MBA Knowledge Library · Business Analytics

Descriptive, Predictive, and Prescriptive Analytics

A dashboard can describe a problem without predicting what comes next or identifying a sensible response. Match the type of analysis to the decision, and keep assumptions visible as you move from observations to recommendations.

← Business Analytics hub · Browse all resources

Three questions, three analytical jobs

Descriptive analytics summarizes what happened. Predictive analytics estimates a future or otherwise unknown outcome. Prescriptive analytics compares actions and recommends a choice against an objective and constraints. A fourth category, diagnostic analytics, investigates why an outcome occurred. These categories can work together in one analysis.

Start with the decision and the data definition

Before choosing a model, write the decision, available actions, reporting period, unit, and consequence of being wrong. Define each measure. “Sales” might mean orders, invoices, recognized revenue, or collections; those definitions can produce very different conclusions.

Create a compact metric dictionary: name, business meaning, formula, source, cutoff, owner, and known exclusions. Check missing records, duplicate transactions, changing classifications, and selection effects. A more elaborate model does not repair an inconsistent denominator.

One worked example across all four types

A hypothetical support team wants to decide next week’s staffing. It has 480 complete requests waiting at the start of the current week, receives 600, and completes 560. There are no cancellations or transfers.

Descriptive: reconcile the backlog

Ending backlog = 480 + 600 − 560 = 520 requests. Backlog increased by 40, or 8.33% of the opening backlog. This describes the week; it does not establish why throughput fell short.

Diagnostic: investigate the gap

Break arrivals and completions down by request type, available staff hours, system downtime, rework, and missing-information status. If complex requests increased, compare output within comparable work groups. A drop in requests per employee could reflect more difficult work rather than weaker performance.

State alternative explanations and seek evidence. A chart showing high backlog on busy days does not prove that adding employees is the only effective intervention. Process changes, better intake, or exception routing may also matter.

Predictive: estimate next week’s arrivals

The last three comparable weeks had 540, 600, and 570 arrivals. A simple three-week moving average forecasts (540 + 600 + 570) ÷ 3 = 570 arrivals. For planning, the team also evaluates 510 and 630 arrivals as illustrative lower and higher scenarios. These are chosen scenarios, not a statistically estimated confidence interval.

With 520 opening backlog and 560 planned completions, the base-case ending backlog is 520 + 570 − 560 = 530. The scenarios produce 470 and 590. A forecast is conditional on factors such as comparable demand, working days, staffing, and request mix.

Prescriptive: compare feasible staffing choices

Suppose the manager wants ending backlog at or below 500. Each temporary analyst can complete 40 additional acceptable requests during the week for an incremental cost of $1,000. At most two are available. Existing capacity remains 560; onboarding, quality review, and work mix are assumed not to reduce that capacity in this simplified model.

Under the base case, one analyst is the lowest-cost listed option meeting the 500-backlog target. Under the high case, even two analysts leave 510. The manager must consider another feasible action, a revised service commitment, or explicit acceptance of the shortfall. The model makes the tradeoff visible; it does not authorize a staffing decision.

Check a forecast before trusting it

1. Use a baseline: compare the model with a simple recent-value or seasonal benchmark.
2. Separate training and evaluation periods: evaluate on observations not used to fit the model.
3. Respect time: do not let information unavailable at the forecast date leak into the inputs.
4. Measure errors: inspect direction, magnitude, and performance by segment.
5. Monitor after launch: changes in behavior, processes, or data can make the model less useful.

For an original error illustration, forecasts of 100, 120, and 110 compared with actuals of 90, 130, and 100 have absolute errors of 10, 10, and 10. Mean absolute error is 30 ÷ 3 = 10 units. That does not mean every future error will be 10. Compare error to the cost of the business decision, not just a model score.

Turn analysis into a useful recommendation

A brief should state: the decision, evidence, plausible explanations, forecast assumptions, options, tradeoffs, recommended action, owner, and review date. Separate an observed fact from an estimate or judgment. “Backlog ended at 520” is an observation; “arrivals may be 570” is a forecast; “add one analyst” is a conditional recommendation.

Common mistakes

• Calling every dashboard “predictive.”
• Presenting a scenario range as a confidence interval.
• Claiming causation from a correlation.
• Optimizing cost while omitting quality or service constraints.
• Measuring accuracy on the same observations used to fit the model.
• Reporting a precise output without exposing uncertain inputs.

Frequently asked questions

Apply this framework

Choose one recurring decision. Write one descriptive measure, one diagnostic question, one simple forecast, and two possible actions. Identify the assumptions that could reverse the recommendation, then set a date to compare the forecast and action with actual results.

Sources and example notes

• IBM: Predictive Analytics
• IBM: Prescriptive Analytics

Conceptual references are linked above. Examples, scorecards, calculations, checklists, and practical recommendations are original educational illustrations. They are not research findings, vendor endorsements, or universal benchmarks. Adapt decisions to actual evidence, policies, and obligations.