What is forecast value add? Formula and how to use it

Forecast value add (FVA) measures whether a step in your forecasting process actually makes the forecast better. It compares the accuracy of a forecast at each stage, statistical model, sales override, consensus meeting, against a simple naive baseline, and assigns a positive or negative score to whoever touched it.

Planning teams running weekly S&OP cycles rarely question whether their process helps. They assume more inputs mean a better number. FVA is the test that checks that assumption, and the results are often uncomfortable.

Key Takeaways

• FVA compares forecast accuracy at each process step against a naive benchmark (often last period’s actuals or a seasonal random walk), using MAPE or a similar error metric.
• A study of more than 300,000 forecasts found that 52% performed worse than a simple naive forecast, according to research by Steve Morlidge cited in the Journal of Business Forecasting.
• Research covering more than 60,000 forecasts across four supply chain companies found that around 75% of statistical forecasts were manually adjusted, and small positive adjustments tended to reduce accuracy rather than improve it.
• FVA has real limitations. It tells you a step made things worse, not why, so treat it as a diagnostic starting point rather than a final verdict.
• The fastest way to start is comparing your current forecast against a naive baseline for one product category over one quarter, before adding more sophistication.

What is forecast value add (FVA)?

Forecast value add is a metric that evaluates whether a specific step or contributor in the forecasting process improves accuracy compared to doing nothing. Each touchpoint, statistical model, sales input, marketing adjustment, executive override, gets measured against a naive benchmark using an error metric such as MAPE. If the touchpoint reduces error, it has positive FVA. If it increases error, it has negative FVA.

The idea traces back to Michael Gilliland’s 2010 book The Business Forecasting Deal, and it answers a specific question mid-market planning teams tend to skip: is this meeting, this override, this extra layer of review, actually making the number better, or just making everyone feel involved?

Diego managed replenishment forecasting for a DTC beauty brand launching four new product lines each quarter. His weekly planning call added a sales override to nearly every forecast, on the assumption that the team closest to retailers knew best. When Diego finally measured FVA on the past year’s launches, the override made things worse three times out of four, mostly by inflating demand for lines with soft early reviews. He didn’t stop taking sales input. He started asking for the reasoning behind each override before it went in.

Teams running AI-powered demand planning with a strong statistical baseline tend to find that kind of override scrutiny easier, simply because the baseline itself is harder to beat.

The forecast value add formula

The core calculation is simple. For any two adjacent steps in your process, FVA equals the accuracy of the later step minus the accuracy of the earlier one, where accuracy is typically expressed as MAPE (mean absolute percentage error) or a similar error rate.

FVA (Step B vs Step A) = Accuracy of Step A - Accuracy of Step B

A positive result means Step B reduced error compared to Step A. A negative result means Step B made things worse. Most teams run this comparison at several stages: naive forecast versus statistical forecast, statistical forecast versus sales-adjusted forecast, and sales-adjusted forecast versus the final consensus number that goes into the buying plan.

The metric itself is common ground; the harder decision is picking what counts as your naive baseline.

How to calculate FVA step by step

1. Define your baseline. A naive forecast is usually a random walk (this month equals last month) or a seasonal random walk (this month equals the same month last year). Pick whichever a planner would default to if the forecasting process disappeared tomorrow.
2. Generate the statistical forecast. Run your normal forecasting software or model on the same historical data used for the baseline, without any manual changes yet.
3. Capture every override. Record the forecast after each contributor touches it, sales, marketing, operations, and the final consensus meeting, as its own separate version.
4. Wait for actuals. Once the period closes, calculate the error rate (MAPE or a comparable metric) for every version against real demand.
5. Calculate FVA at each step. Subtract each version’s error from the prior version’s error. A positive number means that step helped. A negative number means it hurt.
6. Track it over time. One bad period doesn’t condemn a process step, and one good period doesn’t vindicate it. Look at FVA across several cycles before deciding to change anything.

For definition and list snippets, this covers the two most common search formats. The harder part is what comes next: interpreting what a negative score actually tells you.

What counts as a good FVA benchmark

Positive FVA at every stage is the goal, but the size of the gap matters more than the sign. A statistical forecast that only beats the naive baseline by half a percentage point of MAPE probably isn’t worth the software license and analyst hours behind it. A sales override that beats the statistical baseline by ten points across several cycles is a genuine signal worth protecting.

Research on judgmental adjustment gives a useful reference point here. In a study covering more than 60,000 forecasts across four supply chain companies, researchers found that forecasters adjusted roughly 75% of statistical forecasts, and small positive adjustments (nudging demand up slightly) tended to reduce accuracy, while larger negative adjustments (cutting demand down meaningfully) tended to add value. That asymmetry is worth checking in your own FVA data before assuming every override is equally suspect.

Statistic callout: A study of more than 300,000 forecasts found that 52% performed worse than a simple naive forecast, per research by Steve Morlidge cited in the Journal of Business Forecasting (Spring 2016 issue, via the Institute of Business Forecasting).

That statistic is the reason FVA exists. If half of forecasting effort across an industry sample failed to beat doing nothing, the assumption that more process automatically means more accuracy doesn’t hold up. It also explains why forecast error keeps showing up downstream: see the hidden cost of inaccurate sales forecasts for how that error translates into markdown risk and missed sales over a season.

Naomi ran demand planning for a mid-market home décor brand with a dozen wholesale accounts and a growing DTC channel. Her team held a 90-minute consensus call every Monday, and every forecast that came out of it got adjusted by at least three people. When she finally ran FVA across two quarters, the statistical baseline beat the naive forecast by six points of MAPE, a solid result. The consensus call, on the other hand, added error back in almost every week, mostly from optimistic sales input on categories that were already slowing down. She didn’t cancel the meeting. She narrowed what it was allowed to adjust.

Where forecast value add breaks down

FVA is a useful diagnostic, not a complete answer. A few honest limitations are worth naming before you build a process around it.

It tells you that a step hurt accuracy, not why. A sales override that consistently adds error might reflect a genuine blind spot in the statistical model (a promotion the system didn’t see coming) or it might reflect political pressure to hit a revenue target. FVA can’t tell the two apart on its own; someone still has to ask the question.

It can be gamed. If forecast accuracy becomes a scorecard metric tied to department performance, contributors have an incentive to shade their inputs toward whatever makes their FVA look good rather than what they actually believe about demand. This is the same dynamic behind Goodhart’s Law: once a measure becomes a target, it stops measuring what it was meant to measure.

It treats demand as a single number when the future is genuinely uncertain. A forecast that lands closer to the average outcome isn’t automatically the more useful one if the business needs to plan for a range of scenarios, not just the midpoint. Lokad’s research on FVA makes this critique directly, arguing that a purely accuracy-focused lens misses the point when the real goal is minimizing the financial cost of being wrong, not just the percentage error.

None of this means FVA isn’t worth running. It means treating a negative score as the start of a conversation, not the end of one.

Across Metreecs’ work with retailers running seasonal buying cycles, one pattern shows up often: the statistical baseline is usually more accurate than the consensus process gives it credit for, and the override that hurts the most is the one added latest, right before the buying plan locks, when there’s the least time left to sanity-check it.

Using FVA to fix your forecasting process, not just score it

The point of running FVA isn’t to produce a leaderboard of who forecasts best. It’s to find the two or three steps in your process that are quietly making things worse and either fix them or remove them.

Start narrow. Pick one product category and one quarter, calculate FVA at each stage of your current process, and see where the biggest negative number shows up. If it’s the statistical model, the fix is usually in the model, not the people around it. Retail demand has enough seasonal and promotional structure that a generic time-series approach often needs recalibration by product and location rather than by category alone, a point covered in more depth in how to combine internal and external data for more accurate forecasting. If it’s a specific override stage, the fix is narrowing what that contributor is allowed to change, not removing them from the process entirely.

One mistake we repeatedly see is treating every manual adjustment as equally risky. The research above suggests the opposite: small positive nudges are the ones to watch closely, while larger, well-reasoned negative adjustments (someone catching a real demand drop the model missed) tend to hold up. Product x location demand forecasting, updated as new sell-through data comes in, gives planners a stronger statistical baseline to adjust from in the first place, which raises the bar for what counts as a genuine override rather than noise, and it shows up directly in inventory optimization results once fewer bad overrides make it into the buying plan.

Elias ran supply planning for a franchise network of quick-service restaurants using a managed forecasting service rather than an internal data team. His FVA review after one summer promotion cycle found that the naive seasonal baseline alone beat the franchise-level consensus forecast in nine of twelve regions. The consensus process wasn’t useless, it caught two regions with a genuine local event the baseline missed, but it was adding noise everywhere else. Narrowing the override to only those two regions cut the review meeting from ninety minutes to twenty.

If your FVA analysis keeps landing on the same conclusion, that the naive forecast and the statistical model are both doing fine and the consensus process is where accuracy leaks out, that’s worth a direct conversation with whoever owns that meeting.

FAQ

What is a good FVA score? There’s no universal target number, since it depends on your baseline and category. The useful signal is direction and consistency: a step with positive FVA across several cycles is adding value, and a step with negative FVA across several cycles is worth investigating or removing, regardless of the exact percentage.

How is FVA different from forecast accuracy (MAPE)? MAPE measures how far off a single forecast was from actual demand. FVA measures the change in that error between two versions of the forecast, so it tells you which step or person caused the improvement or the damage, not just how accurate the final number turned out to be.

Do I need special software to calculate FVA? No. FVA only requires you to save each version of the forecast (naive, statistical, post-override, final consensus) and compare their error rates once actuals come in. A spreadsheet works for a pilot; dedicated forecasting or planning software makes it easier to automate at scale across many products and locations.

Can FVA be negative for the statistical forecast itself? Yes, and it happens more often than most teams expect. If the statistical model is poorly tuned, missing recent promotional history, or built on too little data for a new product, it can underperform even a basic naive forecast. That’s a signal to fix the model’s inputs, not to add more manual overrides on top of it.

Should sales and marketing input be removed from forecasting entirely? Not necessarily. The research on judgmental adjustment shows overrides can add real value, particularly larger, well-reasoned ones. The goal of FVA isn’t to eliminate human input, it’s to find out which specific inputs are helping and keep those, while narrowing or removing the ones that consistently add error.

How often should I re-run FVA analysis? Quarterly is a reasonable cadence for most mid-market retailers, though high-velocity categories with frequent promotions benefit from a monthly check. One bad cycle doesn’t condemn a step, so look for a pattern across at least three to four measurement periods before changing your process.

Conclusion

Forecast value add gives planning teams a way to test an assumption almost everyone makes without checking it: that more process, more meetings, more overrides, automatically means a better forecast. The research says otherwise often enough that it’s worth measuring rather than assuming. Start with one category, compare your current process against a naive baseline, and let the numbers tell you which steps to keep.

Book a demo to see how Metreecs strengthens the statistical baseline your team’s overrides get measured against.

Sources

• Steve Morlidge study of over 300,000 forecasts, cited by Eric Wilson in "How To Use Forecast Value Added Analysis," Journal of Business Forecasting (Spring 2016), Institute of Business Forecasting: https://demand-planning.com/2018/02/12/what-is-forecast-value-added-analysis/, source for the 52%-worse-than-naive statistic.
• Fildes & Goodwin, "Effective forecasting and judgmental adjustments: an empirical evaluation and strategies for improvement in supply-chain planning," International Journal of Forecasting: https://www.researchgate.net/publication/223443452_Effective_forecasting_and_judgmental_adjustments_an_empirical_evaluation_and_strategies_for_improvement_in_supply-chain_planning, source for the 60,000-forecast, four-company adjustment study.
• Lokad, "Forecast Value Added" knowledge base article: https://www.lokad.com/forecast-value-added/, source for the FVA process description and the balanced critique of its limitations.