Forecast Value Add (FVA): The Only Forecasting Metric That Settles Arguments
Forecast Value Add (FVA) measures whether each step in your forecasting process, the statistical model, the planner’s override, sales input, the S&OP consensus, made the forecast better or worse than the step before it. It is the only forecasting metric that answers the question everyone actually argues about: is all this touching helping? Most organizations that measure it honestly for the first time find that a large share of their manual adjustments subtract value. That discovery changes forecasting culture in a way no algorithm upgrade ever has.
What is Forecast Value Add, exactly?
Score every stage of the forecast against the same actuals, and score a naive baseline (last period, or seasonal naive) alongside them. The FVA of a stage is its accuracy minus the accuracy of the input it received. Positive means the stage earns its keep. Negative means you are paying people to make the forecast worse, month after month, with everyone’s full approval. That is the standard definition used by the Institute of Business Forecasting and in Lokad’s write-up of the method, and the naive benchmark is the part people leave out most often.
We score the stages with WAPE, because a metric that weights by volume is the only fair way to compare a stage that touched thirty SKUs with one that touched three thousand. If you want the argument for that choice in full, we made it in MAPE vs WMAPE vs WAPE. The formula itself is standard, and Rob Hyndman’s note on WAPE is the short version.
A worked FVA stairstep
The numbers below are a worked example built to show the shape of a real result, not a client data set. Every stage is scored against the same actuals using WAPE at the same operational lag, which is the only way the stages stay comparable.
| Process stage | WAPE | FVA vs the stage before | What it tells you |
|---|---|---|---|
| Seasonal naive baseline | 38% | baseline | The number every other stage has to beat to justify existing |
| Statistical model | 26% | +12 pts | The model earns its licence fee several times over |
| Planner override | 23% | +3 pts | Small and positive, which is what a healthy override layer looks like |
| Sales input | 29% | -6 pts | The step that pays for itself in nothing. Optimism arriving as a number |
| S&OP consensus | 25% | +4 pts | Consensus repairs part of the damage, then stops |
Read the last column down and you have the finding that starts the difficult meeting. The process ends at 25% WAPE having already passed through 23% two stages earlier, so everything after the planner override is net negative. The forecast was better before three senior people touched it.
How do you run an FVA analysis without new software?
- Log the forecast at each stage for three to six months. Most planning tools already store this, and a spreadsheet handles it below a few hundred SKUs.
- Score every stage with the same metric at the same lag. WAPE, at your operational lag.
- Compare each stage to its input, and everything to naive.
- Publish the stairstep. Then hold on, because the meeting that follows is a difficult one.
Nothing in that list needs a licence. The reason FVA gets sold as a software feature is that vendors need a reason to charge for it, not that the arithmetic is hard.
How do you read a negative FVA result?
A negative-FVA sales override is not a reason to humiliate sales. It is a reason to ask what sales knows that the number does not capture, such as promotions or distribution gains, and whether that knowledge can enter the process as data instead of opinion. The override often carries one real signal wrapped in ten points of optimism, and the fix is extracting the signal rather than deleting the input. Add a de minimis rule while you are at it. Adjustments below a threshold are not worth their meeting time even when they are positive.
The pattern is not unique to one company either. The Foresight reality check on FVA is the standard reference for how often judgmental adjustments fail to beat the input they were applied to.
Why is FVA getting attention now?
FVA is the new kid on the block, and it turned up at the moment the arithmetic got cheap. Compute that used to need a project sponsor is close to a rounding error now, so evaluations nobody thought were worth the trouble have become an afternoon’s work. The method itself has barely changed. The cost of running it has.
Underneath that sits a bigger shift. Forecasts have always been wrong, and for decades the accepted answer was to pay for the error somewhere else: cash tied up in inventory, longer wait times, service levels that slipped by agreement. That trade held because there was no better option going. There is now. Neural networks put accuracy in reach that we would not have promised a few years ago, and a model can filter historical noise on its own rather than a planner deciding by hand which weeks counted as unusual. We would still read the vendor numbers with a hand on the brake, for the reasons we set out in AI demand forecasting: what the 20 to 50% better claims leave out.
So the planner’s contribution has to justify itself against a stronger baseline than it used to. That is the whole point of the metric. When someone adjusts a machine-generated forecast, is the adjustment worth anything, or is it a bias the business pays for two quarters later?
Does FVA only judge the planner?
No, and reading it that way throws away half the output.
Run the analysis, then root-cause it, and planners get to see the behaviour behind their own bias. That is a change-management problem far more than an analytics one. People adjust for reasons, usually ones that made sense in a previous job or under a demand pattern that has since moved, and they stop when they can see the pattern in their own numbers instead of being told about it in a meeting.
The other half of the output points at the model. Where a machine forecast keeps getting beaten by the humans downstream, you are usually looking at gaps in the training data, a gap between what the data says and what is happening on the ground, or an input feed that broke its own standard operating procedure months ago without anyone noticing. All three have systemic fixes. None of them gets fixed by asking planners to try harder.
How do FVA and forecast bias fit together?
FVA and bias analysis are the same investigation from two angles. FVA finds which stage hurts. Bias analysis finds which direction, and usually whose incentive. Run together, they turn “whose forecast is right”, an argument nobody wins, into “which process steps earn their cost”, which has answers. The bias half of that pair is in how to actually reduce forecast bias, and if the stairstep above points at the consensus meeting rather than any one stage, the problem is probably the one we described in why S&OP fails.
