In a study of more than 300,000 forecasts, 52% lose to a random walk—the plainest test there is, which simply carries last period’s value forward unchanged [3]. That is the number I would put in front of a planning team about to renew a demand-sensing budget. The question is not how clever the extra layer is, but whether its value added is positive.

The big picture:

A planning stack is a pile of layers. You have a model’s baseline, sales input, a team consensus pass, and a sensing tool that re-forecasts inside the month. Vendors sell each layer on accuracy, but operators rarely score them against the layer below.

What strikes me is how plain the missing test is. Forecast value added (FVA) is the change in error that one step of the process brings compared to a baseline test [3]. The usual test is the simplest one around—last period’s value, or the same week a year ago. When a step’s error is worse than that baseline, the step extracts value rather than adding it.

Scoring steps is rare work. In a May 2026 survey of 661 planners, 16% did not formally measure forecast accuracy at all [5]. Another 51% said they sometimes go ahead with a forecast they know is unreliable, because no better option exists [5].

By the numbers

  • 52% — Worse than random: In a study of more than 300,000 forecasts, reported in a practitioner account of Gilliland’s Business Forecasting, 52% did worse than carrying the last value forward [3].
  • 147,000 — Forecasts reviewed: Edits to a model’s baseline helped just over half of stock keeping units (SKUs) across six datasets, and edits that pushed the number up were the most likely to hurt [1].
  • 60,000 — Supply-chain forecasts: Edits raised accuracy on average at three of four firms studied, with large edits helping and small ones often causing harm [2].
  • 13.8 points — Consensus layer loss: In a worked case cited from Gilliland’s 2010 book, a consensus forecast lost to its naive test in 11 of 13 weeks, carrying an average error of 13.8 points [4].

What I’d watch:

The part I keep circling is what changes once a team scores its own process. The moves I am watching all sit downstream of that one number.

  • The frozen lag: A score only matters at the point where the plan is locked. For the baseline, that is months out. For a sensing layer, it is the nearest one to four weeks when operators make actual decisions [6].
  • Edit direction: Edits that push a number up are the ones most likely to make it worse. A sales team raising the forecast is not proof of a better number [1].
  • Edit size: Small tidy-up tweaks damage accuracy most often, while big adjustments tend to help. Most override rules let the wrong kind through [2].
  • The dollar tether: A consumer packaged goods brand cut its 30-day mean absolute percentage error (MAPE) from 24% to 21% using a machine-learning tool, releasing $8.5M of finished-goods stock [7]. A gain that does not turn into stock or cash is not a business result [4].
  • The sensing test: Every layer has to beat the layer beneath it. Otherwise, it adds cost and complexity for nothing [6].

The catch:

My read is that value added is a check-up, not a final ruling. Four limits matter.

  • The tool has its own critics, and their point is about money. A forecast can be 40% more accurate while the firm makes 75% less profit overall [4].
  • One bad read is not a reason to retire a step. The cited advice on a consensus pass that lost 11 of 13 weeks was to find out why it behaves that way, not to shut it down immediately [4].
  • The studies measure three different things on three different samples. Blending them into one narrative misleads.
  • A good read does not insure a company against a demand shock. It simply prices the error a better forecast can remove.

At a glance

  • The Big Shift: Planning teams now run several forecast layers, and forecast value added (FVA)—the change in error one layer contributes against a simple test—is the check that shows which layers earn their keep.
  • Why It Matters: In a study of more than 300,000 forecasts, 52% were worse than carrying last period’s value forward. A layer that cannot beat that baseline adds cost and complexity rather than accuracy, while a real 3-point error cut freed $8.5M of stock for one brand [7].
  • What I’d Watch: How teams score their own process, and what they do with the answer.
  • The frozen lag: Scoring only counts at the point the plan locks and operators make the call [6].
  • Edit direction and size: Small edits and upward edits are the two adjustments that most often damage accuracy [1][2].
  • The dollar tether: Whether a gain in points turns into stock or cash, or only into a better-looking dashboard [4].
  • The Catch: Value added is a check-up with limits. Accuracy is not profit, one poor quarter is not proof, and a better forecast does not take demand risk away.

Related reading

Sources

[1] Fildes, Goodwin & De Baets, “Forecast value added in demand planning”, International Journal of Forecasting 41(2):649–669 (2025) — https://ideas.repec.org/a/eee/intfor/v41y2025i2p649-669.html (retrieved 2026-10-15) [2] Fildes, Goodwin, Lawrence & Nikolopoulos, “Effective forecasting and judgmental adjustments: an empirical evaluation and strategies for improvement in supply-chain planning”, International Journal of Forecasting 25(1):3–23 (2009) — https://research.manchester.ac.uk/en/publications/effective-forecasting-and-judgmental-adjustments-an-empirical-eva (retrieved 2026-10-15) [3] Gibbons, “How To Use Forecast Value Added Analysis”, Demand Planning / Institute of Business Forecasting (2018-02-12), reporting Steve Morlidge’s study of more than 300,000 forecasts as cited in Gilliland, Tashman & Sglavo, Business Forecasting: Practical Problems and Solutions — https://demand-planning.com/2018/02/12/what-is-forecast-value-added-analysis (retrieved 2026-10-15) [4] Lokad, “Forecast Value Added” (reference page, reviewed Q2 2024; cites Gilliland, 2010) — https://www.lokad.com/forecast-value-added (retrieved 2026-10-15) [5] DOSS, “Shelf Life Stress: What’s Keeping Inventory Teams up at Night in 2026?” (research report dated 18 May 2026; 661 respondents) — https://www.doss.com/research/top-inventory-stressors-affecting-teams (retrieved 2026-10-15) [6] Infomineo, “Demand Sensing vs Demand Forecasting: How to Run Both Without Making Your Plan Worse” (2026-09-21) — https://infomineo.com/services/data-analytics/demand-sensing-vs-demand-forecasting-how-to-run-both-without-making-your-plan-worse (retrieved 2026-10-15) [7] Field notes — context/growth_os/customer-truth.md §demand_sensing_advanced_sop Anecdote 1: a consumer packaged goods brand cut 30-day forecast MAPE from 24% to 21% with machine-learning demand sensing and permanently trimmed $8.5M of excess finished-goods inventory.