MAPE stands for Mean Absolute Percentage Error, and it’s the most widely used way to measure how accurate a forecast is. It tells you, on average, how far off your predictions were from what actually happened, expressed as a percentage. A MAPE of 15% means your forecasts were off by 15% on average.
How MAPE Is Calculated
The formula takes each forecast period, calculates how far off the prediction was from the actual value, converts that gap into a percentage of the actual value, and then averages all those percentages together. In plain terms: for every data point, you subtract the forecast from the actual, ignore whether the error was positive or negative (that’s the “absolute” part), divide by the actual value to get a percentage, then average everything.
Say you forecasted sales of 90, 105, and 200 units over three months, and actual sales turned out to be 100, 100, and 180. The individual percentage errors would be 10%, 5%, and about 11%. Average those, and your MAPE is roughly 8.7%.
The reason MAPE is so popular is that the percentage format is immediately understandable. You don’t need to know anything about the scale of the data to interpret it. A MAPE of 12% means the same thing whether you’re forecasting ten units or ten million dollars.
What Counts as a Good MAPE
General benchmarks from the forecasting literature break it down like this:
- Under 10%: Highly accurate forecasting
- 10% to 20%: Good, reasonable forecast
- 20% to 50%: Acceptable but room for improvement
- Over 50%: Poor accuracy
These thresholds are rough guides, not universal rules. Some industries consider 20% excellent because their data is inherently volatile (fashion retail, for instance), while others like utilities or mature consumer goods expect single-digit MAPE. What matters most is how your MAPE compares to a baseline, like a simple naive forecast or your previous model’s performance.
The Under-Forecasting Bias
MAPE has a quirk that catches many people off guard: it systematically favors forecasts that are too low over forecasts that are too high. This happens because of the math. When you divide by the actual value, overshooting and undershooting by the same absolute amount produce different percentages. If actual demand is 100 and you forecast 80, the error is 20%. But if you forecast 120, the error is still 20 in absolute terms, yet the percentage is also 20%. The asymmetry shows up more dramatically at other values and compounds across a dataset.
The net effect, as researchers have documented, is that MAPE “rewards the timid forecaster and punishes the bold one.” A model that consistently undershoots will often score better on MAPE than one that sometimes overshoots, even when their total errors are similar in size. This matters in practice. If you optimize a model purely to minimize MAPE, you can introduce a consistent low bias. In supply chain contexts, one simulated study showed that a MAPE-optimal forecast led to what the researcher called a “catastrophic stock situation,” because the forecast systematically underestimated demand and caused stockouts.
When MAPE Breaks Down
MAPE has a fundamental problem with zeros. Because the formula divides by the actual value, any period where actual demand is zero produces an undefined result. This makes MAPE unusable for intermittent demand, which is common in spare parts, slow-moving inventory, or any dataset with periods of no activity. Even values close to zero can inflate MAPE dramatically, making one quiet period dominate the entire average.
Small actual values create a related issue. If your actual value is 2 and your forecast is 4, that’s a 100% error by MAPE’s math, even though you were only off by 2 units. In datasets where values fluctuate between small and large numbers, the small-value periods will disproportionately inflate MAPE and make accuracy look worse than it functionally is.
How MAPE Compares to Other Metrics
Mean Absolute Error (MAE) simply averages the raw error amounts without converting to percentages. If your forecasts are off by 5, 10, and 15 units, your MAE is 10 units. The advantage is simplicity and no division-by-zero risk. The disadvantage is that “10 units” means very different things depending on whether you’re selling 50 units or 50,000. MAE works well when you’re evaluating a single product or dataset where the scale is obvious.
Root Mean Square Error (RMSE) also uses raw units but squares the errors before averaging, then takes the square root. This penalizes large errors more heavily than small ones. If occasional big misses are more costly than frequent small ones, RMSE captures that better than MAPE or MAE. Like MAE, it’s scale-dependent, so comparing across different products or datasets requires extra context.
MAPE’s main advantage over both is comparability. Because it’s a percentage, you can meaningfully compare forecast accuracy across products with wildly different volumes. Forecasting 10,000 units of Product A with 8% MAPE is directly comparable to forecasting 50 units of Product B with 12% MAPE in a way that MAE and RMSE don’t allow.
Alternatives That Fix MAPE’s Weaknesses
Weighted Absolute Percentage Error (WAPE) addresses the small-value distortion by changing how the averaging works. Instead of giving every time period equal weight in the average, WAPE sums up all the absolute errors across periods and divides by the sum of all actual values. This means high-volume periods naturally carry more influence than low-volume ones. A 2-unit error when actual demand was 2 no longer dominates the calculation the way it does in standard MAPE. WAPE is popular in retail and supply chain forecasting for exactly this reason.
Symmetric MAPE (sMAPE) tries to solve the under-forecasting bias by dividing the error by the average of the actual and forecast values instead of just the actual. This makes overshooting and undershooting more equally penalized. However, sMAPE introduces its own quirks and isn’t truly symmetric in all cases, so it hasn’t fully replaced MAPE in practice.
The Mean Absolute Scaled Error (MASE) takes a different approach entirely. It compares your forecast errors to the errors you’d get from a naive forecast (one that simply uses the previous period’s value as the prediction). A MASE below 1 means your model is better than naive, above 1 means it’s worse. MASE handles zeros, doesn’t favor under-forecasting, and works across different scales, which is why it’s gained traction in academic forecasting competitions.
Using MAPE Effectively
MAPE works best when your actual values are consistently well above zero, your data doesn’t have extreme swings between tiny and large numbers, and you need a single percentage to communicate forecast quality to stakeholders who aren’t statisticians. It’s particularly useful for comparing accuracy across product lines or business units with different volumes.
Where it gets you into trouble is when you use it as the sole optimization target, because of the under-forecasting bias, or when your data includes zeros or near-zero values. In those cases, pairing MAPE with a complementary metric like MAE or WAPE gives you a more honest picture. Many forecasting teams report MAPE for executive communication while using RMSE or MASE internally to actually tune their models.