A demand planner receives a sales file and sees a sudden dip in the third quarter. Should the entire annual forecast be lowered? First ask what the figures measure. Units shipped are not always the units customers wanted. During a stockout, recorded sales can be a lower bound on demand rather than a reliable estimate of it. That distinction determines whether the dip should change the baseline or be investigated as a constrained observation.
Example
A time series is a sequence of observations at consistent intervals. Trend is a sustained movement in the level. Seasonality is a pattern recurring at the same position in a known cycle, such as a repeated quarterly rise. An exceptional event is a temporary influence with its own explanation. One high observation does not establish seasonality; recurrence and context matter. Tag promotions, shortages and changes in product definition before attributing movements to customer preference. With a short history, use a provisional assumption and state its uncertainty instead of fitting an elaborate model to a few points.
Consider a hypothetical store item. Sales in the four quarters of year one were 80, 100, 120 and 160 units. Year two recorded 88, 110, 90 and 220. Operations reports five days out of stock in quarter three; marketing reports a price promotion in quarter four. The first two matched-quarter comparisons each show 10% growth: 88/80−1 = 10%, and 110/100−1 = 10%. We can use 10% as a first working assumption for underlying demand growth, not as an established law. Applied to the third quarter, it gives 132 units. The 42-unit difference from recorded sales of 90 is not proof of 42 lost orders. The growth assumption may be wrong, and some customers may have bought later. Record 132 as a conditional baseline, then inspect unfilled orders, website demand and stock availability to estimate the shortfall.
The fourth-quarter baseline under the same assumption is 160×1.10 = 176. The observed 44 units above that level might be associated with the promotion, but a before-after difference alone cannot establish its causal effect; purchases could have shifted forward from the next quarter. For a no-promotion scenario, 176 is a clearer starting point than 220, with a wide uncertainty range because the history is short. For a repeat promotion, record its timing, channels and terms and test the effect separately. Preserve the raw values 90 and 220 with stockout and promotion flags; use a separate analytical series so another planner can reproduce the reasoning.
Recorded sales are what the firm fulfilled, not necessarily what customers wanted when stock was unavailable. To estimate unmet demand, examine abandoned baskets, backorders or visits with stock timestamps. Compare similar days rather than the best peak day and retain a range: some buyers switch products or defer purchase. After a promotion, check whether the lift pulled sales forward from later weeks instead of representing lasting category growth.
Sales record fulfilled transactions, not necessarily all demand during a stockout. Examine backorders and abandoned baskets against stock timestamps, then compare similar days rather than peak days. Retain a range because buyers may substitute or defer. After a promotion, test whether a spike pulled sales forward before treating it as permanent growth. Keep the raw sales series with stockout and campaign flags so the analytic adjustment can be reproduced.
