Forecast Value Added: Where in Your Process Does the Forecast Actually Get Better?

Why Forecast Quality Matters — In Concrete Numbers

Forecast quality determines how smoothly sales, production, and order fulfilment can work together. When the forecast is wrong, three things show up on the P&L:

• Stock-outs hit revenue. Lost orders. Customers calling sales to complain. Sales reps spending the next week firefighting instead of selling.
• Expediting hits margin. Late corrections trigger express freight and unplanned rush orders. Every one of those eats the margin on the order.
• Over-forecasting hits working capital. Excess inventory typically ties up 15–30% more working capital than it should. Money sitting in warehouses is money that isn't doing anything else.

Forecasting is not a back-office statistics exercise. It is the upstream signal that determines how much money operations leaves on the table downstream.

What We Currently Don't See

Almost every S&OP process measures total forecast accuracy — the MAPE of the final number that gets released. That tells you whether the overall process worked. It does not tell you which step in the process improved the forecast and which one made it worse.

The consequence: planning effort gets spread evenly across every step. The statistician spends hours tuning the model. The demand planner spends hours adjusting it. The consensus meeting runs for two hours. Management reviews it. Sales overrides parts of it. All of these steps consume effort. Some of them improve the forecast. Some of them make it worse. You can't tell which is which.

What FVA Actually Measures

Forecast Value Added asks one simple question for each step in the planning process: did this step make the forecast more accurate, less accurate, or leave it unchanged?

A worked example from a typical industrial S&OP cycle:

The statistical model started at 45% MAPE. The planner adjustment improved it. The consensus meeting improved it further. The management release step then made it worse — by exactly as much as the consensus meeting had improved it. All the effort the consensus team put in was undone in the last step.

Without FVA, you see a 40% MAPE at the end and conclude "the process works." With FVA, you see that one specific step is destroying value — and you can have a precise conversation about why.

Why the Last Step Often Destroys Value

In our engagements, four root causes show up consistently when a process step has negative FVA:

1. Sandbagging. Sales plans low so they can beat their target. The forecast that goes into production isn't the best estimate — it's the politically safe number.
2. Gut-feel overrides. A manager adjusts the forecast based on what they heard from a key account last week. Sometimes that signal is real and improves things; often it's anecdote dressed up as data.
3. Political compromises. The consensus number is the midpoint between sales and operations, not the best statistical estimate. It satisfies both camps and is less accurate than either.
4. Outdated assumptions. Adjustments are based on information that was true three months ago but isn't anymore. Nobody refreshes them; everyone keeps applying the same correction.

None of these are individual failures. They are structural patterns. Negative FVA isn't a personal verdict — it's a process signal.

The Naive Benchmark: Without It, FVA Is Worthless

FVA is a relative metric. To compute it you need a reference line — something that says "this is what the forecast would look like with zero effort." That reference line is a naïve forecast. Three flavours, picked by the shape of the demand:

The choice of benchmark is a business decision, not a technical one. For seasonal products, Seasonal Naive is the fair yardstick — anything else gives the statistical model an unearned advantage. Rule of thumb: if the statistical forecast beats Seasonal Naive by less than 5%, the modelling effort rarely pays for itself. The naïve forecast belongs in every dashboard and every review, as a permanent comparison line.

Is the Difference Real, or Just Noise?

A +2% FVA isn't necessarily an improvement. It might be statistical noise. Three methods, increasing in robustness:

• Bootstrap confidence intervals. Resample the historical periods to get a 95% confidence interval around the FVA. If the interval includes zero, the difference isn't meaningful.
• Diebold-Mariano test. The established statistical test for whether two forecast error series come from the same distribution. Returns a p-value for the null hypothesis "equal accuracy."
• Equivalence testing. Asks the opposite question: are these two forecasts statistically equivalent? Useful when you want to know whether a simpler process is just as good as a more complex one.

Practical consequence: observe at least 6–12 periods before changing a process. Anyone who restructures S&OP on the basis of a single month is optimising noise.

FVA Is a Change-Management Conversation

The hardest part of an FVA rollout isn't the math. It's the conversation. The wrong framing turns FVA into a blame engine; the right framing turns it into a diagnosis.

Three rules that make this work in practice:

1. Never base a decision on a single value. Always use a rolling 6–12 month window. Otherwise you're measuring noise.
2. Make the aggregation level transparent. A step that's negative at SKU level can be positive at product-family level. Both views matter; show both.
3. Always present action options. Don't just say "this step is bad" — say: eliminate it, coach the team running it, or build tooling that supports it better.

Frame FVA as a diagnosis, not a verdict.

How to Roll It Out: Three Phases, Each With Visible Value

No big-bang implementation. Three phases, each delivering something usable before the next one starts.

Where FVA Comes From

FVA was popularised in the early 2000s by Mike Gilliland at SAS. His core thesis, paraphrased:

"Most steps in the forecast process don't add value — and many actively destroy it."

Today FVA is a standard component of S&OP reviews at large industrial companies. The thesis has held up across two decades of practice, helped along by another observation from the M-competition series: simple, transparent forecasting methods are often surprisingly competitive — and FVA forces an honest measurement of that, instead of letting complexity be sold as value.