Simple models remain the cold-start benchmark

The ladder is a governance sequence, not a ranking of algorithms by age or technical fashion. Each rung adds one identifiable capability: transfer from analogues, filtering of a local level, explicit treatment of intermittent occurrence, use of product attributes, or combination across methods. The candidate then repeats the same launch replay against every eligible lower rung. Complexity that does not improve the decision remains in research.

This is consistent with practical machine-learning guidance to keep the first model simple and establish baseline behaviour before adding complexity. It also reflects a systems principle: a statistical gain is insufficient if the change degrades realised system performance. For apparel planning, the release review therefore includes forecast error, calibration, planning outcome and the reliability of the surrounding pipeline.[S33]

A fair replay holds constant the target event, product and channel grain, horizons, availability treatment, launch exclusions and information cut-offs. Preprocessing, attribute selection and parameter tuning occur inside the training fold. Whole target launches remain out of training when the claim concerns an unseen item. These conditions make the baseline win meaningful rather than an artefact of unequal information.

The result pack separates product families, channels, horizons and availability conditions. It reports central error and bias beside uncertainty quality. It also reports how often each method was eligible and how often it failed back to a lower rung. A candidate that performs well only after quietly dropping difficult launches has not beaten the baseline on the same decision set.

Support size limits the conclusion. When independent launches are few, the correct gate outcome is to retain the approved prior or simple baseline and collect more evidence. A close result should not be resolved by adding metrics until the challenger wins. The release protocol fixes the primary comparisons and the required planning consequence before model scores are computed.

Machine-learning systems incur maintenance costs through data dependencies, configuration, feedback and monitoring. Research on hidden technical debt specifically warns against tiny accuracy benefits purchased with large increases in system complexity. A small industrial team should apply that warning directly: every additional feature, pipeline and model dependency must justify its continued operation.[S39]

The complexity budget records runtime, failure modes, data owners, reproducibility, planner interpretability and rollback effort. It does not impose a universal threshold. Instead, it asks whether the observed improvement is large and persistent enough to change a production decision after those burdens are included. A statistically attractive result that does not alter a viable plan has little claim to deployment.

A baseline service should pass small, deterministic tests before it enters a replay. A flat valid series should produce a stable naive forecast. A level shift should affect naive and smoothing forecasts in the expected order. A sparse sequence should trigger the named intermittent estimator only after its zeros pass the availability test. A repeated cycle should make the seasonal rung eligible only at the approved lag. An unseen item with attributes but no target rows should use the transfer path, not an empty local model.

These tests check meaning as well as code. Each one states the input records, information cut-off, expected eligibility, forecast and fallback. When a result changes, reviewers can tell whether the cause was a corrected rule, a changed parameter search or an accidental alteration to the target. This is especially important for simple methods because their transparency is valuable only when the implemented calculation matches the written definition.

The pack also needs negative tests. Missing availability should prevent a zero-demand interpretation. An absent analogue identifier should prevent an attribute forecast from borrowing an unrecorded reference class. A zero naive denominator should produce an undefined MASE rather than an apparently excellent score. A failed component in a combination should trigger the frozen membership rule. Each failure should leave a visible status in the forecast register.

Ownership completes the control. The demand owner approves target and availability definitions. The forecasting owner approves model specifications and replay results. The planning owner confirms that the output can change a real commitment. These roles may be performed by a small team, but the approvals remain distinct records. A simple model then becomes a stable reference against which both future models and future data changes can be assessed.

The derived control sequence in Exhibit 39 makes this recommendation operational. It does not claim that the sequence has produced a result at a GFT client. It specifies the evidence path that a cold-start service should implement and test.

A baseline is consequently a permanent operating asset, not a disposable stage before serious modelling begins. It anchors evaluation when data are scarce, provides a fallback when advanced models drift and gives planners a visible reason to trust or reject a revision. The forecasting service becomes more capable only when the evidence, rather than the architecture, justifies the next rung.