Forecasting starts with the decision it serves, not with the most elaborate algorithm available. How many units will be needed next month, at which location, and with what procurement lead time? This lesson compares two short-horizon numerical baselines and separates error size from bias. A naïve forecast sets the next period equal to the most recent observation. A three-period average smooths fluctuations but can lag a rising series. Both are useful benchmarks; neither dominates every demand pattern.
Example
Test a method on periods whose outcomes were unavailable when its forecasts were made. Using a later month's value inside an earlier forecast leaks future information into the evaluation. Keep a time-stamped record of each forecast, the data available at issue time, its horizon, and the subsequent actual. Define signed error here as actual minus forecast. Repeated positive errors mean underforecasting; repeated negative errors mean overforecasting under this convention. Mean absolute error averages the magnitudes of the differences and retains the item's unit. State the sign convention because some organizations reverse it.
Worked example: actual demand in months one to six is 100, 110, 120, 130, 140 and 150 units. Compare forecasts issued before month five, then updated before month six. The naïve forecast for month five is 130, giving an error of 140−130 = +10. Once 140 is observed, the month-six forecast is 140, giving 150−140 = +10. Mean absolute error is 10 units and mean signed error is +10: this method underforecasts in the small test window.
The three-period average for month five is (110+120+130)/3 = 120, with an error of +20. Before month six, use values then known: (120+130+140)/3 = 130, with an error of +20. Mean absolute error is 20 and mean signed error is +20. On this short window the naïve benchmark is better by error magnitude, but two periods cannot establish general superiority. A simple trend method might be worth testing, using fresh held-out periods rather than repeatedly tuning on the same test observations.
Why not rely on percentage error in every case? If actual demand is zero, dividing by the actual is undefined; near zero, the percentage can exaggerate a small unit difference. For procurement of one item, mean absolute error is directly interpretable in units. However, raw unit errors from an item selling 10 units and one selling 10,000 should not be compared as if their scales were the same. Nor should a stockout-constrained sales value be treated as unconstrained demand in an accuracy calculation; flag such periods and review them separately.
Compare forecasting methods on periods excluded from tuning, with information available at each forecast date. A positive actual-minus-forecast error indicates underforecasting, but signed errors can cancel and hide poor accuracy. Use absolute error for size, signed error for bias and item-level checks for critical products rather than one average across unlike items. A price or channel change can invalidate last season's model.
Compare methods on held-out periods using only facts available at each forecast date. Positive actual-minus-forecast error signals underforecasting, but signed errors can cancel. Use absolute error for size, signed error for bias and separate checks for critical items. Price, channel or stockout changes can break historical comparability. A forecast is conditional on information, not a permanent accuracy score.
