TL;DR & Quick Summary
MAE, WAPE and MAPE measure the same forecast misses in three different ways. On the identical five-item example below, MAPE says 36.3%, WAPE says 7.6%, and MAE says 23.4 units. None of them is wrong. They answer different questions.
MAE (mean absolute error): the average miss in units. Best for one item.
WAPE (weighted absolute percentage error): total miss divided by total demand. Best for comparing across items, sites and periods.
MAPE (mean absolute percentage error): the average of each item's percentage miss. Dominated by small items and undefined when demand is zero.
Rule of thumb: report WAPE as the headline, add bias for direction, and use MAE where planners think in units.
Key Takeaway: The metric decides what your team chases. MAPE sends attention to tiny items. WAPE sends it to the items that carry the volume.
Get Started: Want your forecasts measured properly? Schedule a strategy call with Cogniq AI or see our predictive analytics services.
The Definitions
For each item and period you have an actual value and a forecast. The error is the difference. The three metrics combine those errors differently.
MAE adds up the sizes of the errors and divides by how many there are:
MAE = average of |actual − forecast|
MAPE turns each error into a percentage of its own actual, then averages those percentages:
MAPE = average of (|actual − forecast| ÷ actual)
WAPE adds up all the errors and divides by all the actual demand:
WAPE = sum of |actual − forecast| ÷ sum of actual
def mae(actual, forecast):
return sum(abs(a - f) for a, f in zip(actual, forecast)) / len(actual)
def mape(actual, forecast):
return 100 * sum(abs(a - f) / a for a, f in zip(actual, forecast)) / len(actual)
def wape(actual, forecast):
return 100 * sum(abs(a - f) for a, f in zip(actual, forecast)) / sum(actual)
That is the whole difference: MAE averages units, MAPE averages percentages, and WAPE divides totals. The first is in units, the other two are percentages.
One Example, Three Answers
Five items, one period:
| Item | Actual | Forecast | Miss (units) | Miss (% of its actual) |
|---|---|---|---|---|
| A | 1,000 | 950 | 50 | 5.0% |
| B | 400 | 440 | 40 | 10.0% |
| C | 120 | 100 | 20 | 16.7% |
| D | 10 | 15 | 5 | 50.0% |
| E | 2 | 4 | 2 | 100.0% |
The total miss is 117 units against total demand of 1,532 units.
| Metric | Result | How it got there |
|---|---|---|
| MAE | 23.4 units | 117 ÷ 5 |
| MAPE | 36.3% | average of 5.0, 10.0, 16.7, 50.0 and 100.0 |
| WAPE | 7.64% | 117 ÷ 1,532 |
The forecast missed the business's total demand by about 7.6%. That is the figure that matters for stock and service. But MAPE reports 36.3%, because items D and E have large percentage errors and count as much as item A, which carries two-thirds of the volume. Those two small items account for 7 of the 117 units of error.
A planner who is told "MAPE is 36%" will spend the afternoon on items D and E. A planner who is told "WAPE is 7.6%" will look at where the 117 units came from, which is mostly A and B.
When Each Metric Is the Right Choice
Use MAE for a single item or a single unit
If you're reviewing one product and the people reading think in units, MAE is the clearest statement: "we miss by about 40 cases a week". It also compares two forecasting methods fairly on the same item, because nothing else is changing.
Its weakness is that it can't be compared across items of different sizes. An MAE of 40 means something very different for an item selling 50 a week than for one selling 5,000.
Use WAPE to compare and to report
WAPE puts every item on one scale, weights by volume, and stays defined when individual items sell nothing in a period. That makes it the sensible headline across a catalogue, between sites, or over time. If you only report one number, report this one.
Its weakness is that it hides which items cause the error. Break it down by segment, such as fast, medium and slow movers, to see where the misses are.
Use MAPE sparingly
MAPE is easy to explain, and it has a place when every item in the comparison matters equally and demand is always well above zero. Its problems are real, though:
- It breaks at zero. If an item sold nothing in a period, the percentage can't be calculated, and many tools silently skip it.
- It's dominated by small items. Items with tiny volumes produce huge percentages and swamp the average.
- It's asymmetric. Forecasting 100 when the actual was 50 is a 100% error. Forecasting 50 when the actual was 100 is a 50% error. The first is penalised twice as hard, so MAPE favours forecasts that run low.
In our earlier tests, MAPE couldn't be calculated for any of the 20 slow-moving items in a 60-item catalogue, which is how it quietly dropped the hardest items from the result. That experiment is in forecast accuracy metrics in Python.
Squared Errors: Where RMSE Fits
RMSE squares each error before averaging, then takes the square root. Squaring makes large misses count much more than small ones.
Compare two forecasts with the same total error:
- Four exact forecasts and one miss of 100 units: MAE is 20.0, RMSE is 44.7
- Five misses of 20 units each: MAE is 20.0, RMSE is 20.0
If a single big miss is far more costly than several small ones, such as a stockout of an expensive part, RMSE reflects that and MAE doesn't. If every unit of error costs roughly the same, MAE is simpler and easier to explain.
Which Metric for Which Question?
Match the metric to the question being asked:
| Question | Best metric |
|---|---|
| "How far off are we on this one product, in units?" | MAE |
| "How accurate is the whole forecast, across products of different sizes?" | WAPE |
| "Did accuracy get better or worse this quarter?" | WAPE, on the same segments each time |
| "Are big misses much worse for us than small ones?" | RMSE, alongside MAE |
| "Do we tend to run high or low?" | Bias |
| "Is the forecast better than just repeating last week?" | Forecast value add against a naive benchmark |
A common situation shows why the choice matters. A planning team reports MAPE each month and sees it rise from 28% to 33%. Panic follows. But WAPE, calculated from the same data, is flat at 9%. The rise in MAPE came entirely from a handful of slow items that sold two or three units fewer than forecast. The business wasn't forecasting worse. A different metric would have saved a week of investigation.
All Three in One Function
For analysts who want all of them at once, this returns each metric and skips items with zero actual demand only for MAPE, so you can see how many were left out:
def error_report(actual, forecast):
n = len(actual)
errs = [abs(a - f) for a, f in zip(actual, forecast)]
nonzero = [(a, e) for a, e in zip(actual, errs) if a != 0]
return {
"MAE": sum(errs) / n,
"WAPE_%": 100 * sum(errs) / sum(actual),
"MAPE_%": 100 * sum(e / a for a, e in nonzero) / len(nonzero),
"MAPE_items_skipped": n - len(nonzero),
"bias_%": 100 * (sum(forecast) - sum(actual)) / sum(actual),
}
error_report([1000, 400, 120, 10, 2], [950, 440, 100, 15, 4])
# MAE 23.4, WAPE 7.64, MAPE 36.3, bias -1.50
Run on the five items in our example, it returns the same figures as the table above. Add a sixth item with zero actual demand and MAPE_items_skipped becomes 1, which is the number most tools leave out.
Direction Matters Too: Add Bias
All three metrics ignore which way the errors lean. A forecast that is always 10% too low and one that is sometimes high and sometimes low can have the same WAPE. They create very different stock problems: the first causes steady stockouts, while the second mostly causes noise that safety stock can absorb.
Bias is the signed version:
bias = (sum of forecasts − sum of actuals) ÷ sum of actuals
In our five-item example the bias is −1.5%, so the forecast ran slightly low overall. That's fine here, but you'd want to know if it were −10% every month. Our guide to forecast accuracy metrics explains how to combine error and bias, and why a forecast that beats a naive benchmark matters more than any single number.
A Simple Reporting Standard
A practical set that works for most planning teams:
| Report | Purpose |
|---|---|
| WAPE by segment | How big the misses are, weighted by volume |
| Bias by segment | Whether the forecast leans high or low |
| MAE for key items | What planners think in: units |
| Forecast value add against a naive forecast | Whether the forecast is worth running |
Segment by something meaningful: fast, medium and slow movers, or the ABC-XYZ groups in our ABC-XYZ inventory analysis in Python.
Which Horizon Are You Measuring?
Accuracy depends heavily on how far ahead the forecast was made. A forecast made one week ahead will be more accurate than one made eight weeks ahead. Measure at the horizon your decisions use: if suppliers need orders six weeks ahead, the forecast that matters is the one made six weeks before the sales period.
Common Mistakes
- Reporting MAPE as the headline. It will rank small items as your biggest problem.
- Comparing MAE across items of different sizes. Use WAPE for that.
- Averaging each item's WAPE. Compute WAPE on totals, otherwise it turns back into something like MAPE.
- Ignoring zero-demand periods. Check how your tool handles them, and report how many items a metric excluded.
- Showing a single number for the whole catalogue. Segment first.
- Forgetting bias. Two forecasts with the same WAPE can lean in opposite directions.
What to Do Next
If your current reports use MAPE, don't throw them away. Add WAPE and bias next to it, and look at what changes. Where the rankings differ, you've found the items the old metric was misleading you about. For intermittent items that often sell nothing, see Croston, SBA and TSB in Python. For seasonal items, see seasonal demand forecasting in Python.
Conclusion
MAE, WAPE and MAPE aren't competing answers to one question. MAE gives the average miss in units, WAPE gives the total miss as a share of demand, and MAPE gives the average of each item's own percentage. On the same five items they gave 23.4 units, 7.6% and 36.3%. Use WAPE as your headline, add bias for direction, keep MAE for the single-item conversation, and treat MAPE with care.
Schedule a strategy call with Cogniq AI to have your forecasts measured against a naive benchmark, or explore our predictive analytics services.



