Forecast Value Added (FVA)
Forecast Value Added (FVA) measures whether human intervention in the forecasting process — adjustments by Sales, Marketing, or Consensus — improves or degrades the quality of the forecast produced by the AI model (Baseline).
Each FVA variant compares a process stage (Sales / Marketing / Consensus) against the Baseline (AI) as the reference point.
Golden sign rule
| FVA value | Interpretation |
|---|---|
| Positive | The stage adds value: it improves the metric vs. the Baseline |
| ≈ 0 (dead zone ±0.05) | The stage is neutral: no meaningful improvement or degradation |
| Negative | The stage destroys value: it worsens the metric vs. the Baseline |
FVA differences between −0.05 and +0.05 are considered neutral: the statistical noise of small samples can produce variations in that range without a real underlying effect.
FVA variants
BIAS FVA
BIAS FVA = |BIAS_Baseline| − |BIAS_stage|
Compares the absolute bias of the Baseline against that of the stage. Absolute values are used because the direction of bias matters less than its total magnitude.
- Positive → "Corrects bias": the stage reduced systematic bias.
- Negative → the stage introduced or amplified bias.
Example: Baseline BIAS = −200 → |−200| = 200. Marketing BIAS = −50 → |−50| = 50. BIAS FVA = 200 − 50 = +150 (Marketing corrects bias).
MAE FVA
MAE FVA = MAE_Baseline − MAE_stage
Compares the mean absolute error of the Baseline against that of the stage.
- Positive → "Reduces error": the stage produced forecasts closer to actual demand.
- Negative → the stage moved the forecast further from actual demand.
Example: Baseline MAE = 150. Marketing MAE = 90. MAE FVA = 150 − 90 = +60 (Marketing reduces error).
ACCURACY FVA
ACCURACY FVA = Accuracy_stage − Accuracy_Baseline
The subtraction order is reversed compared to BIAS FVA and MAE FVA because in Accuracy a higher value is better. A positive result still means "the stage is better than the Baseline."
- Positive → "Improves accuracy": the stage increased accuracy in percentage points.
- Negative → the stage reduced accuracy.
Example: Baseline Accuracy = 82 %. Marketing Accuracy = 91 %. ACCURACY FVA = 91 − 82 = +9 pp.
Combined reading of all three variants
| BIAS FVA | MAE FVA | Diagnosis |
|---|---|---|
| Positive | Positive | Solid value add: corrects bias and reduces error |
| Negative | Negative | Destroys value: introduces bias and increases error |
| Positive | Negative | Noisy correction: fixes bias but disperses errors |
| Negative | Positive | Rare; may indicate a statistical artifact — review carefully |
Full example
SKU "Tablet Z" — Sales vs. Baseline — Actual demand: 1,000 units
| Metric | Baseline (AI) | Sales | FVA |
|---|---|---|---|
| Forecast | 1,150 | 1,040 | — |
| BIAS | −150 | −40 | +110 ✔ |
| MAE | 150 | 40 | +110 ✔ |
| Accuracy | 85 % | 96 % | +11 pp ✔ |
All three variants are positive: the Sales stage adds solid value over the AI Baseline for this SKU.
Review FVA by stage at the close of each version. A persistently negative FVA for a given stage is a signal that the manual adjustment process may be hurting forecast quality and warrants a process-level diagnosis.