Forecast accuracy metrics look like math. They are actually policy. The metric you pick decides which errors the planning team chases and which ones it tolerates, and that decision shapes inventory, service levels, and working capital for the entire fiscal year.
Most teams pick MAPE because it is the default in every textbook and every planning tool. For a meaningful share of those teams, MAPE is the wrong choice and the team has been optimizing for the wrong thing for years without noticing.
This is a working guide to the three metrics that matter — MAPE, WAPE, and MAE — and how to choose between them for a real demand portfolio.
The three metrics in one paragraph each
MAPE (Mean Absolute Percentage Error) averages the percentage error across every forecast period and every SKU. A 50-unit forecast that comes in at 40 units is a 20 percent error. A 10,000-unit forecast that comes in at 9,500 units is a 5 percent error. MAPE treats those two errors as equally important and reports the simple average across the portfolio.
WAPE (Weighted Absolute Percentage Error) weights every error by the actual volume of that period. The 10,000-unit miss matters two hundred times more than the 50-unit miss in the final number. WAPE is what you get when you sum the absolute errors across the portfolio and divide by the sum of the actuals — it answers “what percent off was the total volume forecast?”
MAE (Mean Absolute Error) is the average absolute error in units. No percentages, no weighting. A 10-unit miss is a 10-unit miss whether the actual was 50 or 50,000. MAE is reported in the unit of the thing being forecast — cases, eaches, dollars — and is most useful when you are looking at a single SKU or a tight cluster of similar SKUs.
Why MAPE is the default — and why that is a problem
MAPE became the default because it is unit-free, so you can compare error across SKUs with very different volumes. A 12 percent MAPE on a slow-mover and a 12 percent MAPE on a fast-mover look like the same number, which makes portfolio dashboards easy to read.
That readability is a trap. MAPE has two well-known pathologies that punish exactly the kind of demand portfolios most planning teams actually run.
First, MAPE blows up on small-volume periods. A SKU that sells two units in a week, forecast at zero, contributes a 100 percent error to the average. A SKU that sells zero units, forecast at one, contributes an infinite error and is usually quietly dropped from the calculation, which biases the metric without anyone noticing. If you have a long tail of slow-movers — and most CPG, pharma, and retail portfolios do — MAPE will report your worst accuracy on the SKUs that produce the least revenue, and you will spend planning cycles fixing them.
Second, MAPE treats every SKU as equally important. The 80/20 rule applies to almost every product portfolio, which means roughly 20 percent of SKUs produce 80 percent of the revenue. MAPE averages across all of them. A team chasing a lower MAPE will spend the same effort improving forecasts on a SKU that does $50,000 a year and one that does $5,000,000 a year, because both contribute one observation to the average.
If the goal of the forecast is to drive inventory and revenue decisions, MAPE is misaligned with the goal. The number it reports is not the number the business cares about.
Why WAPE fixes the misalignment
WAPE weights every error by actual volume, so a miss on a high-volume SKU dominates the metric and a miss on a slow-mover barely registers. WAPE answers a question the business actually asks: “across our whole portfolio, what percent of total volume did we miss?”
For most operating teams, that is the right question. Inventory exposure, service-level risk, and revenue forecasting all scale with volume. A 5 percent WAPE on a $100 million portfolio means $5 million of forecast error to absorb across safety stock and operational buffers. That number maps cleanly to working capital planning. A 5 percent MAPE on the same portfolio is a much fuzzier statement because it depends on how the portfolio is weighted in the average.
WAPE is also stable in the presence of slow-movers and zero-volume periods. The denominator is the sum of actuals, which dominates any individual small-volume noise.
For demand planning teams that report up to a finance audience, WAPE is almost always the right primary metric.
When MAE earns its keep
MAE has a narrower role but a real one. Use MAE when you are diagnosing a single SKU or a tight cluster of similar SKUs — same channel, same volume band, same seasonality profile — and you want to see the actual error in the unit you ship in. Cases, eaches, kilos, dollars.
MAE is also the right metric when you are tuning a forecasting model on a single series and want to compare algorithms head to head without the noise of percentage-based metrics. Whether the best algorithm for a given series is a classical statistical method or a machine-learning one is its own question — we cover it in ML vs statistical forecasting. The 45-algorithm selection inside the Geneva Forecast API uses error metrics in this style under the hood — pick the algorithm with the lowest absolute error on the held-out portion of the series.
MAE is the wrong metric for portfolio-level reporting. The unit answer does not aggregate across heterogeneous SKUs in a way that means anything to an executive.
A practical metric stack
Most operating teams should run two metrics, not one.
WAPE at the portfolio level. This is the headline number for the S&OP review, the demand review, and the finance handoff. It scales with volume, it is stable, and it maps to dollars.
MAE at the SKU diagnostic level. When a planner drills into a problem SKU, the absolute error in units is the right diagnostic. It tells the planner how big the actual miss is and whether the safety stock policy is absorbing it.
MAPE has a narrow legitimate use: comparing forecast accuracy across teams or business units where the SKU mix is so different that volume weighting would mask process differences. Even then, treat the MAPE number as a directional signal, not a target.
The bias trap
None of the three metrics catch forecast bias. A forecast that runs 5 percent low every period has the same MAPE, WAPE, and MAE as a forecast that runs 5 percent high every period — and the inventory consequences are opposite. High-bias forecasts produce stockouts. Low-bias forecasts produce overstock.
Run a fourth metric in parallel: forecast bias, computed as the signed mean error divided by the signed mean actual. A bias close to zero means the forecast is centered. A persistent positive or negative bias means the model is wrong in a direction, and no amount of WAPE optimization will fix the inventory exposure that creates. We cover detection and correction in depth in how to detect and correct forecast bias.
Most planning teams report accuracy and ignore bias. The teams that report both find roughly a third of their portfolio sitting on a structural bias they had not noticed.
Run a real accuracy test
The fastest way to see whether your current metric is masking accuracy problems is to run your last twelve months of forecasts and actuals through all four metrics and compare. If WAPE and MAPE disagree by more than a few points, your portfolio has a volume-weighting story your current dashboard is hiding.
The Geneva Forecast API returns the chosen algorithm and confidence interval for every forecast, and the free tier is enough to test a SKU sample against your current forecast in an afternoon.
For teams that want a structured assessment — current forecast accuracy, bias, and the dollars at risk in the gap — Contact RoadMap for a forecast accuracy diagnostic. We do these in two weeks, with your data, and you keep the analysis whether or not we work together after.
Leave a Reply