Descriptive, Predictive, and Prescriptive Analytics: Differences and Examples

Descriptive, predictive and prescriptive analytics are three stages of the same journey: first you see what happened, then you estimate what is likely to happen, then you decide what to do about it. Understanding descriptive vs. predictive vs. prescriptive analytics helps finance and operations teams decide where to invest next. Finance and operations teams move through them in roughly that order, and each stage builds on the one before. Most companies are comfortable with the first, dabble in the second and rarely reach the third, even though the third is where the value of analytics shows most directly in the budget. This guide explains each type with finance and operations examples, adds the often-forgotten diagnostic stage, works through a trend forecast and an order quantity decision with checked arithmetic, and sets out how to move from one stage to the next. It belongs to business analytics for finance and operations, follows what is business analytics and leads to business analytics vs. business intelligence and data-driven decision making. The forecasts and decisions it produces feed budgeting and forecasting.

The stages at a glance

Some authors fold diagnostic into descriptive and speak of three stages, and others add more. The labels matter less than the habit of asking each question in turn.

Descriptive analytics

Descriptive analytics summarizes the past. Monthly sales, accounts receivable aging, headcount, inventory turns, on-time delivery and the income statement are all descriptive. It is the foundation: if the numbers are wrong or inconsistent, nothing built on them is reliable. Good descriptive analytics has agreed definitions, reconciled data and a regular schedule. Its limits are that it looks backward and tells you what, not why. A monthly report showing that gross margin fell from 38 percent to 36 percent is descriptive. It prompts the next question.

Diagnostic analytics

Diagnostic analytics breaks a result into causes. A standard method in finance is the variance bridge: split the change in margin into price, volume, mix and cost effects. Another is segmentation: compare customers, products or regions to see where the change is concentrated. If the two-point fall in margin came from a rise in freight costs on one lane and a shift in sales toward a lower-margin product, the response differs from the response to a general fall in prices. Diagnostic work is often where analysts add the most value, since it turns a number into a story that managers can act on. See budget variance analysis for the finance version.

Predictive analytics

Predictive analytics estimates what is likely to happen using patterns in past data. The methods range from simple (a moving average, a trend line) to complex (machine learning). The right choice depends on the amount of data, the stability of the pattern and the cost of errors. Simple methods often do as well as complex ones for stable products, and the extra complexity pays off mainly when there are many drivers, a lot of data and non-linear effects. See machine learning for forecasting and when to use machine learning vs. simple statistical models.

A worked trend forecast

Monthly unit sales for six months (hypothetical): 100, 104, 109, 111, 118, 121. Number the months 1 to 6.

• Mean of months: 3.5. Mean of sales: 663 / 6 = 110.5.
• Sum of products of deviations: (-2.5)(-10.5) + (-1.5)(-6.5) + (-0.5)(-1.5) + (0.5)(0.5) + (1.5)(7.5) + (2.5)(10.5) = 26.25 + 9.75 + 0.75 + 0.25 + 11.25 + 26.25 = 74.5.
• Sum of squared deviations of months: 6.25 + 2.25 + 0.25 + 0.25 + 2.25 + 6.25 = 17.5.
• Slope = 74.5 / 17.5 = 4.26 units per month.
• Intercept = 110.5 - 4.2571 × 3.5 = 95.6.
• Forecast for month 7 = 95.6 + 4.2571 × 7 = 125.4 units.

In Excel the same result comes from =FORECAST.LINEAR(7, sales_range, month_range). The trend line fits the data well, with sales rising by about four units a month. The forecast is a central estimate, and it needs an indication of uncertainty. The standard deviation of the forecast errors in the past is a practical measure. Suppose recent forecast errors have a standard deviation of about 10 units. A range of roughly 125 plus or minus 20 (about two standard deviations) would cover most outcomes, so the plan should be able to handle demand between about 105 and 145 units. See statistical forecasting in Excel.

Prescriptive analytics

Prescriptive analytics recommends an action. It combines predictions with costs, constraints and goals to find the best option. The classic example in operations is the order quantity for a product bought once for a season: too little stock loses margin on missed sales, too much leaves leftovers sold at a loss.

A worked order quantity example

Using the forecast, expected demand is 125 units with a standard deviation of 10, assumed to be roughly normally distributed. Each unit sold earns a margin of $12, so an unsold opportunity (understock) costs $12. Each leftover unit costs $4 after salvage (overstock cost).

• Critical ratio = understock cost / (understock cost + overstock cost) = 12 / (12 + 4) = 0.75
• The 75th percentile of demand corresponds to z = 0.674
• Order quantity = 125 + 0.674 × 10 = 131.7, about 132 units

The best order is above the expected demand because missing a sale is three times as costly as having a leftover (12 versus 4). If the margin were only $4, the critical ratio would be 0.5 and the best order would be exactly the expected 125. The analysis turns a forecast and two costs into a quantity, and shows how the recommendation changes when costs or uncertainty change. A rule of thumb that says "order the forecast" would be wrong here. Inventory planning tools use the same logic at scale. See demand planning and inventory optimization.

Moving from one stage to the next

Each stage depends on the one before.

1. Fix descriptive first. Agree definitions, reconcile to the accounts and automate the regular reports. Until then, a forecast built on the data would inherit its errors.
2. Add diagnostics to the monthly review. For every major variance, ask for the bridge. This builds the habit of looking for causes.
3. Start predictive with one high-value case. Demand, cash receipts or churn are common choices. Compare the model with a simple baseline, such as last year's number or a moving average. If it does not beat the baseline, do not use it.
4. Add prescriptive logic where decisions repeat. Order quantities, price points and allocation rules are repeated decisions where a rule based on costs and forecasts pays back quickly.
5. Close the loop. Compare predictions and decisions with outcomes, and use the errors to improve the models.

Where the stages show up in the finance function

Limits and cautions

Predictive models learn from the past and fail when conditions change in ways the past did not contain. They give estimates with uncertainty, and presenting a single number as certain invites trust it does not deserve. Prescriptive recommendations are only as good as the costs and constraints entered; a wrong assumption about the cost of understocking changes the answer. Also keep the human in the loop for decisions with large consequences. The model supports judgment, it does not replace it.

Common mistakes

• Jumping to predictive models before the data is reliable.
• Presenting forecasts without ranges.
• Never comparing a model with a simple baseline.
• Optimizing the wrong objective.
• Not tracking whether recommendations were followed and how they turned out.
• Treating the stages as a ladder where earlier ones can be abandoned.

One decision, four stages: setting a price

Take a product priced at $50 with a variable cost of $30 that sells 1,000 units a month. The same decision looks different at each stage.

• Descriptive: the product contributes (50 - 30) × 1,000 = $20,000 a month.
• Diagnostic: sales have been flat for a year, and the margin per unit is below that of similar products.
• Predictive: an earlier price test suggested an elasticity of about -1.2, so each 1 percent price rise reduces units by about 1.2 percent.
• Prescriptive: compare candidate prices using the predicted volumes.

Check: at $52.50 units fall 6 percent (1.2 × 5) to 940, and (52.50 - 30) × 940 = 21,150; at $55, units fall 12 percent to 880, and 25 × 880 = 22,000; at $57.50, 18 percent to 820, and 27.50 × 820 = 22,550; at $60, 24 percent to 760, and 30 × 760 = 22,800. The arithmetic says to keep raising the price, but the elasticity was estimated near the current price and is unlikely to hold at +20 percent. A sensible recommendation is to test $55, measure the real response and repeat. The example shows how each stage adds something: the descriptive stage sets the baseline, the predictive stage supplies the response and the prescriptive stage compares options. It also shows the limit of a model: it should not be trusted far outside the range of the data it was built on. See price elasticity of demand.

FAQs

What is the difference between descriptive, predictive and prescriptive analytics?

Descriptive analytics reports what happened, predictive estimates what is likely to happen and prescriptive recommends what to do, usually by combining predictions with costs and constraints.

Which type of analytics should a company start with?

Start with descriptive, so that numbers are reliable and consistent, then add diagnostics, then predictive and prescriptive analytics for the decisions with the greatest value.

Is machine learning required for predictive analytics?

No. Trend lines, seasonal averages and regression are predictive methods and often work well. Machine learning helps when there are many drivers and a lot of data.

What is an example of prescriptive analytics in finance or operations?

Choosing an order quantity by balancing the cost of understocking against the cost of overstocking, given a demand forecast and its uncertainty.

What is diagnostic analytics?

Analysis that explains why something happened, using drill-downs, segmentation and variance bridges.