FVA: Calibrating the Freedom to Adjust the Forecast

Every mature forecasting process passes through many hands: statistics, demand planner, sales, marketing, finance, S&OP. Each link adjusts the previous number with its own reading of the world. And in almost every company, no one measures who improved and who worsened the forecast.

The outcome is familiar: consensus becomes a political auction, where the loudest area in the room moves the number the most. Accuracy drops, but accountability dissolves. FVA exists precisely to make that chain auditable.

FVA is a process metric, not a model metric. It compares the forecast error at each stage of the adjustment chain against the error of the immediately preceding stage. The question is simple: did this intervention reduce or increase the average error?

The base reference is usually an automatic statistical forecast, the baseline. From there, each additional layer (planner adjustment, commercial override, S&OP alignment) is measured against the previous step and against the baseline itself, producing a clear trail of added or destroyed value.

To make the metric auditable, it must be anchored against what each stage is being compared to. Consolidated practice organizes the process in four steps, each measured against the immediately preceding step.

The 12-month moving average is the concrete example of a naive baseline. Seeing consensus lose to that average is the harshest alert the method delivers: it means the entire adjustment structure, meetings and reviews cost money and worsened the forecast.

The calculation is grounded in MAPE, or another consistent error metric, computed by SKU, region or family, across closed forecasting cycles. For each adjustment layer, the version of the forecast is recorded, and at the end of the horizon it is compared with actual sales.

Positive values indicate the intervention added value. Negative values indicate the adjustment worsened accuracy. Aggregated readings by area and participant reveal consistent patterns across many cycles, not isolated cases.

After a few cycles, the FVA panel naturally separates three profiles within the process:

The sensitive point: in many operations, the statistical baseline alone beats the consensus forecast. When this shows up recurrently, the adjustment process is costing the company money.

Collaboration has a cost. It mobilizes people, meetings and time across several areas to revise numbers. This cost is rarely accounted for when the consensus process is defended as an end in itself.

For highly predictable items, with low coefficient of variation, the theoretical accuracy of the statistical baseline is already high. The margin for collaboration to improve is small, and often negative. Forecastability is the accuracy ceiling attainable given the demand pattern: stable items have a high ceiling and little room for manual gain.

Crossing predictability with item importance makes the decision obvious. Low volume and low value do not justify mobilizing a committee to move the number. The ABC-XYZ matrix defines where collaboration is investment and where it is dead cost.

Not every item deserves a meeting. Collaborating on an item the model already gets right is not rigor. It is waste. The decision to collaborate should be segmented: concentrate human effort where potential FVA is high (erratic items, high value, launches, promotions) and let the statistical engine run alone where collaboration does not pay for itself.

FVA is not about punishing people, it is about calibrating the freedom to adjust within the process. The rule is direct: those who historically add value keep autonomy to intervene. Those who historically destroy value must provide explicit justification, additional approval, or have their adjustment range restricted.

This changes the nature of the consensus meeting. Instead of debating feelings, evidence is discussed: what was the area's average FVA across recent cycles, which overrides worked, which turned into noise. S&OP stops being an opinion forum and becomes a data-driven decision forum.

Deploying FVA is more a process decision than a technology one. It requires versioning the forecast at each stage, defining measurement points, publishing results transparently and reviewing the autonomy rules on a regular cadence. Quarterly usually works well.

In NPLAN's Demand modules, each adjustment layer is recorded with author, timestamp and delta against the baseline. FVA stops being a manual spreadsheet exercise and becomes an operational indicator, generated automatically at every forecasting cycle.