TL;DR & Quick Summary
Most demand planning teams report one accuracy number, usually MAPE, and treat it as the score. That single figure hides the two things that actually determine whether your inventory position is any good: whether your errors lean in a consistent direction, and whether your forecasting process beats simply reusing last week's number.
MAPE — intuitive, but breaks on low-volume and intermittent demand
WAPE — volume-weighted, the better default for mixed catalogues
Bias — the signed average error, and the metric that decides your stock position
Forecast Value Add — the only metric that asks whether the process is worth running
Key Takeaway: Error size tells you how noisy your forecast is. Bias tells you whether it will systematically stock you out. Report both, plus a naive benchmark, or you are measuring nothing useful.
Get Started: Want your current forecast measured against a naive benchmark before investing further? Schedule a Strategy Call with Cogniq AI or explore our predictive analytics services.
The Four Metrics
| Metric | What it measures | Best for | Main weakness |
|---|---|---|---|
| MAPE | Average percentage error, items weighted equally | Stable, similar-volume items | Breaks near zero demand |
| WAPE | Total error ÷ total actual demand | Mixed volumes, intermittent demand | Less intuitive to explain |
| MAE | Average error in units | Single product, operational planning | Not comparable across products |
| Bias | Average signed error | Every forecast, always | Says nothing about error size |
MAPE: Intuitive, and Frequently Wrong
Mean Absolute Percentage Error averages the percentage error across items, weighting each equally.
Its appeal is that it produces a number anyone understands: "we were 14% out." Its problem is the division by actual demand.
Consider two forecasts:
| Forecast | Actual | Absolute error | Percentage error |
|---|---|---|---|
| 2 units | 1 unit | 1 unit | 100% |
| 1,010 units | 1,000 units | 10 units | 1% |
The first miss is one unit and scores 100%. The second is ten units and scores 1%. Average them and you report 50.5% — a figure describing neither situation.
Across a catalogue with many slow movers, MAPE becomes dominated by items that barely matter. It is a real distortion, not an edge case: MAPE works well when actual values stay comfortably above zero and planners want an intuitive average, but low-volume and zero-demand observations distort it badly (Baeldung).
There is a subtler problem too. MAPE penalises over-forecasting more heavily than under-forecasting, because the error is divided by actual demand. Optimising against it therefore nudges a process toward forecasting low — which produces stockouts.
WAPE: The Better Default
Weighted Absolute Percentage Error divides total absolute error by total actual demand, rather than averaging individual percentages.
WAPE = Σ|forecast − actual| ÷ Σ actual
This makes high-volume periods and products contribute proportionally more, which is usually what you want commercially. Getting your best-selling line 5% wrong matters more than getting a slow mover 80% wrong.
WAPE is recommended over MAPE where sales are low or intermittent, and is a preferable default where volume varies widely across the catalogue (Xorosoft, Baeldung). It is also defined when demand is zero, which MAPE is not.
The trade-off is explanatory. "Our WAPE is 12%" needs a sentence of context in a way "our MAPE is 12%" does not. That is a presentation problem, and worth accepting.
Bias: The Metric That Decides Your Stock Position
Bias is the average signed error — errors keep their direction rather than being made absolute.
Bias = Σ(forecast − actual) ÷ Σ actual
Positive means you consistently forecast high. Negative means consistently low. Near zero means your misses cancel out.
This matters more than error size, and the reason is arithmetic. Random error averages out across a portfolio and across time. Bias does not — it accumulates every period, in the same direction.
Compare two forecasting processes:
| Process | WAPE | Bias | Result |
|---|---|---|---|
| A | 20% | 0% | Noisy, but roughly the right total stock |
| B | 10% | −8% | Half the error, permanent stockouts |
Process B looks twice as good on the headline number and is worse to operate. Being consistently 8% low means the shortfall recurs every period, and safety stock has to permanently absorb a known, predictable gap — which is an expensive way to compensate for a correctable defect.
Bias also has a diagnosable cause in a way that random error does not. Persistent negative bias usually means a sandbagged sales forecast, an unmodelled growth trend, or promotional lift being excluded. Persistent positive bias usually means optimistic targets being fed in as forecasts. Both are process problems that show up nowhere in an absolute-error metric.
Forecast Value Add: The Only Question That Matters
Every metric above tells you how well the forecasting process performed. Forecast Value Add asks whether it should exist.
Compare your forecast against a naive benchmark — most simply, last period's actual demand:
FVA = (naive benchmark error) − (your forecast error)
Positive means your process adds value. Zero or negative means it does not, and you would get the same or better results using last week's number and redeploying the planners.
This check is uncomfortable and frequently skipped, which is exactly why it is worth running. A forecasting process consuming several people's time can be measurably worse than a one-line rule, and no accuracy metric will reveal that — because accuracy metrics assume the process is worth doing and only measure how well.
Two rules for running it honestly:
Use a naive benchmark you would actually accept. Last period's actual for stable demand; the same week last year for seasonal products. Choosing a deliberately weak benchmark to make your process look good defeats the purpose.
Run it per product family. In most businesses, forecasting adds real value on some families and destroys it on others — typically the intermittent ones. Knowing which is which lets you stop forecasting the families where a simple rule wins.
How to Measure It Honestly
The metric matters less than the discipline of the test. Three rules keep the numbers meaningful.
Measure at the horizon you actually commit at. If your supplier lead time is eight weeks, an accuracy figure computed one week ahead is describing a decision you never make. Forecast error grows with horizon, sometimes steeply, so a one-week number is flattering and useless. Report accuracy at the lead time you buy against — and if you buy at several horizons, report several.
Lock the forecast before the period starts. Comparing a forecast that was revised mid-period against the actual it was revised toward measures nothing except your ability to update a spreadsheet. Snapshot the forecast at the moment of commitment and score that number, not the final version.
Score at the level you decide at. Accuracy improves as you aggregate — a forecast that is poor per SKU per store can look excellent at national level, because errors cancel. If you place orders per SKU per location, that is where the number must be computed. Reporting an aggregate figure while ordering at detail level is the most common way teams convince themselves a weak forecast is strong.
The Ways These Numbers Get Gamed
Not usually deliberately, but the incentives are real and worth naming:
- Aggregating up until the number looks acceptable
- Shortening the horizon to a week nobody buys at
- Excluding "exceptional" periods — promotions, disruptions, seasonal peaks — which are precisely the periods where forecasting is worth doing
- Reporting MAPE on a catalogue with many slow movers, where the metric rewards systematic under-forecasting
If an accuracy figure improves sharply without a change in method, check which of these moved first.
What to Report
Three numbers, together, per product family:
- WAPE — how large the misses are, weighted by what matters
- Bias — whether they lean, and which way
- FVA — whether the process beats doing nothing clever
Plus one rule: segment before you report. A blended figure across a whole catalogue averages your stable high-volume lines together with your intermittent slow movers and describes neither. Report by family, and expect the numbers to look very different across them.
On targets: there is no universal "good" MAPE, and figures quoted without context are unreliable. Achievable accuracy is set by how volatile the underlying demand is. Stable high-volume products routinely forecast within single digits; promotion-driven and intermittent items may never get below thirty. Compare against your own naive benchmark and your own prior performance — not a vendor case study on a different catalogue.
How This Connects to Inventory
Accuracy metrics are not an end in themselves. They feed two decisions:
Safety stock sizing. Safety stock should be set against forecast error distribution, not against a padded point forecast. If you know your WAPE and its variability per family, you can size buffers against a service-level target rather than guessing. Our predictive analytics for inventory management guide covers this in detail.
Where to invest effort. Families with high WAPE, high bias and positive FVA are worth improving. Families with negative FVA should be moved to a simple rule and left alone. That triage saves more planner time than any modelling improvement.
There is a connection to demand amplification here too: padding a biased forecast with extra buffer at every tier is one of the mechanisms that produces the bullwhip effect, covered in our bullwhip effect guide. Fixing bias upstream removes the reason for the padding.
For the broader forecasting approach, our AI demand planning and inventory forecasting guide covers method rather than measurement.
Conclusion
Reporting one accuracy percentage is how forecasting teams end up optimising a number that has stopped describing anything useful. MAPE rewards under-forecasting and breaks on slow movers. Error size without bias tells you nothing about whether you will be permanently short. And no accuracy metric at all can tell you whether the forecasting process is beating a one-line rule.
Report WAPE with bias, segmented by family, and check Forecast Value Add at least once a year. The third one is the uncomfortable check, and it is the one most likely to change what your team does on Monday.
Schedule a Strategy Call with Cogniq AI and we will measure your current forecast against a naive benchmark before recommending anything — including when the honest answer is that a simple rule already beats what you have.
