Document what a sales forecast cannot predict

A sales forecast estimates an outcome under its data, target and assumptions. It does not automatically recover unmet demand, explain why sales changed, estimate causal promotion lift or anticipate events absent from every input. Writing these limits is part of model quality because downstream users otherwise fill the gaps with stronger claims than the model supports.

Turn scope into a capability matrix

The local time-series lab encodes five claims for its teaching fixture.

from timeseries_cases import limitation_case

result = limitation_case()
assert result["supported"] + result["unsupported"] == 5
assert result["supported"] == 2
assert result["unsupported"] == 3
print(result["supported"], result["unsupported"])

The matrix is specific to the lab’s inputs. Another system may include inventory availability, competitor intelligence or a causal design, but each additional capability needs its own evidence.

Sales are not always demand

When inventory or capacity runs out, recorded sales are censored by supply. A model trained on sales can learn that stockout periods have low demand. Recovering unconstrained demand requires availability signals and defensible assumptions or experiments. If these are absent, label the target “observed sales” and state the limitation.

Prediction is not causal lift

A regression coefficient or forecast difference with a promotion flag does not alone estimate what would have happened without the promotion. Campaign selection, pricing and seasonality can confound the comparison. Causal lift needs an experiment or a suitable identification strategy with assumptions that can be defended. A forecasting model can still answer the conditional planning question: “What outcome is expected if this approved calendar occurs?”

Unknown interventions need scenarios

Historical uncertainty intervals generally represent variation under modeled conditions. They do not guarantee coverage for an unannounced competitor entry, policy shutdown or supply shock. Build named scenarios when decision-makers can describe plausible interventions. Keep scenario assumptions beside the outputs so users do not confuse them with probability statements.

Publish limitations with the forecast

Record target definition, data cutoff, geographic and product scope, horizons, known covariates, excluded cases, performance slices, break behavior and monitoring plan. State who reviews overrides and when the forecast should not be used. Update the document with the model; a stale limitation sheet is a stale contract.

This honesty improves adoption because planners know when to rely on the model and when to add judgment. The Data Science course develops this communication beside model construction, evaluation and deployment.

Exercise

Write ten stakeholder statements beginning “the forecast predicts…” Classify each as supported prediction, conditional scenario, causal claim, missing-target claim or unknowable intervention. For every unsupported statement, name the data or study design required to support it.

Continue learning

This article is part of the Forecasting and time-series analysis sequence. Use the neighbouring tasks when you need the prerequisite or the next application.

• Review the preceding task in Build a forecast handover for an inventory planner.
• Return to the cluster foundation in Time-series train-test splits: reproduce a real forecast date.

Reference: Forecasting: Principles and Practice on judgmental forecasting.