Forecast Value Added: Does Each Step Earn Its Keep?

A four-step forecasting process (Naive → Statistical model → Analyst override → Executive consensus), scored on portfolio MAPE across a simulated 50-SKU manufacturer. FVA is a placebo test: every step is compared to the dirt-simple naive benchmark. Positive FVA = lower error = value added. Move the sliders to see who adds value and who destroys it.

Net process FVA (final vs naive)
−0.6 pp

The four-step pipeline

STEP 1
Naive
Seasonal naive — copy the same month last year. The benchmark everyone must beat.
14.6% MAPE
benchmark
STEP 2
Statistical
STL + ETS on 3 years of history. The real model: trend + robust seasonality.
12.3% MAPE
STEP 3
Analyst
Demand-planner judgmental overrides — small, mostly-noise tweaks.
13.6% MAPE
STEP 4
Executive
S&OP consensus — systematic upward optimism. The final number.
15.3% MAPE

Accuracy stairstep (MAPE by step)

Lower is better. The dashed line is the naive benchmark — anything above it is worse than doing nothing.

FVA waterfall (value added / destroyed)

Green = the step removed error (added value). Red = the step added error (destroyed value).

% of SKUs beating naive (baseline simulation, 50 SKUs — reference, not driven by sliders)

Statistical
78%
Analyst
64%
Executive
58%
The dark tick marks the 50% coin-flip line. Each human step slides the portfolio closer to it.

Advisory

Active drivers

    Recommendations

      References & data

      1. Gilliland, M. (2010). The Business Forecasting Deal. Wiley & SAS Business Series. ISBN 978-0-470-57443-0. (Origin of FVA as a standard metric.)
      2. Gilliland, M., SAS Institute — Forecast Value Added Analysis: Step-by-Step (white paper).
      3. Morlidge, S. (2013), “How Good Is a ‘Good’ Forecast?” Foresight, Issue 30. Study of 8 businesses: 52% of forecasts failed to beat naive (via Petropoulos et al. 2022, IJF).
      4. Fildes, Goodwin, Lawrence & Nikolopoulos (2009), “Effective forecasting and judgmental adjustments,” IJF 25(1), 3–23. 60,000+ forecasts, 4 companies: small adjustments often hurt; large downward ones more often help.
      Baseline figures from a reproducible simulation (50 SKUs × 48 months, set.seed(42), STL+ETS via R forecast). Synthetic data, not a real company. The what-if sliders are a teaching model calibrated so the default state reproduces the simulated baseline exactly.