Quantitative Methods

Forecasting Demand and the Limits of a Time Series

Learn what a demand forecast can and cannot tell you, then stop ordering against the point forecast, which is the single most expensive habit in operations.

  • Advanced
  • 13 min total
  • 14 chapters

What decision this helps you make: What quantity actually goes on the purchase order, and whether the forecasting model you are paying for beats the free benchmark it must clear.

What this topic is

Demand forecasting is the estimation of future demand from history: its trend, its seasonality, and whatever else you can observe. It is expressed as a distribution over outcomes rather than a single number. A time series model learns from the past of the series itself. That is its strength, because seasonality and momentum are genuinely repeatable, and it is its limitation, because it can only extrapolate causes it has already observed.

Why it matters

Almost every operational commitment a business makes is a bet on a forecast: how much stock to order, how many staff to roster, how much capacity to reserve, how much cash to hold. Two failures dominate. The first is a model that never beat the free benchmark (last year, same week) while consuming a licence fee and a team. The second, larger and more universal, is ordering to the point forecast. That is a conditional average, and therefore the wrong quantity whenever the cost of being short differs from the cost of being long. That is most of the time.

Who should learn it

Operators who place orders or set capacity against a forecast, planners choosing between forecasting approaches, and anyone being asked to fund a demand-planning system.

What you will understand

  • How to backtest a forecast honestly with a rolling origin, and which accuracy measure not to use
  • Why forecast error grows with the square root of the horizon, and what that does to an eight-week order
  • How to convert a forecast distribution into an order quantity using the cost of being short against the cost of being long
  • The four things a time series cannot see, and what to do about each

Prerequisites

Common misconception

"The forecast says 1,240 units, so order 1,240." A point forecast is an estimate of the conditional mean. Ordering to the mean gives you roughly a coin flip on being in stock, and it silently assumes that a unit short costs the same as a unit long. In most businesses it does not. A lost sale costs the whole contribution margin plus some probability of losing the customer, while an extra unit costs holding and eventual markdown. When those two numbers differ, the correct order quantity is a specific quantile of the forecast distribution, not its centre, and the gap between the two is often 20% of the order.