The number everyone quotes
Ask any demand planner how accurate their forecast is and you will get a percentage. That percentage is almost always MAPE, the Mean Absolute Percentage Error. It is the default metric in every planning tool, every ERP report, and every board deck. And it is the metric most likely to be misinterpreted by the person reading it.
MAPE is not a bad metric. It is a useful metric used outside its limits. The problem is that MAPE behaves predictably when demand is steady and high, and behaves badly when demand is low, intermittent, or zero. Since most manufacturers have a long tail of slow-moving SKUs, MAPE on those SKUs is often meaningless, sometimes infinite, and potentially misleading when interpreted without volume and intermittency context.
This article is a practitioner's guide to reading MAPE correctly. It covers what the formula actually computes, where it breaks, what a "good" MAPE actually looks like in different industries, and which metrics you should track alongside it to avoid being lied to by a single number.
The formula, plainly
MAPE is the average of the absolute percentage errors between your forecast and the actual demand, across all periods and SKUs in scope. Formally:
For each period, compute the absolute error divided by the actual demand, express it as a percentage, then average across all periods. The result is a single percentage that represents your average miss rate, weighted equally across all observations regardless of size.
The appeal is obvious. A single percentage is easy to communicate. "Our MAPE is 18%" sounds precise. Executives can track it over time. Tools can benchmark against it. Equal weighting can be useful for some questions, but it does not express volume or financial impact.
Where MAPE breaks
MAPE has three well-documented failure modes. Every demand planner hits them. Most do not realise they have hit them because the metric still produces a number, it just produces a meaningless one.
First, MAPE explodes when actual demand is low. If you forecast 10 units and the actual is 1, your percentage error is 900%. If you forecast 10 and actual is 100, your percentage error is 90%. These examples have different absolute errors: 9 units and 90 units. For an equal 9-unit error, forecasting 109 against an actual of 100 produces 9%, while forecasting 10 against an actual of 1 produces 900%. The denominator changes the interpretation. This means slow movers, which have low actual demand, will dominate your MAPE even if they are a small part of your revenue. A single slow mover with a 500% error can drag an otherwise healthy 15% MAPE up to 25%.
Second, MAPE is undefined when actual demand is zero. Division by zero. Many tools silently skip these periods, which means you are not actually computing the average across all periods. You are computing the average across non-zero periods. If 30% of your SKUs have zero demand in a given month (common for spare parts, seasonal items, new products), your MAPE is computed on a biased sample.
Third, MAPE weights all errors equally regardless of volume. A 50% error on a SKU that sells 2 units a month contributes the same as a 50% error on a SKU that sells 2,000 units a month. In revenue terms, the second error is 1,000 times more important. MAPE cannot see this. A weighted metric like WMAPE (Weighted MAPE) can.
What a good MAPE actually looks like
There is no universal good MAPE. The number depends on your industry, your product mix, your forecast horizon, and your level of aggregation. A 15% MAPE might be excellent for one company and terrible for another. Context is everything.
Use your own baseline at the same item-location grain and forecast horizon. Evaluate later observations that were not used to fit the model, and separate segments with different demand behavior. A universal 25% or 40% threshold is not a reliable verdict across businesses.
The forecast accuracy chapter in Forecasting: Principles and Practice explains percentage-error limitations and alternatives. A useful scorecard states the test period, zero-demand treatment, forecast horizon, aggregation, and comparison baseline. There is no substantiated industry benchmark asserted here.
What to track alongside MAPE
No single metric tells the truth about forecast quality. MAPE should always be read alongside three other signals.
Bias. MAPE tells you the size of your errors but not their direction. Bias measures whether you systematically over-forecast or under-forecast. A MAPE of 20% with zero aggregate bias means positive and negative errors balance overall; it does not prove random errors or a healthy forecast. If positive bias is defined as forecast minus actual, a positive bias indicates net over-prediction, which is a different and more fixable problem. Track bias per SKU, per category, and per planner. Persistent bias is more actionable than random error.
WMAPE. Weighted MAPE weights each error by the actual demand volume, so a miss on a high-volume SKU counts more than a miss on a low-volume SKU. Using WMAPE defined as the sum of absolute errors divided by the sum of actual demand gives larger-volume observations more influence. It is still undefined if aggregate actual demand is zero. If your MAPE is 30% but your WMAPE is 12%, it means your errors are concentrated in low-volume SKUs, which still needs review for service and cost consequences. If your MAPE is 30% and your WMAPE is 28%, your errors are in high-volume SKUs, which calls for review of operational impact.
Forecast value added (FVA). FVA measures whether your manual overrides actually improve the forecast or make it worse. It compares the accuracy of the statistical baseline to the accuracy of the final published forecast after overrides. If your statistical baseline has a MAPE of 22% and your final forecast (after planner overrides) has a MAPE of 25%, your planners are destroying value. This happens more often than planners admit. Track FVA per planner to find where overrides help and where they hurt.
The reading checklist
Before you report a MAPE number to anyone, including yourself, answer these five questions.
One. What level of aggregation is this MAPE computed at? Aggregation changes the metric and can hide offsetting errors; it does not guarantee a lower value in every dataset. Comparing the two is meaningless. Always state the level.
Two. What forecast horizon is this MAPE for? Forecast difficulty changes with horizon; a shorter horizon is not guaranteed to produce a lower error in every sample. Always state the horizon.
Three. How are zero-demand periods handled? If they are skipped, the reported metric describes only non-zero periods. If they are included with a small epsilon added to the denominator, state the epsilon.
Four. What is the bias? A MAPE number without a bias number is half a diagnosis.
Five. What is the WMAPE? If you only report MAPE, you are hiding where the errors actually sit.
The takeaway
MAPE is a useful metric. It is not a verdict. Read it alongside bias, WMAPE, and FVA, always at a stated level of aggregation and horizon, and you will have a forecast accuracy picture that actually reflects reality. Read it alone, and you will have a number that sounds precise and tells you nothing.
The best demand planners do not optimise for a lower MAPE. They optimise for a forecast that the business can plan around. Those are not always the same thing.
Evaluate the metric in a pilot
Use the forecast pilot checklist to define the data and test period, or explore GrepEye Supply Chain to review a planning workflow.