What a forecast is for
A demand forecast is an estimate of how much of a product customers will want over some future period. Almost every supply chain decision leans on one: how much to buy, how much to make, how many workers to schedule, how much space to book on a truck or ship. Because most of these decisions must be made before demand is actually known, a forecast is the best available guide, not a promise.
Start with the data you have
Most simple methods begin with history, such as weekly sales of an item over the past year. A few cautions apply before any math.
- Sales are not always demand. If an item was out of stock for a week, sales fell, but demand did not necessarily. Using raw sales can understate what customers wanted.
- One-off events distort history. A big promotion, a one-time bulk order or a disruption can mislead a method that assumes the past was normal.
- The level of detail matters. Forecasts for a whole product family tend to be more reliable in percentage terms than forecasts for a single item at a single store, because ups and downs partly cancel out in larger groups.
Patterns to look for
Demand history often contains a mixture of:
- Level: the typical amount sold.
- Trend: a steady rise or fall over time.
- Seasonality: a repeating pattern tied to the calendar, such as higher sales before holidays.
- Random variation: noise that cannot be explained.
Good forecasting tries to capture the first three and accept the last. Retailers dealing with surges around the peak season, for example, usually treat seasonal patterns separately from the underlying trend.
A few simple methods
Naive forecast. Predict that next period will equal the last one. It is crude, but it makes a useful benchmark: a more complicated method should beat it to be worth the trouble.
Moving average. Average the last few periods. Suppose, hypothetically, weekly demand for an item was 100, 120, 110 and 130 units. The three-week moving average for next week uses the last three values: (120 + 110 + 130) divided by 3 is 360 divided by 3, or 120 units. Averaging smooths out noise, but it reacts slowly to real changes and ignores any trend.
Exponential smoothing. Give recent periods more weight than older ones. A simple version updates the forecast like this: new forecast equals alpha times the latest actual plus one minus alpha times the old forecast. With alpha of 0.2, an old forecast of 120 and an actual of 130, the new forecast is 0.2 times 130 plus 0.8 times 120, which is 26 plus 96, or 122 units. A higher alpha reacts faster but is also more easily thrown off by noise.
Seasonal methods. These adjust a base forecast using an index for each part of the year. If an item typically sells 20 percent above average in one month, the forecast for that month can be raised accordingly, provided the seasonal pattern has been reliable.
Judgment and market input. Sales teams, buyers and customers know things the history does not show: a planned promotion, a competitor’s launch, a large contract. Combining statistical forecasts with informed human input often works better than either alone, though human adjustments can also add bias if they are not tracked.
Why forecasts are always wrong
No method can know the future, so forecasts carry error. Customers change their minds, competitors act, weather shifts and supply problems interrupt sales. Even a well-built forecast should be expected to miss by some amount in each period.
The goal is therefore not a perfect forecast but a useful one, along with an understanding of how big the misses typically are. A plan that assumes the forecast is exactly right will be brittle.
Measuring forecast error
Common ways to judge a forecast include:
- Forecast error: actual minus forecast for each period.
- Mean absolute error: the average size of the errors, ignoring direction.
- Percentage error: error as a share of actual demand, helpful for comparing items of different size, though it behaves poorly when demand is near zero.
- Bias: whether errors lean consistently in one direction. A forecast that is always too high leads to excess stock, and one that is always too low causes shortages.
Tracking these over time helps a team see which methods work for which products. Fast sellers with a stable pattern are often easier to forecast than new products or items with lumpy demand.
Living with uncertainty
Since forecasts miss, businesses build in protection. Extra buffer inventory, covered under safety stock, absorbs unexpected demand. Shorter lead times and flexible suppliers reduce how far ahead a business must guess. Regular planning meetings, such as sales and operations planning, let teams revise forecasts as news arrives.
Larger companies increasingly apply advanced analytics and machine learning to forecasting, and these tools can pick up patterns that simple methods miss, such as the effect of price or weather. They still rely on good data, they still produce errors, and their output benefits from human review.
Practical habits
- Keep a record of past forecasts so accuracy can be checked.
- Start simple, and add complexity only if it demonstrably helps.
- Forecast at a level of detail that matches the decision being made.
- Treat forecasts as ranges where possible, not single numbers.
- Revisit assumptions when something significant changes.