What forecast value added measures
An accuracy figure on its own says little. A WAPE of 4.6% is excellent for a volatile new product and poor for a stable staple. FVA asks a sharper question: did this step of the process improve the forecast, compared with doing nothing or with the step before? The idea was popularised by Michael Gilliland of SAS in The Business Forecasting Deal. This guide and the Tabellab course are independent and not affiliated with SAS.
- FVA > 0: the step reduced the error and added value.
- FVA < 0: the step made the forecast worse. The company spent time and got a less accurate number.
Each process step gets its own forecast version: the naive forecast, the statistical forecast, the planner override, the consensus from the S&OP meeting. Keep every version, not only the final number, or FVA cannot be measured.
Step 1: build the naive forecast
The naive forecast is the simplest forecast possible: next month = this month's actual. In the grid, C3 is =B2, filled down. For a seasonal item, use the seasonal naive instead, the actual of the same month last year. Whatever you choose, it must cost almost nothing to produce, because it is the placebo every other step has to beat.
Compare only the periods that every version covers. December 2024 has an actual but no naive forecast, so it stays out of every sum. All errors in the example run over January to December 2025.
Step 2: measure every version with the same metric
Use one error metric for all steps. The example uses WAPE, the sum of absolute errors divided by the sum of actuals: =SUMPRODUCT(ABS(D3:D14-$B$3:$B$14))/SUM($B$3:$B$14). With the actual column locked, the same formula copies to the right for every forecast version.
Over the year the SKU sold 67,000 units. The absolute errors add up to 4,900 units for the naive forecast, 2,750 for the statistical forecast and 3,100 for the consensus, which gives the WAPEs of 7.3%, 4.1% and 4.6%. WAPE is safer than MAPE when some periods or items are small, as the MAPE guide shows.
Worked example: an FVA stairstep for one SKU over 12 months
Monthly actuals of one fictional household-care SKU and three forecast versions (yellow inputs): the naive forecast (previous month's actual), the statistical forecast and the consensus forecast after the S&OP meeting. Rows 16 to 21 calculate WAPE, accuracy, FVA against the naive forecast and against the previous step, bias and the number of months each step beat the naive forecast.
| A | B | C | D | E | |
|---|---|---|---|---|---|
| 1 | Month | Actual | Naive | Statistical | Consensus |
| 2 | Dec 2024 | 4,800 | |||
| 3 | Jan 2025 | 5,100 | 4,800 | 4,950 | 5,300 |
| 4 | Feb 2025 | 4,700 | 5,100 | 4,900 | 5,000 |
| 5 | Mar 2025 | 5,300 | 4,700 | 5,150 | 5,200 |
| 6 | Apr 2025 | 5,000 | 5,300 | 5,350 | 5,500 |
| 7 | May 2025 | 5,600 | 5,000 | 5,350 | 5,700 |
| 8 | Jun 2025 | 5,900 | 5,600 | 5,700 | 6,100 |
| 9 | Jul 2025 | 5,400 | 5,900 | 5,600 | 5,900 |
| 10 | Aug 2025 | 5,200 | 5,400 | 5,450 | 5,500 |
| 11 | Sep 2025 | 5,800 | 5,200 | 5,600 | 5,800 |
| 12 | Oct 2025 | 6,100 | 5,800 | 5,950 | 6,200 |
| 13 | Nov 2025 | 6,600 | 6,100 | 6,300 | 7,200 |
| 14 | Dec 2025 | 6,300 | 6,600 | 6,650 | 6,500 |
| 15 | |||||
| 16 | WAPE (Jan–Dec 2025) | 7.3% | 4.1% | 4.6% | |
| 17 | Accuracy (1 − WAPE) | 92.7% | 95.9% | 95.4% | |
| 18 | FVA vs naive (points) | 3.2% | 2.7% | ||
| 19 | FVA vs previous step | 3.2% | -0.5% | ||
| 20 | Bias (forecast − actual) ÷ actual | -2.2% | -0.1% | 4.3% | |
| 21 | Months the step beat naive | 9 | 8 | ||
- C3
=B2Naive forecast: previous month's actual - D16
=SUMPRODUCT(ABS(D3:D14-$B$3:$B$14))/SUM($B$3:$B$14)WAPE of the statistical forecast - D18
=$C$16-D16FVA of the statistical forecast vs naive - E19
=D16-E16FVA of the consensus vs the statistical forecast - E20
=SUMPRODUCT(E3:E14-$B$3:$B$14)/SUM($B$3:$B$14)Bias of the consensus - D21
=SUMPRODUCT(--(ABS(D3:D14-$B$3:$B$14)<ABS($C$3:$C$14-$B$3:$B$14)))Months the statistical forecast beat naive
Fictional, illustrative data for one SKU. FVA is shown in percentage points of WAPE. The FVA concept is credited to Michael Gilliland of SAS. Tabellab is not affiliated with SAS.
Step 3: the FVA stairstep
The classic FVA report lists the steps in process order with two FVA columns:
- FVA vs naive = WAPE naive − WAPE of the step: statistical +3.2 points, consensus +2.7 points.
- FVA vs the previous step = WAPE of the step before − WAPE of this step: statistical vs naive +3.2 points, consensus vs statistical −0.5 points.
Both columns matter. Judged only against the naive forecast, the consensus looks fine, because it still beats doing nothing. The step-by-step column shows that the meeting took a 4.1% forecast and made it 4.6%. The value came from the statistical model, and the last step gave part of it back.
Look at bias, not only accuracy
Bias shows the direction of the error: =SUMPRODUCT(E3:E14-$B$3:$B$14)/SUM($B$3:$B$14). The statistical forecast is almost unbiased at −0.1%. The consensus is +4.3%: it was above the actual in 10 of the 12 months and never far below. The biggest miss is November, when the promotion forecast of 7,200 units met an actual of 6,600.
This is the most common FVA finding in practice. Overrides and consensus meetings tend to add optimism, especially around promotions and new listings. A tracking signal or a bias limit per step catches it early.
How to act on the result
Count how often each step beats the naive forecast month by month: the statistical forecast does in 9 of 12 months, the consensus in 8. Then decide by evidence, not by job title:
- Measure FVA at the level you forecast (SKU or SKU × location) and over enough periods, at least a few months, before drawing conclusions.
- Allow overrides only above a size threshold or with a documented reason, and track the FVA of each override type.
- A negative FVA is not a reason to blame planners. It is a reason to change the process or drop a step that costs time and adds error.
Common mistakes
- Comparing different periods. Including December 2024 for the statistical forecast but not for the naive one gives each step a different test. Use only the months every version covers.
- A different metric per step. MAPE for one step and WAPE for another cannot be subtracted. One metric, one formula, copied across.
- Only FVA vs naive. The consensus beats naive by 2.7 points and still destroys 0.5 points compared with the statistical forecast. Always show the step-by-step column too.
- Keeping only the final forecast. Without the saved statistical and override versions, FVA cannot be calculated afterwards. Store every version every cycle.
- Judging on one month. A single month can go either way by luck. Look at a full year or a rolling window, and at how many months a step won.
- Ignoring bias. A step with a slightly worse WAPE and a strong upward bias costs more than the WAPE suggests, because it builds excess stock.
Frequently asked questions
What is forecast value added?
Forecast value added (FVA) is the change in a forecast error metric caused by one step of the forecasting process, such as a statistical model, a planner override or a consensus meeting. A positive FVA means the step made the forecast more accurate.
How do you calculate FVA?
Measure every forecast version with the same metric over the same periods, then subtract: FVA = error of the earlier forecast − error of this step. With WAPE of 7.3% for the naive forecast and 4.1% for the statistical forecast, the FVA of the model is +3.2 points.
What is a naive forecast?
A naive forecast uses the last actual as the next forecast, or the same period last year for seasonal items. It costs almost nothing, so every forecasting step should be more accurate than it.
What is a good FVA?
Any step with a positive FVA adds value, and steps with a negative FVA should be changed or removed. There is no universal target, because the possible gain depends on how forecastable the demand is.