Skip to content
Predict.ai

Accuracy & evaluation

Choosing a forecast accuracy metric

Pick the metric that matches the cost of being wrong. WAPE for volume, MAE when units are the pain, RMSE when big misses hurt more than small ones, bias when you always run heavy or light.

Updated Aug 19, 2026·10 min read

The metric is a job description

You do not “choose accuracy.” You choose which misses are allowed to dominate the meeting. Percent errors care about relative miss. Unit errors care about volume. Squared errors care about the rare disaster. Bias cares about the lean. Pick the one that matches the cost, then stop shopping when the bake-off gets uncomfortable.

Pick the pain, then the metric

MetricUse it whenBe careful when
WAPEVolume plansZeros, mixed SKU sizes
MAEEvery unit costs the sameComparing different scales
RMSERare spikes are the disasterNoisy series, vanity 'stats'
MASEMany series, one leaderboardWrong seasonal period
BiasAlways heavy or lightUsing it alone
A metric is a job description. If you optimize one and report another, you will promote the wrong champion and argue about it quarterly.

Percent errors

MAPE is the celebrity. WAPE is the one that survives a catalog. sMAPE was invented because MAPE kept exploding; there are two formulas, so paste yours in the footnote. None of these replace a signed bias. You can look “12% accurate” and still bleed cash if you are always heavy on yogurt.

Unit errors

MAE is the average miss in the original units. Ops can feel it. RMSE squares first, so a 10-unit miss is more than twice a 5-unit miss. Use RMSE when the tail is the product (an SLA, a stockout that takes down a line). Use MAE when every extra case is roughly the same headache.

Scale-free scores

MASE divides your MAE by a naive seasonal baseline. Below 1 means you beat last week (or last year). It is how you compare a big series and a small one without a fake percent. Get the period right — daily often 7, monthly 12 — or you will crown a model for beating a stupid baseline.

Don’t skip bias

Mean signed error, or percent bias. Slice it. Models that look unbiased at chain level are often drunk on one brand. You can debias with a multiplier if the lean is stable. If it flips every month, you have a regime, not a constant.

A practical pick

Volume plans: WAPE + bias. Unit-cost worlds: MAE + bias. Many series on one board: MASE, then glance at WAPE on the A items. Tail risk: look at a high quantile or RMSE, not only the average. And run a seasonal naive on the same folds so “good” has a floor.

FAQ

Which forecast accuracy metric should I use?
If the series has zeros or mixed scales, start with WAPE plus bias. If every unit costs the same, MAE is honest. If a rare spike is the disaster, look at RMSE or a tail pinball loss, not a single average.
What is MASE?
Mean Absolute Scaled Error. It divides your MAE by the MAE of a naive seasonal baseline. Below 1 means you beat 'same as last week/year.' It is handy when you compare series of different sizes.
Should finance and ops share one metric?
They should share a dashboard, not necessarily one number. Finance may care about revenue WAPE; ops may care about units on the A items. One metric that makes both groups shrug will not change a PO.

Keep going

Try it on your data.

Connect an outcome and its history. Predict.ai finds the drivers, races the models, and keeps the forecast live — in the workspace, over the API, and through agents.