Forecast accuracy
How close the forecast was to what actually happened, measured as absolute error against actuals.
01 What it is
Why anyone looks at this number
How close the forecast was to what actually happened, measured as absolute error against actuals.
Every downstream plan — stock, staffing, cash — is built on the forecast. Its error is the real reason for safety stock, overtime and unused credit lines.
02 The formula
How it is worked out
Forecast accuracy = 1 − ( Σ|actual − forecast| ÷ Σactual ) grain : month × product or account unit : percentage source : Planning system compared against the ERP actuals
Lock the forecast version by date. Comparing actuals against a forecast that was revised last week measures nothing and always scores well.
03 Worked example
The same number, with real inputs
| Forecast | 4,000 units |
| Actual | 3,600 units |
| Calculation | 1 − (400 ÷ 3,600) |
| Result | 88.9% |
11% error on a four-week lead time is roughly the safety stock you are carrying. Improving accuracy is usually cheaper than holding the stock that hides the error.
04 What moves it
Four things that actually change this number
Forecast horizon
Accuracy at one month ahead and six months ahead are different metrics.
Level of aggregation
Everything is accurate at group level; the pain is at SKU and location.
Promotions and events
Unplanned demand spikes that nobody told planning about.
Bias
Persistent over- or under-forecasting, which is a behaviour, not an error.
05 Where the number lives
The system, the record and the fields
| System of record | Key record | Fields you need |
|---|---|---|
| Planning system compared against the ERP actuals | Forecast version joined to Sales order / Shipment | forecast_qty, actual_qty, forecast_version, lag |
Lock the forecast version by date. Comparing actuals against a forecast that was revised last week measures nothing and always scores well.
06 How it goes wrong
Three ways this metric misleads people
Measuring only bias
Over-forecasting one month and under-forecasting the next averages to zero error and hides both.
Fix: Measure absolute error for accuracy and signed error for bias, separately.Aggregating up until it looks good
National monthly accuracy is meaningless if you replenish weekly by location.
Fix: Measure at the level you actually make decisions.No locked version
Continuous revision means you always measure yesterday's forecast against today's actual.
Fix: Snapshot the forecast at the decision date and score that one.08 Questions
Frequently asked
What is a realistic forecast accuracy?
It depends entirely on the level and the horizon. High-volume staples one week out can exceed 95%; new products at SKU level six months out may be under 50%. Compare like with like, and to your own trend.
Which error measure should we use?
Weighted absolute percentage error is the most common and is hard to game, because large-volume misses count more than small ones. Avoid simple percentage error at SKU level, where small denominators produce meaningless numbers.
One definition, everywhere it is used
SCIKIQ stores this metric once and serves it to every dashboard, board pack and agent that asks.