need an additional transfer mechanism, such as an analogue-supplied path or a separately learned initial state. That mechanism becomes part of the model and must be validated with it.

Once a short valid history exists, N-BEATS can test whether a flexible sequence representation beats naive, smoothing or recurrent alternatives at the required horizons. The comparison should use exactly the same observations and availability corrections. Giving the deep candidate a cleaned target while the baseline receives raw sales would confound architecture with data preparation.

The absence of covariates also creates a clean ablation. If the univariate model performs as well as a mixed-input alternative, the additional attributes and exogenous pipelines have not demonstrated enough value to justify their operation. If it fails where availability or price changes drive the series, the result shows why the launch context cannot be discarded merely to fit an architecture.

Generic interpretation should not become a causal story. A basis component or attention-like visualisation may describe how a model formed a prediction. It does not establish that a fabric, promotion or calendar event caused demand to change. The planning review should keep predictive diagnostics separate from claims about commercial mechanisms.

The N-BEATS register should state the minimum valid history, treatment of leading zeros, required sequence length, forecast horizons and fallback when those conditions fail. A model that returns a number despite an empty or availability-corrupted input is not being flexible. It is concealing an unmet assumption.

For the pre-launch decision, the eligible control remains the frozen analogue or attribute prior. N-BEATS enters the ladder only when its required sequence exists or when a complete, separately validated transfer design gives it a legitimate entry point. This keeps the comparison honest about which problem each architecture can actually solve.

A global model learns the population it is given. When the target assortment, price structure, channel or market moves away from that population, old relationships can become unreliable even if the data pipeline remains technically healthy. Concept-drift research distinguishes changes in the relationship between inputs and target from merely observing new input values.[S38] Both matter, but they call for different diagnostics.

Visual transfer has an additional dependency: the historical catalogue must remain relevant, and its demand signals must remain correctly interpreted. The cited visual cold-start architecture transfers corrected demand from similar products, so a shift in the neighbour population or correction assumptions can alter the forecast before any local sales trend becomes obvious.[S05] Similarity monitoring therefore belongs beside sales and error monitoring.

The first group of measures describes the incoming population. It tracks attribute coverage, distance to training examples, price-band mix, channel mix, launch timing and the share of items outside documented training ranges. These measures can warn that the model is extrapolating. They do not prove that accuracy has already deteriorated.

The second group describes observation quality. It tracks stockouts, listing exposure, missing attributes, image versions, returns and late-arriving corrections. A change in any of these can alter the target or its measurement without a change in underlying customer interest. Data-quality alarms should prevent automatic retraining until the demand owner has interpreted the cause.

The third group describes realised performance. It tracks bias, error, interval coverage and planning outcomes by product family, channel, horizon and availability condition. Aggregate sales can remain stable while the model fails on a newly important segment. Segmented results are therefore essential to locating a transfer failure.

A drift alarm should open a review rather than trigger unquestioned retraining. Adding recent data can help when the new launches belong to the population the service is meant to serve. It can hurt when those launches are stockout-contaminated, structurally different or too few to support a stable update. The response may be to reweight the corpus, revise a feature, widen uncertainty, gather a pilot or revert to the frozen baseline.

Every candidate must also clear the same leakage controls used during its initial replay. A new feature selected after inspecting the held-out result cannot remain in an unchanged performance claim. A new analogue set requires a new frozen evaluation. The leakage literature's requirement for clean separation and legitimate features applies to each revision, not only to the first model.[S37]

The derived control path in Exhibit 44 connects corpus admission, unseen-launch evaluation, planning value and monitored fallback. It treats rejection and reversion as successful controls when evidence is inadequate. It is a proposed GFT operating sequence, not evidence of a completed implementation.

Statistical improvement remains only one admission condition. Practical machine-learning guidance prioritises realised system performance when a model change improves loss but degrades the resulting system.[S33] For a launch forecast, the final review asks whether the candidate improved a defined planning decision after accounting for availability, overrides and operating failures. If not, the baseline remains the better service.

The chapter's answer is consequently conditional. Global models are valuable when they encode a transferable population, give unseen products a legitimate entry point and preserve a route for local learning. Their strength is shared evidence under inspection. When that population shifts or cannot be established, the disciplined action is to stop borrowing and return to a prior or baseline whose uncertainty the planning team can still defend.