Business statistics: descriptive and inferential methods

Descriptive and inferential statistics separate what your observed sales rows show from what a sample suggests about a wider population. Keep your statistics descriptive when the question concerns only observed rows; use inferential methods to extend a sample result beyond them, with stated assumptions and uncertainty.

The distinction matters whenever a management report turns a selected set of orders into a claim about the whole business. A correct average can still support an incorrect conclusion if the selection excludes the customers or transactions the conclusion refers to.

What each method measures

An observed total can be exact for the file while the file itself is incomplete. The Census Bureau distinguishes sampling error from nonsampling error: missing coverage, reporting mistakes and processing errors can affect results independently of sampling. Reconcile the export before debating statistical precision.

Define the sales population first

Suppose a US industrial distributor wants the mean net value of completed domestic orders during September 2026. The analysis excludes canceled orders and sales tax, uses USD, and treats each completed order as one observation. Related invoice lines are combined under the order identifier before selection.

For the illustration, assume the analyst selects a simple random sample without replacement from a much larger eligible order list. The sampled fraction is negligible, orders are sufficiently independent, and the underlying order-value distribution is approximately normal. These are assumptions for the example, not conclusions that the following ten rows establish.

Record the eligible population, selection method, exclusions and period with the extract. Run the checks before trusting a spreadsheet total to catch missing values, duplicate records and inconsistent units. A sample of invoice lines would answer a different question from a sample of orders.

The same rows, described directly

The selected order values are shown in ascending order for readability; sorting is not how they were selected. The final column supplies the calculation behind the sample variance.

Calculations retain unrounded values; displayed dollar results use two decimals where needed. The identity mean times count equals total gives 1,250 × 10 = 12,500. The deviations sum to zero, which checks that they were calculated around the same mean.

This supports a direct statement: the selected orders total $12,500 and average $1,250. It does not establish the September population total. The worksheet guide to descriptive statistics covers implementation of these summaries without changing their scope.

Extend the sample result with an interval

Under the stated assumptions, use a two-sided t interval to estimate the population mean. NIST gives the mean confidence-interval formula and explains that its confidence level describes repeated-sampling coverage.

Mean confidence interval = Sample mean ± t critical value × Sample standard deviation / Square root of sample count

For 95% confidence, the critical value uses 10 - 1 = 9 degrees of freedom. The t-distribution percentile at 0.975 is 2.262157, rounded here to six decimals.

The point estimate is still $1,250. The new information is the uncertainty interval for the wider population mean. This interval does not say that most orders fall between its endpoints, or guarantee a range for next month's sales.

NIST's guidance on mean confidence limits explains the approximate-normality assumption and cautions about small or heavily skewed samples. Ten observations make the arithmetic easy to inspect; they are not a blanket recommendation for a production analysis. A statistician should review the design when orders are clustered by customer, sampling fractions are substantial or values have a long right tail.

Which result belongs in a business decision?

For reconciliation, use the complete eligible population when it is available. If all September orders are already in a reliable export, compute their actual total and mean rather than estimating them from a subset. The descriptive result is the answer to that completed-period question.

For a manual review that can inspect only selected orders, inference may support a population estimate. Keep the selection method and interval beside the estimate so managers can assess its precision. Do not multiply a sample mean into an asserted company sales total without validating the population count, design and resulting uncertainty.

Prediction is adjacent: it concerns an unobserved future outcome. A September population estimate alone does not account for October pricing, seasonality or customer mix. Likewise, a difference between groups does not establish that a commercial policy caused it.

The sales analyst owns the row definitions and descriptive calculations. The finance owner checks agreement with the reporting scope. A qualified statistical reviewer owns assumptions for inference, and the business manager decides what action the evidence justifies. The analysis-tool overview describes available statistical procedures; choosing a procedure follows the question and design.

Where it goes wrong

• Convenience selection presented as random. Reviewing only easy-to-find orders can omit entire channels. A confidence calculation does not repair that selection.
• Order lines treated as independent orders. Repeated lines change the observation unit and can understate uncertainty. Aggregate and identify the intended unit first.
• Spread confused with precision. Standard deviation describes variation among order values; standard error describes uncertainty in an estimated mean under the design.
• A complete period treated as a future guarantee. Knowing all observed transactions removes sampling uncertainty for that scope, while future business conditions remain uncertain.
• Rounding before calculation. Use full precision for variance, standard error and endpoints; round only the displayed results so the checks remain consistent.

Inspect your own sales file in Covirage

Bring an order export with identifiers, dates, net values and any supplied sample-selection fields. Covirage maps the columns, and its tools compute and check figures from those supplied rows; the external AI model explains the results with every figure traced to its rows. Create a free Covirage account and bring a sample of your own data, keeping the population and selection assumptions alongside any estimate.

See self-service analytics with Covirage for the supporting approach.