Forecast value added is a process audit

Forecast value added, or FVA, is most useful as a process audit. It asks whether a change to a baseline improved the forecast under a stated error measure. It does not ask whether a person is good at forecasting, and it does not turn an override into a moral score.

The minimum record has three numbers: the system or statistical baseline, the adjusted or final forecast, and the actual outcome. Without the first number, the team can score the final forecast but cannot tell whether the intervention added information or merely moved the error.

Why does Forecast Value Added begin with a statistical baseline?

Lawrence, O’Connor, and Edmundson studied judgmental sales forecasting in a field setting across manufacturing organizations. Their database contained around 24,000 forecasts and 4,500 actual sales values. The company forecasts were not uniformly more accurate than a simple naive forecast, defined as the most recent actual value. The authors also identify bias and serial correlation in the errors.

The lesson is not that every business should use a naive forecast. It is that a forecast comparison needs a declared comparator. A baseline can be a statistical model, a naive method, a prior approved forecast, or another method appropriate to the series. The choice changes the meaning of FVA and must be frozen before the intervention is judged.

Mantrala and Raman make a separate governance boundary visible. Their analytical treatment of the Gonik incentive scheme links bonus income to a management quota, a salesperson’s submitted forecast, and actual sales. Under a stochastic one-period model, the plan parameters can elicit a selected probability of fulfillment while preserving the incentive to sell. It is not field evidence of an accuracy gain. It clarifies why a forecast and a quota should remain separate objects: if the number is only a target, an FVA audit cannot tell whether it carried information or expressed an incentive.

Separate accuracy from bias

Fildes, Goodwin, and De Baets bring the measurement discipline up to date. Their analysis combines 147,131 forecasts and actuals from 10 organizations and 22 business units across six datasets. They compare system forecasts with judgmentally adjusted final forecasts and analyze accuracy and bias as separate dimensions.

The reported overall medians are deliberately modest. Across the company and business-unit summaries, 51.5% of SKUs improved on the reported FVA measure and 55.6% improved on the reported bias measure. Those figures are not a pooled pass rate for every forecasting team. They are medians across varied datasets. Accuracy and bias can also move in opposite directions, so one score cannot stand in for both.

Direction is evidence, not a policy

Across the six datasets, the 2025 study finds that downward adjustments generally improve accuracy and bias, while upward adjustments have a mixed record and often reduce accuracy. The result is useful because it makes direction worth recording. It is not a license to approve every downward change or ban every upward one.

The same study reports variation across datasets and across companies or business units. Its analysis uses one-step-ahead weekly or monthly SKU forecasts, not discrete B2B opportunities. A sales pipeline may contain private information that a statistical system cannot see. A deal also does not have the same repeated time-series structure as a stock keeping unit.

So the valid transfer is a design hypothesis: preserve the baseline, record the intervention, and test the result. The invalid transfer is to copy a directional rule into a commercial forecast and call it empirical proof. The existing number you call piece is useful as a prompt for that question. The defensible evidence review sets the stopping rule for deciding when the repeated comparison has earned another cycle. The FVA audit is the evidence boundary that keeps the prompt from becoming a universal sales claim.

Run the audit before changing the process

Use a repeated window rather than a single memorable miss:

1. Freeze the baseline at the moment the forecast is issued.
2. Record every intervention separately, including no-change cases where the process allows them.
3. Preserve direction, size, timing, reason, author, and supporting information.
4. Define the actual outcome window before the result is known.
5. Compare accuracy and bias separately, then segment by horizon, product, business unit, and process.
6. Add review time and system effort before deciding whether the difference earned its cost.

The process should also retain the forecast that would have reached the decision without the intervention. A final number alone cannot show whether the human contribution helped. A baseline and an actual alone cannot show what the intervention changed.

What this evidence does not prove

The 2025 evidence does not observe every organizational process or motivation behind an adjustment. It is limited to one-period-ahead demand forecasts and its company results vary. The 2000 field study is a manufacturing forecasting context. Neither study validates a B2B revenue-forecast override protocol.

FVA also does not prove that a more accurate forecast produced a better business decision. Inventory, capacity, service, and commercial choices may value different errors. State the decision purpose and the error measure before interpreting the result.

Boundary

Forecast value added is a disciplined comparison, not a verdict on judgment. It becomes useful when the team can show the untouched baseline, the intervention, the actual outcome, the metric, and the cost of the process over repeated periods. Only then can the organization decide which interventions earned their effort and which ones only moved accountability.

Evidence base. The analytical frame also draws on these additional sources: Mantrala and Raman 1990. The links identify the exact works; they support the mechanisms and boundary conditions discussed here, not every claim in isolation.