Descriptive vs predictive analytics: compare past and estimates

Descriptive and predictive analytics answer different business questions: what the recorded results were, and what an unknown outcome is estimated to be. Descriptive analysis needs reconciled actuals; predictive analysis also needs a dated estimate, assumptions and an honest accuracy check. The same monthly sales file can support both comparisons without mixing their meanings.

What each approach measures

Descriptive analysis summarizes observations. A sales total, a regional share and a return rate describe the records included in the calculation. They remain subject to missing rows, late credits and the business definitions used to prepare the export.

Predictive analysis estimates an unknown outcome using information available when the estimate is made. That outcome is often future demand or revenue, although prediction can also estimate a current value that has not yet been observed. An estimate is evidence for planning, not a posted transaction.

The broader overview of descriptive, diagnostic, predictive and prescriptive analytics covers the full set of approaches. Here the practical question is narrower: how do observed results and saved predictions differ when a manager reviews the same measure?

Work the same rows both ways

A fictional US industrial supplier reviews monthly units shipped. Its sales operations team preserved each one-month-ahead prediction before the target month began. The actual quantities come from the shipment export after that month closed.

The first calculation is descriptive: total the actual units and compare the observed months. The second evaluates predictive work: compare each saved estimate with the outcome it was intended to predict.

Descriptively, shipments totaled 1,000 + 1,200 + 1,400 = 3,600 units. The monthly average was 3,600 / 3 = 1,200 units. September shipments exceeded July shipments by (1,400 − 1,000) / 1,000 = 40.0%. None of those calculations estimates October.

For the saved predictions, July's signed error was 1,000 − 1,100 = −100 units. August and September were above prediction by 50 and 100 units. The signed errors net to 50 units, but absolute errors total 250 units.

Mean absolute error = Sum of absolute prediction errors / Number of matched observations

Here mean absolute error is 250 / 3 = 83.3 units per month. Calculate from unrounded values; displayed averages and percentages are rounded to one decimal place. Forecasting: Principles and Practice explains why forecast evaluation must use observations unavailable when fitting the forecasting method, and defines mean absolute error from absolute errors.

Prove the comparison before interpreting it

Start with the actuals. The shipment rows included for July through September must total 3,600 units, using the same rules for canceled shipments and returns as the monthly source report. Do not compare gross shipments in one file with net shipments in the other.

Then match each prediction to the correct target month, product scope and unit. A prediction for orders booked cannot be evaluated against shipments without a documented bridge between the two. Preserve the source issue date so a later revision cannot quietly replace the estimate being evaluated.

The row check is actual units = predicted units + signed error. It holds in each month and in total: 3,600 = 3,550 + 50. Absolute errors must be summed separately; taking the absolute value of the net error would conceal offsetting misses.

A familiar business total can also be checked with a percentage:

Weighted absolute percentage error = Sum of absolute errors / Sum of actual quantities

The example gives 250 / 3,600 = 6.9%. The denominator is the actual volume across the same matched observations. See the forecast accuracy and bias calculation guide for a complete spreadsheet workflow rather than treating this short comparison as a forecasting manual.

Use history to challenge a prediction

The supplier now receives the following planning estimates. These are supplied inputs, separate from the closed shipment records.

The planning total is 1,500 + 1,550 + 1,600 = 4,650 units. It belongs in a capacity discussion as expected demand. It must not be added to the historical total and labeled actual shipments.

The manager should ask what supports the increase: open orders, changed customer commitments or a statistical estimate. A rising historical series alone does not establish that the increase will continue. NIST's introduction to exploratory data analysis emphasizes examining underlying structure, unusual observations and assumptions. Those checks help expose weak inputs before an estimate becomes a purchasing decision.

An analyst can also ask for lower and higher planning cases. Label such cases as assumptions unless their uncertainty has been statistically calibrated. A management sensitivity range is not automatically a prediction interval.

Decide which evidence belongs in the review

Use descriptive results when reviewing completed performance, reconciling reported sales or identifying where changes occurred. Compare regions or product lines on the same net basis, then investigate differences. The descriptive analytics tools comparison focuses on selecting a workflow for those reproducible summaries.

Use predictive results when the decision must be made before the outcome is known. Procurement may need an estimate of upcoming units, and operations may need to prepare capacity. The decision owner should see the estimate's issue date and assumptions, alongside the errors of comparable prior predictions.

Assign the actuals to their source owner, the estimate and its revisions to its forecasting owner, and the business action to the purchasing or operations manager. For a broader explanation of forecasting methods and uses, see predictive analytics. Choosing an action under constraints is covered separately in prescriptive analytics.

Where it goes wrong

• Replacing predictions after the fact. A revised estimate cannot fairly represent what the company knew when it committed inventory or capacity.
• Comparing different horizons. Keep one-month-ahead predictions separate from longer-horizon estimates when assessing performance. Their uncertainty and business uses differ.
• Letting errors cancel. A small net error can hide large misses in opposite directions. Retain signed and absolute errors.
• Mixing actuals and estimates. Keep unfinished periods visibly separate and exclude them from claims about closed-period results.
• Calling a target a forecast. A desired sales number expresses an objective. It becomes a prediction only when presented as an evidence-based expectation.

Compare your actuals and supplied predictions in Covirage

Bring the closed-period export and a file of dated predictions with matching periods and business identifiers. Covirage maps the columns, its tools compute and check the comparisons, and 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 after setting up your business.

See self-service analytics with Covirage for the supporting file-based workflow.