The forecast accuracy formula compares forecast demand with actual demand, period by period, and summarises the gap as a percentage. The most useful versions are WAPE (total absolute error divided by total actual demand) and bias (whether forecasts lean high or low). Together they tell planners how large the errors are and which way they point.
What does forecast accuracy measure?
Forecast accuracy measures how close forecasts were to what customers actually bought, over a defined set of items, locations and periods. It is reported as an error percentage, or as accuracy, which is 100% minus that error.
Two questions sit behind every result. How big were the misses, and did they lean in one direction? A forecast that is 10 units too high one week and 10 units too low the next has the same absolute error as one that is 10 units too high every week, but the second one is far more damaging, because it builds stock steadily. That is why accuracy and bias are always read together.
How is forecast accuracy calculated?
Start with the error for each period, then aggregate. Use the same definitions as the forecast accuracy calculator so the numbers agree.
Error = actual demand – forecast
- MAE (mean absolute error) = sum of |error| / number of periods
- MAPE = average of (|error| / actual) x 100, skipping periods where actual is zero
- WAPE = sum of |error| / sum of actual x 100
- Bias = (sum of forecast – sum of actual) / sum of actual x 100; positive means over-forecasting
- RMSE = square root of the average squared error
- Tracking signal = sum of errors / MAE; values far from zero suggest persistent bias
Accuracy is commonly reported as 100% minus WAPE. Whichever measure you choose, write the definition down next to the number, because “forecast accuracy” means different things in different reports.
Worked example with six periods
This example uses small, illustrative numbers. Actual demand over six weeks was 100, 120, 90, 110, 130 and 95 units. The forecasts were 105, 110, 100, 115, 120 and 100.
- Errors (actual – forecast): -5, +10, -10, -5, +10, -5.
- Absolute errors add up to 45, so MAE = 45 / 6 = 7.5 units.
- Total actual demand is 645, so WAPE = 45 / 645 = 6.98% and accuracy = 93.02%.
- MAPE averages the period percentages (5.00%, 8.33%, 11.11%, 4.55%, 7.69%, 5.26%) and comes to 6.99%.
- Total forecast is 650, so bias = (650 – 645) / 645 = +0.78%, a slight over-forecast.
- RMSE = square root of 62.5 = 7.91, and the tracking signal = -5 / 7.5 = -0.67.
The result: roughly 93% accurate with almost no bias. That is a healthy picture for this item, though six data points are too few to call it stable.
MAPE or WAPE: which should you report?
For most planning work, report WAPE. It weights errors by volume, so a miss on a fast mover counts for more than a miss on an item that sells two units a week. MAPE treats every period equally and breaks down when actual demand is zero or very small, since a tiny denominator inflates the percentage.
| Measure | Strength | Weakness |
|---|---|---|
| MAPE | Easy to explain; scale-free across items | Distorted by low or zero demand; cannot be calculated when actual is zero |
| WAPE | Volume-weighted; stable with small values | Hides performance on slow movers when pooled with fast ones |
| Bias | Shows direction of error | Errors can cancel out and hide large misses |
| RMSE | Penalises large misses | Expressed in units, harder to compare across items |
How do you read the result?
A good result depends on the product. Steady, high-volume items can be forecast much more accurately than intermittent, promotional or new items, so compare each item with its own history and with a simple baseline such as last period or a moving average. If a sophisticated method barely beats the baseline on identical history, keep the simpler one.
Three habits make the number more useful:
- Measure at the level where decisions are made, usually item and location, and also at aggregate level to see how errors net out.
- Measure over a horizon that matches the lead time. A forecast made one week out is not the one that drove a purchase order placed eight weeks earlier.
- Watch the tracking signal. If it keeps drifting in one direction, correct the bias before tuning anything else.
Forecast variability also feeds buffer sizing, so the error you measure here connects directly to safety stock. More on planning practice is in the demand planning guides.
Common mistakes to avoid
- Comparing the forecast with shipments instead of demand, which bakes past stockouts into the “actuals”.
- Mixing forecast versions, for example scoring the latest revision rather than the one used for the decision.
- Averaging item-level MAPE without weights, so a slow mover with 3 units counts as much as a bestseller.
- Reporting accuracy without bias, which hides a consistent lean that causes steady overstock or stockouts.
- Judging a model on too few periods or on a window that includes an unusual event.
- Changing the aggregation level between reports and then wondering why accuracy “improved”.
Where AI and analytics help
Machine learning models can add features such as promotions, price and weather that simple methods ignore, and they can forecast many items at once. They should be judged by the same measures on the same holdout history as any other method. Strong accuracy in testing does not guarantee strong accuracy on new products or after a market shift, so keep monitoring bias after launch.
Common questions
What is a good forecast accuracy percentage?
It depends on the product and how stable demand is. Steady, high-volume items can reach much higher accuracy than intermittent or promotional ones. A useful target is relative: beat a simple baseline on the same history, and improve on your own trend over time.
Should I use MAPE or WAPE?
WAPE is usually the better choice for planning. It weights errors by volume and copes with zero or small values. MAPE treats every period equally, which can overstate error on low-volume periods. Report the one your stakeholders understand, and state the definition.
What does a positive bias mean?
A positive bias means the forecast is higher than actual demand on average, which tends to cause excess stock. A negative bias means under-forecasting and a higher risk of stockouts. Bias is expressed as a percentage of actual demand over the period measured.
How many periods do I need?
At least a few full cycles of the demand pattern. Two periods is the technical minimum for the calculation, but a result from six data points is only indicative. For seasonal items, include at least one full year so the seasonal peak and trough are both tested.
What is the tracking signal for?
The tracking signal flags persistent bias. It is the cumulative sum of errors divided by MAE. A value that stays close to zero means errors balance out, while one that keeps growing in the same direction suggests the forecast needs correcting.