Turning a request into an answerable analytical question

An operations manager asks why sales have fallen. The request sounds precise, yet it contains several unresolved questions. Did order volume fall, or did the value of each order change? Is the comparison with the preceding month or the corresponding month last year? Do sales include cancellations and returns? Which decision would change once the answer is known? A data specialist resolves these questions before opening a spreadsheet or choosing a chart. A correct calculation can still support the wrong decision when its question is poorly framed.

A decision is a practical choice between available actions. An analytical question identifies information needed to distinguish those actions. The decision might concern adding evening delivery staff. A relevant question would compare the proportion of orders delivered late across work periods, alongside the number of orders. Counting complaints alone is insufficient: complaints can increase as activity grows even when the late-delivery rate remains unchanged. Establish whether the decision concerns capacity, service quality or pricing, then choose variables that distinguish the relevant explanations.

Write a scope card beginning with the decision maker, deadline and available alternatives. Next define the observational unit: the entity represented by a row, such as an order, customer, visit or operating day. Specify the target population, for example completed orders in two regions, and a time period using a stated time convention. Record exclusions and their reasons. Excluding documented test orders is reasonable. Excluding late orders because they distort the average removes the phenomenon being investigated. Preserve the count and value of excluded records so their effect can be assessed.

Describing events, explaining causes and predicting outcomes are different tasks. A descriptive question summarizes a defined dataset, such as a fall from one thousand orders to nine hundred. A diagnostic question investigates where that change is concentrated and which explanations fit, such as a larger decline in one channel. A causal question requires a design that distinguishes the effect of one factor from other changes. A campaign coinciding with higher sales does not isolate the campaign's effect. A forecast estimates a future outcome with expected error; it is not a promise. Label the question type so a descriptive finding does not become a causal claim during presentation.

Begin with testable explanations rather than a story favored by management. Falling revenue can be investigated through transaction count, average transaction value, returns and product mix. For each explanation, identify an observable implication and evidence that would weaken it. A pricing explanation becomes less convincing when the net selling price is stable and transaction volume has fallen. However, an unchanged list price alone is insufficient because discounts may have changed. Put the necessary data beside each explanation. This prevents collecting large files that do not inform the decision.

The comparison belongs in the question; it is not a cosmetic choice made after results appear. Comparing a holiday week with a full operating week can mix changes in operating time with changes in performance. Values per operating day or comparable periods may help, provided the adjustment serves the question. If the decision concerns total monthly cash receipts, the total remains relevant even when the number of operating days differs. Show a rate alongside it to explain the difference. Do not replace the business question with an easier measure simply because that measure is available.

Agree what constitutes a sufficient answer before calculating. For operational triage, identifying the work period responsible for most additional late deliveries may justify further investigation, provided the number of observations is adequate for that purpose. A substantial investment decision may require alternative costs, uncertainty estimates and a trial. The analyst does not set a financial authorization threshold on behalf of the decision maker. Record the agreed threshold, or state that it is still needed. Discovering a pattern and authorizing action are separate responsibilities.

Check whether the question can actually be answered. Are both numerator and denominator available? Do fields mean the same thing in each period? Does the system record unsuccessful cases or only successful ones? A completed-transactions database cannot establish a purchase completion rate if visits or purchase attempts are absent. Record the gap, narrow the conclusion to the available evidence, or request an appropriate source. Never substitute zero for an unknown quantity or promise an explanation that the available design cannot distinguish.

The output is a card that another analyst can execute without reconstructing your intentions. It contains the question, unit, period, definitions, comparison, required output, limitations and decision owner. If two people would count a boundary case differently, such as an order canceled after payment, the card remains incomplete. Resolve the definition before publishing a measure whose meaning differs between teams.