CPG Scheduler
All posts

Forecast Value Added: Is Your Demand Override Beating a Naive Guess?

forecastinginventory planningstockoutscpg operations

A planner nudges next week's granola forecast up 12% "because the display is going in." It feels responsible—you have information the statistical model doesn't. Two weeks later the display underperforms, you've overproduced, and now there's a shelf-life scramble to move aging inventory before it dates out.

Was the override worth it? Most demand-planning teams have no idea, because they never measure it. Forecast Value Added (FVA) is the discipline that answers the question quantitatively: did each step in your forecasting process—and each human touch—actually make the number better than doing almost nothing?

This is a different question from the ones our other forecasting posts tackle. Tracking signal catches bias. Croston's method handles intermittent SKUs. Forecasting that survives the floor is about robustness. FVA is about process accountability—proving which analytical steps and manual overrides earn their keep, and which ones quietly destroy accuracy.

What FVA actually is

FVA analysis measures the improvement in accuracy and bias achieved by each step of your forecasting process, compared against a simple naive benchmark. Instead of asking "is the forecast accurate?", it asks "is the forecast more accurate than it would be if we did the laziest possible thing?"

The core idea is a "stairstep" of process stages:

  1. Naive benchmark — last period's actuals, or a flat moving average. Cheap and dumb by design.
  2. Statistical baseline — your system-generated forecast (exponential smoothing, ARIMA, whatever).
  3. Analyst override — the judgmental adjustment a planner makes for promos, weather, distribution gains.
  4. Consensus / S&OP number — the final agreed forecast after the cross-functional meeting.

Each rung is supposed to add accuracy. FVA measures whether it actually does. The formula for any step is straightforward:

FVA = (Benchmark error − Process error) ÷ Benchmark error

Positive = the step earned its keep. Negative = it made the forecast worse.

The mindset shift is the whole point: you audit the process, not just the final forecast. A forecast can be "accurate" and still fail FVA if a naive guess would have been just as good—meaning all the labor between the raw data and the final number was wasted.

Choosing the naive benchmark

Your yardstick has to be cheap and dumb, because that's what makes the comparison meaningful. Common choices:

  • Random walk (last actual): next period equals this period. Good default for stable demand.
  • Seasonal naive: this period equals the same period last year. Good for seasonal SKUs.
  • Simple moving average: a flat average of the last N periods. Smooths noise.

The benchmark should require essentially zero effort and zero judgment. If your sophisticated statistical model and three planners in a two-hour meeting can't beat "same as last week," that's the finding—and it's a valuable one.

One technical caution: because judgmental adjustments are made at the disaggregated SKU level across hundreds or thousands of series, the choice of error measure can flip your conclusion about whether adjustments helped. MAPE, MASE, and bias metrics can disagree. Pick your error measure before you run the analysis, so you're not tempted to cherry-pick the metric that flatters your team.

The uncomfortable evidence

Here's why this matters. When researchers pooled roughly 147,000 forecasts across six studies, they found that judgmental adjustments improved bias and accuracy for only just over half of SKUs. In plain terms: roughly half the manual override effort added no value or actively hurt.

It gets more specific, and more actionable:

  • Direction matters. Upward (positive) adjustments were more likely to worsen performance. Downward (negative) adjustments typically improved accuracy—particularly when they were large. Planners tend toward optimism, and optimism is exactly where value leaks out.
  • Timing and size matter. Adjustments made closer to the sales point tend to grow larger and more positive—but these do not consistently enhance accuracy. That's a direct warning against last-minute gut overrides.

Before you conclude "ban all overrides," though, look at the earlier landmark evidence: in a study of 60,000+ forecasts from four supply-chain companies, judgmental adjustments increased accuracy on average in three of the four companies. So the process isn't worthless—it's inconsistent. Judgment adds real value in some teams and on some SKUs, and destroys it in others. That's a management problem, and the only way to manage it is to measure it. FVA is the measurement.

How to run an FVA analysis on a CPG team

You don't need new software to start. You need discipline about logging and a rolling comparison window.

Step 1 — Log every stage. For each SKU each period, record four numbers: the naive benchmark, the statistical baseline, the analyst-adjusted forecast, and the final consensus number. If you only keep the final number, FVA is impossible—you've thrown away the evidence.

Step 2 — Compute error at each stage. Once actuals land, calculate your chosen error metric (say, absolute percentage error and bias) for every stage, per SKU, over a rolling window—13 weeks is a reasonable starting point.

Step 3 — Build the FVA stairstep table. For each SKU or product family, show the average error at each stage plus the FVA% versus the naive benchmark and versus the prior step. This is the artifact that makes the invisible visible.

Step 4 — Segment ruthlessly. Don't average FVA across the whole portfolio—that hides everything. Break it out by SKU, by planner, and by channel. You want to know which people add value on which items, not whether "forecasting" in aggregate is net positive.

Reading the results: four verdicts

Every SKU-planner combination lands in one of four buckets:

  1. Override beats statistical and naive. The judgment is genuinely adding accuracy. Keep it, document what the planner knows, and reinforce the behavior.
  2. Statistical beats naive, override adds nothing. The model is doing the work; the human touch is noise. Stop touching those SKUs—let the model run untouched and redeploy the planner's time.
  3. Override falls below naive. The red flag. You are literally paying skilled people to make forecasts worse than "same as last week." These are your highest-priority intervention.
  4. Everything ≈ naive. Nobody beats the dumb benchmark. That's usually a signature of intermittent or lumpy demand—the SKU doesn't belong in a standard override workflow at all. Route it to Croston's method instead.

That second bucket—"no-touch" SKUs where the model already wins—is often the biggest efficiency unlock. Freeing planners from items where their judgment demonstrably doesn't help lets them concentrate on the SKUs where it does.

From forecast to floor: why FVA is a scheduling issue

Here's the part that makes FVA more than a forecasting-team nicety. A forecast that fails FVA doesn't just sit in a spreadsheet being wrong—it propagates straight onto your production line.

An unaudited override becomes a schedule error:

And the two failure directions map to the two floor disasters. A failed-FVA upward bias means overproduction and shelf-life risk—especially painful for the perishable SKUs covered in shelf-life-constrained scheduling. A failed-FVA downward miss means stockouts and the firefighting that follows. The evidence that upward adjustments are the ones most likely to hurt should make every scheduler nervous, because upward optimism is precisely what fills the warehouse with product that dates out.

This is also why FVA only pays off when it closes the loop. Forecasts that live in spreadsheets don't improve inventory performance. Value appears only when the (audited) forecast actually drives ordering and production decisions, and when the system tracks accuracy over time per item. An FVA number you calculate but never act on is itself a step that adds no value.

Operator takeaways

Start small and make it routine:

  • Pick a naive benchmark today. Random walk or seasonal naive—choose one and start logging it alongside your real forecasts.
  • Log all four stages (naive, statistical, override, consensus). You can't audit what you didn't record.
  • Measure FVA per SKU-planner, not in aggregate. Averages hide the problem; segmentation exposes it.
  • Give the "no-touch" SKUs a free pass. Where the model already beats the human, let the model run and reclaim the time.
  • Watch upward and late adjustments hardest. These are the statistically most likely value-destroyers. A big optimistic override made the week before the sale deserves the most scrutiny.
  • Route the lumpy stuff elsewhere. SKUs where nothing beats naive are intermittent-demand candidates, not override candidates.

FVA doesn't tell you to stop forecasting or to fire your planners. It tells you where judgment earns its keep and where it's quietly costing you a full shift of changeovers and a pallet of expired product. In a CPG operation, that's not a back-office metric—it's the difference between a run plan built on evidence and one built on optimism.

Sources

More from the journal