A feature can have an interesting relationship with the target and still make a retrieval-based forecast worse. That is a useful design question. It is not, on its own, evidence that an entire class of features belongs outside the model.
Exploratory work on livestock forecasting led us to ask whether some lagged economic signals should influence a forecast after analogue selection, rather than being included in the distance used to retrieve analogues. The experiments motivated a hypothesis; they did not validate a universal architecture.
Why the location of a feature matters
In an analogue model, every feature in the distance measure helps define what counts as a similar historical situation. Adding a variable can change the selected neighbours even when the existing features remain the same.
A useful question is therefore more specific than “does this feature correlate with the target?” It is: “does including it retrieve historical examples whose subsequent outcomes are more useful for this decision?”
Marginal correlation is neither necessary nor sufficient for predictive usefulness. A variable may interact with others, identify a particular condition, duplicate existing information or capture an unstable relationship.
Two possible roles for a signal
A state descriptor helps represent the current configuration of the market: prices, recent changes, supply measures or seasonal position. A leading indicator may contain information about a later outcome through a lagged relationship.
These roles can overlap. They are prompts for model design and evaluation, rather than mutually exclusive categories into which every feature can be placed.
A lagged signal could be included in retrieval, used to condition a correction after retrieval, or excluded. The appropriate choice depends on its information timing, stability, interaction with other features and measured performance on new data.
What the exploratory work established
Historical development notes recorded several beef-market feature configurations and a later residual-correction experiment. Milk-price signals were among the features investigated.
The configurations were not clean, isolated milk ablations: oil-price inputs also changed, and tuning budgets differed. Their selected tuning scores therefore cannot establish the causal effect of adding milk. The recorded correlation does not establish the agricultural mechanism or a stable forecast relationship either.
The later correction test worsened performance on the later test period. Limited data and changing regimes are possible explanations, but that result does not establish that the architecture is correct and only the dataset is at fault.
The defensible conclusion is narrower: those experiments did not demonstrate a useful correction, and the feature-placement hypothesis remains to be tested.
A more discriminating test
A follow-up evaluation should specify three candidates in advance: a baseline retrieval model, the same model with the additional feature in its distance, and the same baseline with a separately fitted correction.
Data, timing rules and tuning budgets should be held consistent where possible. Any correction should be fitted only on earlier data. The confirmation period and scoring population should be fixed before inspecting its results.
The test should also retain the retrieved neighbours. If a feature changes performance, examining those selections can help distinguish better retrieval from a change in forecast construction.
The practical engineering lesson
Feature placement is an architectural choice that deserves an explicit test. Preserve enough provenance to know what changed, compare like with like, and let an unsuccessful experiment narrow the claim.
That discipline is useful beyond forecasting. A plausible explanation becomes much more valuable when the system makes it possible to test—and reject—it.