What it means
Almost every operating decision in a business depends on an expectation of demand: purchasing buys materials for expected orders, production schedules to expected volumes, the warehouse holds stock for expected sales, the call centre rosters staff for expected calls, and finance forecasts cash from expected revenue. The demand forecast is the single assumption on which these plans rest, and errors in it propagate through every one of them.
Forecast too high and the business carries excess inventory, idle capacity and cash tied up in stock; forecast too low and it runs out, loses sales, expedites at premium cost and disappoints customers, so forecast accuracy is an economic quantity and improving it is worth money. Forecasting methods fall into three families.
Qualitative methods rely on judgement (sales force estimates, expert panels, customer surveys and structured techniques such as the Delphi method) and are used where there is little history, for new products, or where the future is expected to differ from the past. Time-series methods project the pattern in past demand forward: moving averages smooth out noise, exponential smoothing weights recent periods more heavily, and decomposition separates level, trend and seasonality so each can be projected.
Causal methods model demand as a function of the factors that drive it, such as price, advertising, weather, economic activity or the customer's own sales, using regression and related techniques. In practice a good process combines them: a statistical baseline from the time series, adjusted by causal factors where known and overridden by judgement where there is specific information the statistics cannot see.
The forecast must also be produced at the right level of detail and horizon for the decision it supports, since a factory scheduling next week needs a forecast by product and day while a board planning capacity needs one by product family and quarter over three years. Forecasts at aggregate levels are more accurate than forecasts at detailed levels, because errors in individual items partly cancel out, and forecasts for the near future are more accurate than for the distant future.
A business typically maintains a hierarchy of forecasts reconciled to each other and reforecasts on a rolling basis as new information arrives. Measuring accuracy is what turns forecasting from guesswork into a managed process, using mean absolute percentage error, which shows the typical size of the error regardless of direction, and bias, which shows whether the forecast is systematically too high or too low.
Bias is the more damaging fault, because a consistently optimistic forecast builds inventory month after month, and it usually has an organisational cause such as sales teams forecasting to their targets or managers reluctant to forecast a decline. Tracking accuracy by product, forecaster and method identifies where the process is weak, and comparing judgemental overrides with the statistical baseline shows whether the judgement is adding value or subtracting it.
The financial connection runs through inventory, service and capacity. Safety stock, the buffer held against demand exceeding the forecast, is set by the variability of forecast error, so the less accurate the forecast, the more safety stock and cash are needed for a given service level, while lost sales from under-forecasting cost the gross margin and sometimes the customer, and capacity built for an over-forecast is depreciated whether or not it is used.
A sales and operations planning process, in which sales, operations and finance agree a single forecast and the plans that flow from it, is the standard way of connecting the forecast to the decisions and holding the organisation to one set of numbers.
In practice
Real-world examples
Example 1
A retailer uses last year's daily sales, adjusted for the calendar and for planned promotions, to forecast store replenishment, and reviews the forecast for each store every morning.
Example 2
A software company forecasts new subscriptions from its sales pipeline weighted by stage, and finds that pipeline forecasts made by sales managers run 20% above the outcome.
Example 3
A utility forecasts electricity demand hour by hour from weather forecasts, day of the week and historical load, and buys power ahead on the result.
Think of it
Demand forecasting is predicting what customers will buy-educated guessing about future needs.
Formula
Calculation
Moving average forecast = Average of the last n periods' actual demand
Exponential smoothing forecast = Previous forecast + Alpha x (Previous actual minus Previous forecast), where alpha is between 0 and 1
Seasonal forecast = Base level x Trend factor x Seasonal index
Mean absolute percentage error (MAPE) = Average of (absolute error / Actual) x 100
Bias = Average of (Forecast minus Actual)
Safety stock = Service level factor z x Standard deviation of demand x Square root of lead time
Worked example: methods. A product sold 1,200 units in January, 1,300 in February and 1,400 in March.
- Three-month moving average forecast for April = (1,200 + 1,300 + 1,400) / 3 = 1,300 units; note that the average lags the rising trend
- Exponential smoothing with alpha 0.3: if the March forecast was 1,250 and actual was 1,400, the April forecast = 1,250 + 0.3 x (1,400 - 1,250) = 1,295 units
- Seasonal: average quarterly sales are 1,200 units and fourth-quarter sales average 1,800, giving a Q4 seasonal index of 1.5; with underlying growth of 5% a year, next Q4's forecast = 1,200 x 1.05 x 1.5 = 1,890 units
Worked example: accuracy. Over three months, forecasts were 90, 130 and 105 against actuals of 100, 120 and 110.
- Percentage errors: 10 / 100 = 10%; 10 / 120 = 8.3%; 5 / 110 = 4.5%; MAPE = (10% + 8.3% + 4.5%) / 3 = 7.6%
- Bias = ((90 - 100) + (130 - 120) + (105 - 110)) / 3 = -1.7 units a month, a slight under-forecast on average
Worked example: safety stock. Weekly demand has a standard deviation of 200 units around the forecast, the supplier lead time is four weeks and the target service level is 95% (z = 1.65).
- Safety stock = 1.65 x 200 x square root of 4 = 1.65 x 200 x 2 = 660 units
- If forecast improvement halved the standard deviation to 100 units, safety stock would fall to 330 units, releasing the cash tied up in 330 units of inventory
Worked example: cost of error. A distributor sells $30,000,000 a year and buys about $18,000,000 of goods on its forecast. Over-forecasting purchases by 10% builds $1,800,000 of excess inventory, costing about $360,000 a year to carry at 20%. Under-forecasting by 5% loses 5% of $30,000,000, or $1,500,000, of sales, which at a 40% margin is $600,000 of lost gross profit. Both errors are expensive, and the asymmetry between them determines whether the business should lean high or low.Case study
Seen in the real world
This is a fictional, illustrative scenario. A beverage company with revenue of $120,000,000 relied on its sales force for demand forecasts. Each regional manager submitted a monthly forecast by product, and the forecasts were aggregated for production planning. They ran about 15% above actual sales, because the managers forecast to their targets rather than their expectations, and the production plan followed.
The result was inventory of $24,000,000, regular write-offs of short-dated stock and, oddly, frequent stockouts of the fastest-selling lines, because the excess had been made in the wrong products. Forecast error at product level was 28%. The company introduced a statistical baseline forecast, built from three years of history with seasonal indices and trend, as the starting point for every product. Regional managers could adjust it, but each adjustment needed a reason (a new listing, a promotion, a lost customer) and its accuracy was tracked against the baseline.
A monthly sales and operations planning meeting brought sales, production and finance together to agree one forecast and one plan, and within three months the tracking showed that adjustments helped only where they related to specific known events. A year later, error had fallen from 28% to 14% and the bias had gone. Inventory fell from $24,000,000 to $18,000,000, releasing $6,000,000 of cash and saving about $1,200,000 a year in carrying cost, while stockouts fell from about 3% of revenue to about 1%, worth roughly $1,080,000 a year of gross profit at a 45% margin. The finance director's summary was that the company had not become better at predicting the future; it had stopped confusing its targets with its forecasts.
Watch out
Common mistakes
- Forecasting to targets rather than to expectations, which produces a consistent upward bias, excess inventory and a plan built on hope.
- Producing forecasts without measuring their accuracy, so that nobody knows whether the process is improving or whether judgemental adjustments help or harm.
- Forecasting at too detailed a level for the decision, where errors are largest, instead of forecasting at the aggregate level and allocating down.
Questions
People also ask
What is the difference between a forecast and a target?
A forecast is an unbiased estimate of what will happen; a target is what the business wants to happen. Using one as the other corrupts both: a target used as a forecast builds inventory for sales that do not arrive, and a forecast used as a target removes any stretch.
Which forecasting method is best?
None in general. Time-series methods work well for established products with stable patterns; causal methods where demand is driven by measurable factors; judgement where there is no history or where the future will differ from the past. Good processes combine a statistical baseline with disciplined judgemental adjustment, and measure which is adding value.
How accurate should a demand forecast be?
It depends on the product and horizon. Stable, high-volume products can be forecast within a few percent a month ahead; promotional, seasonal or new products may carry errors of 30% or more. The useful question is whether accuracy is improving and whether inventory and service levels are set with the actual error in mind.