MAPE: Formula, Limits, and Better Alternatives

MAPE forecast accuracy, based on the mean absolute percentage error, is the most widely quoted accuracy measure and one of the most misleading. It is easy to calculate, easy to explain and easy to compare across items, which explains its popularity. It also breaks down when actual demand is zero, gives tiny items the same influence as giant ones and penalizes over-forecasting more than under-forecasting. A team that manages by MAPE alone can end up improving the metric while making inventory worse. This guide gives the formula, works through examples that show each weakness with checked arithmetic, explains when MAPE is still useful and recommends better alternatives for most consumer product portfolios. It is part of demand planning and inventory optimization, follows forecast accuracy metrics and demand planning KPIs, and leads to WMAPE and forecast bias. Accuracy is also measured in finance, as in measuring financial forecast accuracy, and the inventory consequences are covered in cash flow management.

The formula

For each item or period, calculate the absolute error divided by the actual, then average across items or periods and express as a percentage:

MAPE = average of ( |Forecast - Actual| / Actual ) × 100

A forecast of 110 against an actual of 100 has an absolute percentage error of |110 - 100| / 100 = 10 percent. If three items have percentage errors of 10, 50 and 5 percent, MAPE is (10 + 50 + 5) / 3 = 21.7 percent.

Weakness 1: it breaks at zero

If an actual is zero, the percentage error is undefined, because you divide by zero. Intermittent items, new items, and slow movers all have zero periods. Teams handle this by dropping those periods, which flatters the metric, by adding a small constant, which distorts it, or by switching to another measure. None of these is satisfying. If your portfolio has many zeros, MAPE is the wrong tool.

Weakness 2: small items dominate

Because every item counts equally, a tiny item with a large percentage error can swamp the result. Consider two items. Item A: actual 10, forecast 20, an error of 10 units, which is 100 percent. Item B: actual 1,000, forecast 950, an error of 50 units, which is 5 percent. MAPE is (100 + 5) / 2 = 52.5 percent. But in volume terms, the total absolute error is 10 + 50 = 60 units against total actuals of 1,010, a weighted error (WAPE) of 60 / 1,010 = 5.9 percent. The business impact of the large item's 50-unit miss is bigger than that of the small item's 10-unit miss, yet MAPE says the forecast is terrible because of the small item. A planner who "improves" MAPE by spending time on tiny items is working on the wrong thing.

Weakness 3: it penalizes over-forecasting more than under-forecasting

Percentage error is divided by the actual, so the penalty for over-forecasting is unbounded, while the penalty for under-forecasting is capped at 100 percent. If the actual is 100 and the forecast is 50, the error is 50 percent. If the actual is 50 and the forecast is 100, the error is 100 percent, even though the miss is the same 50 units. Check: |50 - 100| / 100 = 50 percent and |100 - 50| / 50 = 100 percent. A model tuned to minimize MAPE will therefore tend to under-forecast, which is the safe way to score well, and can produce a bias toward shortage. That is the opposite of what a retailer or customer wants.

Weakness 4: it ignores the cost of error

MAPE treats a 10 percent miss on a low-margin accessory the same as a 10 percent miss on a flagship item. It also ignores which direction hurts more in your business. If stockouts are costly, an under-forecast of 10 percent is worse than an over-forecast, and if the product is perishable or fashion, the reverse may be true. MAPE says nothing about this. An error measure should reflect the economics.

When MAPE is still useful

• Comparing one item's forecast across time when the item has steady, non-zero demand.
• Communicating with audiences who understand percentages and need a rough sense.
• Homogeneous portfolios with items of similar size and no zeros.
• Quick checks at an aggregate level, where small items are not an issue.

In those cases it is fine, provided you know what it measures and report it with bias.

Better alternatives

WAPE (or WMAPE): total absolute error divided by total actual. It weights by volume, handles zeros at the item level and aligns with the cost of inventory. For most portfolios it is the best replacement. See WMAPE.

Bias: total forecast minus total actual, over total actual. Always report alongside WAPE.

RMSE or the standard deviation of error: useful for safety stock because it emphasizes large errors.

MASE: scales error by that of a naive forecast, so it compares across items and tells you whether you beat the benchmark.

Symmetric MAPE: divides the absolute error by the average of the forecast and actual. It reduces the over-forecast skew but has its own problems, including instability near zero.

Cost-weighted error: multiplies error by unit cost or margin, so that expensive items count more. It aligns accuracy with the money at stake.

An example of a better report

Suppose a category has three items with actuals of 1,000, 300 and 10 units and forecasts of 950, 330 and 25. Errors are 50, 30 and 15 in absolute terms, a total of 95 against actuals of 1,310. WAPE is 95 / 1,310 = 7.3 percent. Bias is (1,305 - 1,310) / 1,310 = -0.4 percent. MAPE for the same data is (5 + 10 + 150) / 3 = 55 percent, driven entirely by the item that sold 10 units. The first two numbers describe the real position: accuracy is good and the forecast is balanced. The MAPE number would send the team chasing a 15-unit miss. Check: 50 / 1,000 = 5 percent, 30 / 300 = 10 percent, 15 / 10 = 150 percent; 950 + 330 + 25 = 1,305.

How to move away from MAPE

Run both measures in parallel for a few cycles and explain the differences. Agree on WAPE and bias as the new standard, with item-level segments. Keep MAPE in the dashboard for continuity if people insist, but do not set targets on it. Reset targets in the new measure, based on history. Document the definition and the level and lag of measurement.

Common mistakes

1. Excluding zero periods and reporting a flattering number.
2. Using MAPE as the sole KPI.
3. Setting targets by MAPE and encouraging under-forecasting.
4. Averaging across items of very different size.
5. Ignoring bias.
6. Comparing MAPE across companies with different portfolios.

FAQs

What is MAPE?

The mean absolute percentage error: the average of the absolute forecast error divided by the actual, expressed as a percentage.

Why is MAPE misleading?

It is undefined when actuals are zero, gives small items equal weight to large ones and penalizes over-forecasts more than under-forecasts.

What is a good MAPE?

It depends on the item mix and level of detail. Compare with your own history and a naive forecast, and prefer a volume-weighted measure.

What should I use instead of MAPE?

WAPE together with bias for most portfolios, and standard deviation of error for safety stock.

Can MAPE be over 100 percent?

Yes. When the forecast is more than double the actual, the percentage error exceeds 100 percent. There is no upper limit for over-forecasts.