Predictive analytics: what it is, with examples from finance and sales

Predictive analytics uses statistical models fitted on historical data to estimate a value you do not know yet: next quarter's revenue, which customers will stop buying, which invoices will be paid late. A prediction is only worth the error it shows on data it was not fitted on. In the worked example below, a straight trend line misses two unseen quarters by 1.26% on average, against 6.75% for simply repeating the last quarter.

What predictive analytics is

Analytics is usually split into four kinds, by the question each answers.

Predictive analytics is the third row. It takes the patterns in past data, fits a method to them, and applies that method to cases or periods where the answer is not known yet.

Examples in finance and sales

• Revenue and demand. Next month's or next quarter's revenue, by product or region, from the history of the same series. See four sales forecasting methods compared.
• Churn propensity. The chance each customer stops buying, from order frequency, recency, the trend in spend and support history. B2B churn without a subscription covers how to define it when there is no contract end date.
• Late payment. The chance each open invoice is paid after its due date, from the customer's payment history, invoice size and terms.
• Win probability. The chance each open opportunity closes, from stage, age, deal size and the rep's history.
• Cash collection timing. When each receivable is likely to turn into cash, which feeds a cash flow forecast.

The methods

Regression. Fits a line, or a surface, through past data so one value can be estimated from others. A straight trend over time is the simplest case: revenue = intercept + slope × quarter number. NIST's handbook describes the idea as splitting the variation in one quantity into a deterministic function of other quantities plus a random part.

Time-series methods. Use the series' own past. Exponential smoothing weights recent periods more than old ones and can carry trend and seasonality; ARIMA describes how each period depends on earlier periods and earlier errors. Both are covered in Hyndman and Athanasopoulos, Forecasting: Principles and Practice.

Classification. Estimates the chance of a yes-or-no outcome: churns or not, pays late or not, wins or not. Logistic regression gives a probability from a weighted sum of inputs; decision trees split the rows by simple rules; gradient-boosted trees combine many small trees, each correcting the last.

Worked: a trend forecast tested on quarters it has not seen

Eight quarters of revenue, in millions of dollars. Fit the trend on Q1-Q6 only and hold Q7 and Q8 back as the test.

On Q1-Q6, the fitted trend has a slope of 0.1851 a quarter and an intercept of 3.9287:

Trend = a + b × t, with b = SLOPE(y, t) and a = INTERCEPT(y, t)

So Q7 is 3.9287 + 0.1851 × 7 = 5.225 and Q8 is 5.410. With quarter numbers 1-8 in A2:A9 and revenue in B2:B9, the forecast for Q7 in C8 is:

=FORECAST.LINEAR(7,B2:B7,A2:A7)

which returns 5.225. Microsoft describes FORECAST.LINEAR as predicting a future value "by using linear regression". The error per quarter in D8 is:

=ABS(B8-C8)/B8

and MAPE is the average of those errors:

MAPE = average of |actual − forecast| / actual

The trend misses Q7 by 1.42% and Q8 by 1.10%, a MAPE of 1.26%. The naive baseline, which repeats Q6's 5.02, misses by 5.28% and 8.23%, a MAPE of 6.75%. The trend beats the baseline on quarters it never saw, so it has earned its place. Only then refit on all eight quarters: slope 0.1946, intercept 3.9029, giving Q9 = 5.655 and Q10 = 5.849.

How to judge a prediction

Hold-out test. Fit on the earlier data, test on the later data, and report the error on the test only. Hyndman and Athanasopoulos put it plainly: "A model which fits the training data well will not necessarily forecast well."

A naive baseline. Compare against "same as last period". The same book treats simple methods as benchmarks: if a new method does not beat them, "the new method is not worth considering."

An error measure people can read. MAPE is a percentage, so it compares across series of different size. How to measure forecast accuracy and bias in Excel adds bias, the tendency to miss in one direction.

A range, not a point. Report a prediction with the range it is likely to fall in. In Excel, FORECAST.ETS.CONFINT returns the half-width of a confidence interval for a FORECAST.ETS forecast.

Predictive analytics tools

Neutral, and in rough order of effort:

• Spreadsheets. FORECAST.LINEAR for a straight trend, FORECAST.ETS for seasonal series (=FORECAST.ETS(target_date,values,timeline)), and LINEST for regression with several inputs.
• BI and data platforms. Many include forecasting and trend lines on charts, run against the data already in the warehouse.
• Python and R. Full control over methods, testing and automation, for teams that can maintain code.
• Specialist tools. Built for one job, such as demand planning or churn scoring.

The right choice depends on how much data there is, how many series, and who will keep the work running after the first version.

Where AI fits

Statistical models fitted on your data make the prediction, and they can be tested for error on periods they have not seen. A language model can explain the result in plain words: which quarters drove the trend, why the range is wide, what would change the estimate. It should not produce the numbers, because its output cannot be tested the same way and does not show the rows behind it. Why the language model must never do the arithmetic sets out the reasoning.

Where it goes wrong

• Testing on the fitted data. A method always looks good on the data it was fitted on. Report only the hold-out error.
• No naive baseline. A method that cannot beat "same as last period" adds nothing.
• Leakage. Using information not known at prediction time, such as the payment date when predicting late payment, gives a test score that will never be repeated in use.
• Extending a trend through a structural break. A lost customer or a price change starts a new series; the old trend does not carry across it.
• A single number without a range. It gets read as a promise.

Predictions you can check

Covirage fits statistical models on your own history with its tools, tests them against a naive baseline on periods they have not seen, and reports the error beside every prediction; the external AI model explains the result and never produces the numbers. See AI analytics for how the two parts divide the work, and what is a good forecast accuracy for what error to accept. For the Excel functions in detail, see the FORECAST function in Excel.