Predictive revenue: inputs, validation and forecast bias

Predictive revenue estimates future revenue using historical sales and signals available when the estimate is issued. Its business value depends on whether those estimates outperform a simple benchmark on later outcomes, with errors small enough for the decision being made. A close company total alone does not prove that the predictions are useful.

Define what the prediction means

Choose the outcome before choosing an approach. Bookings, invoiced sales, recognized revenue and cash collections answer different questions. A prediction of signed contracts cannot be validated against recognized revenue without explaining the timing difference.

Write down the forecast date, target period, reporting currency and segment definition. Also record the horizon: how far ahead the estimate was made. A prediction issued immediately before month-end has a different information advantage from one issued before the quarter begins.

Finance owns the accounting definition and actuals. Sales operations owns the dated pipeline snapshot and stage definitions. The person responsible for the forecasting method owns its assumptions, input history and version. Together they should agree which decision the estimate supports: hiring, spending, capacity or a revenue commitment.

Use inputs that existed at the forecast date

Historical revenue captures trend and seasonality. Pipeline amount, stage, age and expected closing date may help explain new business. Renewal schedules and contracted amounts matter for recurring business. Customer concentration matters when one delayed contract can change the period.

Statistical models fitted on your data can connect revenue with explanatory signals. But future inputs may themselves be uncertain: planned promotions are different from actual promotion results. The authors of Forecasting: Principles and Practice explain the distinction between predictions using information available beforehand and those using later observed inputs.

For the operational work of preserving opportunity information, see CRM forecasting. Adding more columns does not automatically improve a prediction; each column needs a business reason and a timestamp.

Test on later periods without leakage

Use earlier periods to develop the forecasting approach and later periods as a holdout. Select assumptions using development and validation periods, then leave the final holdout untouched until the approach is fixed. Repeatedly adjusting an approach after seeing its final test result turns that test into development work.

Recreate the reporting sequence. At each historical issue date, use only information available then, issue the prediction, and compare it with the later actual. This rolling evaluation is described in the authors' time series cross-validation guidance.

Leakage happens when an input reveals an outcome that would not have been available in practice. Examples include a deal's final closed status, a revised close date entered after the forecast date, or a customer segment assigned using later revenue. Randomly mixing earlier and later periods can also hide changing business conditions.

A complex approach may fit history closely and still perform poorly on new observations. That distinction is central to Google's explanation of overfitting. The validation task is to demonstrate usefulness outside the data used to develop the approach.

A worked holdout: the total hides the errors

A US business services company issued the following September revenue predictions before September began. The baseline uses the preceding comparable period's actual revenue. All amounts are USD; predictions and baseline amounts are supplied inputs, and actuals come from the closed revenue schedule.

Signed error = Prediction - Actual

Bias percentage = Sum of signed errors / Sum of actual revenue

WAPE = Sum of absolute errors / Sum of actual revenue

The total prediction equals actual revenue, so aggregate bias is 0 / 250,000 = 0%. Yet absolute segment errors sum to $40,000, giving WAPE of 40,000 / 250,000 = 16%. The baseline's WAPE is 35,000 / 250,000 = 14%. Both percentages are exact here.

The richer approach loses this comparison despite its perfect aggregate total. Enterprise is overpredicted while the other segments are underpredicted. That pattern could direct sales capacity toward the wrong customer group.

The proof is twofold: signed errors must equal total prediction minus total actual, while absolute errors must be calculated before summing. The actual revenue total must also reconcile to the closed revenue schedule. The spreadsheet guide to forecast accuracy and bias gives the cell formulas for a longer export.

Assess uncertainty and decide what to change

Score each horizon separately. Pooling near-term revisions with earlier predictions can make a weak planning estimate appear reliable. Repeat the segment comparison across enough historical issue dates to include normal seasonality and unusual periods.

Request prediction intervals with their stated coverage level when the forecasting method supplies them. Test how often actual outcomes fall inside those intervals and how wide the intervals are. Very wide ranges may cover outcomes but still be unhelpful for hiring decisions. Scenario ranges based on selected assumptions should be labeled as scenarios, rather than assigned an unsupported probability.

Use persistent segment bias to investigate stage definitions, renewal assumptions or changing customer behavior. Use large absolute errors to limit commitments that depend on precise revenue timing. The worked example of bias across reps and quarters shows why direction needs its own review.

Where it goes wrong

• Revised predictions scored as originals. Preserve the issued version and the inputs available on that date.
• Different revenue definitions. A bookings estimate and a recognized-revenue actual create a timing gap that is not automatically forecasting error.
• Cancellation hidden by totals. Sum absolute segment errors before computing a portfolio error measure.
• A changing customer mix. Check performance by segment as well as across the business, especially after a pricing or market change.
• Uncertainty treated as a promise. Keep assumptions, intervals and the decision's tolerance visible alongside the point estimate.

Validate your supplied predictions in Covirage

Bring a file or Google Sheet containing issued predictions, forecast dates, target periods, segments and matched actuals. Covirage maps the columns; its tools compute and check the errors and reconciliations, 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 export.

See forecast analysis for the supporting workflow.