Demand forecasting is estimating how much of something you will sell over a future period, so you can decide how much to buy and hold.
Every business does it. Most do it implicitly — someone looks at the shelf, thinks about last month, and places an order. Formal forecasting replaces that with a method that can be examined, measured and improved.
The reason it matters is not that accurate forecasts are achievable. They are not, and any process that assumes otherwise will disappoint. It matters because the size of the error determines how much safety stock you need, and safety stock is expensive. Better forecasting is the only way to hold less stock without reducing availability.
What demand forecasting is
A forecast has three parts, and forecasts that go wrong usually omit two of them.
A number. Expected units over a defined period.
A period. Next week, next month, next quarter. Accuracy falls as the horizon lengthens, so the horizon should match what you need it for — a forecast for ordering purposes only needs to reach as far as your lead time plus your review period.
A confidence range. How wrong the number might be. A forecast of 500 units that could plausibly be anywhere from 400 to 600 is a very different planning input from one that could be 200 to 800, even though both have the same central figure. The range is what sizes your buffer, and a forecast without one gives you no basis for deciding how much protection to hold.
Forecasts also differ by level. A forecast for total category demand is far more reliable than one for an individual SKU in an individual colour, because variation cancels out as you aggregate. This is worth knowing because it explains why your top-level numbers can look accurate whilst individual product forecasts are consistently wrong.
The main forecasting methods
Naïve forecasting. Next period equals last period. Genuinely useful as a baseline — if your sophisticated method cannot beat it, the sophistication is not earning its keep.
Moving average. The mean of the last several periods, which smooths out random variation. Simple and effective for stable products, and slow to respond to genuine change because old periods carry as much weight as recent ones.
Weighted moving average. As above, but recent periods count for more. Responds faster to shifts whilst retaining some smoothing.
Exponential smoothing. Weights decline geometrically into the past, controlled by a single smoothing factor. A higher factor responds faster to change and is noisier; a lower one is stable and slow. This is the workhorse of practical forecasting, and variants handle trend and seasonality explicitly.
Trend and seasonal decomposition. Split history into an underlying level, a trend, and a repeating seasonal pattern, project each, and recombine. Necessary for anything genuinely seasonal, which is most consumer goods.
Regression. Relate demand to variables that drive it — price, marketing spend, weather, a leading economic indicator. Powerful where a genuine causal relationship exists, and misleading where a coincidental correlation is mistaken for one.
Qualitative methods. Sales team input, customer intentions, expert judgement. Indispensable where history does not exist — new products, new markets, a channel you have just opened. Also the most prone to optimism, which is why it is worth recording predictions and scoring them afterwards.
Machine learning approaches. Increasingly available in planning software, and genuinely better where you have many products, long histories and multiple influencing variables. They need substantially more data than most operations have, and they do not fix bad inputs.
Types of demand forecasting
Forecasts get categorised several ways, and the labels are worth knowing because people use them as shorthand in planning conversations.
By time horizon.
- Short-term — days to a few months. Drives replenishment, reorder points and staffing. Accuracy here is achievable and the payoff is immediate.
- Medium-term — a quarter to a year. Drives purchasing commitments, supplier agreements and seasonal buys.
- Long-term — a year or more. Drives capacity decisions: warehouse space, equipment, headcount. Precision is neither expected nor useful at this range; direction and magnitude are.
By scope.
- Macro-level forecasting looks outward at the conditions your demand sits inside — the economy, the category, the market. Useful for context and rarely actionable on its own.
- Micro-level forecasting looks at a specific product, channel or customer. This is where operational decisions are actually made.
By stance.
- Passive forecasting assumes continuity: what happened will keep happening, adjusted for trend and season. It suits stable businesses and is the right default for most inventory planning, because it makes no heroic assumptions.
- Active forecasting builds in what you intend to change — a promotion, a new channel, a price rise, a product launch. It is necessary whenever the past is a poor guide, and it is also where optimism does the most damage, because a plan gets mistaken for a prediction.
By method.
- Qualitative — judgement, sales input, expert opinion. Necessary for new products with no history, and the least reliable where data exists.
- Quantitative — moving averages, exponential smoothing, regression, seasonal decomposition. Reliable where history is available and demand is reasonably stable.
- AI-based and machine-learning — models that consider many variables at once and find patterns a simple time series would miss: weather, pricing, competitor behaviour, web traffic, promotional overlap. They earn their keep at scale, on large SKU counts with plenty of history, and they are not a substitute for clean data. A model trained on sales history that includes unflagged stockout periods learns the wrong thing faster and more confidently than a spreadsheet would.
The practical point: most businesses need short-term, micro-level, passive quantitative forecasting for the bulk of their range, with active forecasting layered on for the handful of lines where something is deliberately changing. Reaching for sophisticated methods before the basic one is working is the most common and most expensive sequencing error in this area.
How to forecast demand in practice
1. Clean the history first. This step determines whether everything after it is worth doing.
The critical correction is stockouts. Sales history records what you sold, not what customers wanted. A period where you were out of stock shows low sales, and feeding that into a forecast tells the model that demand is lower than it is — for exactly the product that already failed you. Flag stockout periods and either exclude them or estimate the lost demand.
Also strip one-off events: a single bulk order that will not repeat, a clearance, a data error. Leave genuine recurring patterns like promotions in, but mark them so they are understood as promotions rather than as baseline.
2. Choose the level. Forecast at the level you will act on. If you order by SKU, you need SKU-level forecasts, but you may build them by forecasting the category and splitting by recent share, which is often more stable than forecasting each SKU independently.
3. Pick a method per product group. Stable high-volume lines suit exponential smoothing. Seasonal lines need seasonal decomposition. Erratic slow movers are frequently best served by a simple average and a generous buffer, because sophisticated methods fit noise.
4. Apply judgement, and record it. Adjust for things the history cannot know — a listing win, a competitor exiting, a price change, a promotion. Record the adjustment separately from the statistical forecast so you can review later whether your interventions helped. Many do not.
5. Measure the error. Without this it is not a process, it is a habit.
6. Review and adjust regularly. Monthly for most operations. The point is to catch products whose pattern has changed, not to refine numbers that were fine.
Measuring forecast accuracy
Three measures cover most needs.
MAPE — mean absolute percentage error. The average error as a percentage of actual demand. Intuitive and widely used. Its weakness is that it distorts badly at low volumes: being two units out on a forecast of three is a 67% error that means nothing.
MAD — mean absolute deviation. The average error in units, ignoring direction. Useful because it feeds directly into safety stock calculation.
Bias. The average error with direction retained. This is the one most often skipped and frequently the most informative. If your errors average out near zero, you are as often high as low — noise, which buffers handle. If they are consistently positive or negative, you have a systematic problem, and no amount of safety stock fixes a forecast that is always low.
Measure by product group rather than in aggregate, because a good overall number routinely conceals a few badly forecast lines doing real damage.
Compare against the naïve baseline. If your method is not beating “same as last month”, simplify.
Why it’s never perfect (and that’s fine)
Forecasts are wrong. The useful questions are how wrong, in which direction, and whether the error is stable enough to plan around.
This is why forecasting and safety stock are the same conversation. The forecast gives you the expected demand; the forecast error gives you the variability that safety stock exists to absorb. Improve the forecast and the required buffer falls for the same service level. That is the entire commercial argument for investing in forecasting, and it is measurable.
Accept too that some products are genuinely unforecastable. Very slow movers, one-off project items, anything driven by external events nobody can anticipate. For these, the honest answer is a different strategy — make to order, hold supplier stock, accept longer lead times — rather than a better model.
Forecasting seasonal and promotional demand
Two patterns cause most of the difficulty, and both need handling explicitly rather than hoping a general method absorbs them.
Seasonality is a repeating pattern tied to the calendar. The mistake is treating a seasonal peak as growth. A moving average applied to a product selling four times as much in December will chase the peak upward in January, order far too much, and then chase it back down through spring.
Handling it properly means separating the seasonal factor from the underlying level. Calculate the ratio of each period’s demand to the annual average across several years, and you have a set of seasonal indices — December 2.4, January 0.6, and so on. Forecast the underlying level, then apply the index for the period you are forecasting.
Two cautions. You need at least two full cycles of history before seasonal indices mean anything, and three is better. And seasonality shifts — Easter moves, peak trading weeks migrate as buying habits change, and an index calculated from five-year-old data may be describing a pattern that no longer exists.
Promotions are the larger problem in most businesses, because they are both substantial and self-inflicted. A promotion can multiply demand several times over for a week, and that spike then sits in your history distorting every forecast that follows.
The disciplines that help:
- Flag promotional periods in the data so the statistical forecast can exclude or adjust for them.
- Forecast the promotional uplift separately from baseline demand, using the results of comparable past promotions rather than a percentage somebody feels is right.
- Account for the dip afterwards. Promotions frequently pull demand forward rather than creating it, so the following weeks run below baseline. Forecasting the peak whilst ignoring the trough leaves you over-stocked twice.
- Get the promotional calendar early. The most common cause of promotional stockouts is the warehouse finding out about the promotion at the same time as the customers.
Who owns the forecast
Forecasting fails organisationally more often than it fails technically, and the pattern is consistent.
Sales forecasts optimistically, because targets and forecasts get conflated and nobody wants to forecast below target. Operations forecasts conservatively, because they carry the cost of excess stock and the blame for stockouts. Finance wants a number that matches the plan. Each produces a different figure, and the business ends up running on whichever one belongs to the person who argued hardest.
What works better:
One number, owned by one function. Usually planning or operations, with structured input from sales rather than a separate sales forecast running in parallel.
Separate the forecast from the target. A target is what you intend to achieve. A forecast is what you expect to happen. Conflating them guarantees the forecast is biased upward, and the stock follows.
Score the adjustments. When someone overrides the statistical forecast, record it and review afterwards whether the override improved accuracy. This is uncomfortable and it is the fastest way to find out whose judgement is worth applying and whose is not.
Make the cost visible. Forecast error has a price — excess stock, write-offs, expedited freight, lost sales. Attaching that figure to the process is usually what turns forecasting from an administrative chore into something people take seriously.
How forecasting connects to the rest
- Reorder points use forecast demand over the lead time.
- Safety stock is sized from forecast error.
- Purchasing commits capital against the forecast, which is where errors become expensive.
- Warehouse capacity planning depends on expected volume, and getting it wrong shows up as no space in November.
- Labour planning follows expected order volume rather than units.
- Supplier relationships improve when you share forecasts, because they can plan production and hold appropriate stock.
That last one is underused. A supplier given visibility of your expected requirements can usually shorten and stabilise their lead time, which reduces your safety stock requirement more than a better model would.
A worked example
Take a product with twelve months of history, and work through what a forecast actually involves.
Monthly sales run roughly: 210, 195, 230, 240, 215, 225, 260, 250, 245, 280, 520, 310.
The immediate observation is that November at 520 is not part of the underlying pattern — it is a seasonal peak. Forecasting December from a three-month moving average of September, October and November gives (245 + 280 + 520) ÷ 3 = 348, which over-orders substantially because it treats the peak as the new normal.
Handling it properly:
Separate the seasonal month. Excluding November, the remaining eleven months average about 241 units with a mild upward trend. That 241 is your underlying level.
Calculate the seasonal index. November at 520 against a baseline of 241 gives an index of roughly 2.2. If a previous year showed a similar ratio, you can trust it; with only one year of history, treat it as provisional.
Forecast the level forward. The trend across the year is gently upward — roughly 200 at the start, 260 by autumn excluding the peak — so a reasonable level forecast for next month is around 265.
Apply the index if relevant. December is not the peak month here, so the forecast stays near 265.
Estimate the error. Month-to-month deviation from the underlying trend runs around 20 units, so a standard deviation near 20 is a reasonable working figure. At a 95% service level, safety stock is roughly 1.65 × 20 × the square root of the lead time in months.
The point of walking through it is that most of the work is judgement about what the history means, not arithmetic. Deciding that November is seasonal rather than growth is the decision that determines whether the next order is right, and no method makes that call for you on one year of data.
Where to go deeper
- What is safety stock — sized directly from forecast error.
- What is a reorder point — where the forecast is applied.
- Inventory management techniques — the wider toolkit.
- What is a stockout — the failure forecasting reduces.
- What is inventory turnover — what better forecasting improves.
- How to manage inventory — the process this sits inside.
The one-line takeaway
Demand forecasting estimates future sales so you can decide what to buy — and because the forecast will be wrong, the number that matters most is the size and direction of the error, since that determines how much safety stock you have to carry.
Frequently asked questions
What is demand forecasting? Estimating future demand for a product over a defined period, in order to decide how much stock to buy and hold.
What are the main demand forecasting methods? Moving averages, weighted moving averages, exponential smoothing, trend and seasonal decomposition, regression against causal variables, and qualitative judgement. Most operations use exponential smoothing for stable lines and seasonal decomposition for anything with a repeating pattern.
How accurate should a demand forecast be? There is no universal figure, because accuracy depends on how variable your demand is. The useful test is whether your method beats a naïve forecast of “same as last period”, and whether your error is unbiased rather than consistently high or low.
What is the difference between MAPE and bias? MAPE measures the average size of the error regardless of direction. Bias measures whether errors lean consistently one way. A forecast can have acceptable MAPE and serious bias, which is the more damaging problem because safety stock does not correct a systematic lean.
How do I forecast demand for a new product? History does not exist, so you rely on analogues — similar products at launch — combined with qualitative input from sales and any pre-order or enquiry data. Expect large errors initially and review weekly rather than monthly until a pattern emerges.
Why should I exclude stockout periods from history? Because sales during a stockout record availability rather than demand. Including them teaches the forecast that demand is lower than it really is, on exactly the products that already ran out.
How far ahead should I forecast? Far enough to cover your lead time plus your review period, and no further for ordering purposes. Longer horizons are needed for capacity, budget and supplier planning, but they should be understood as separate exercises with wider tolerances rather than as the number you order against.
Does better forecasting reduce inventory? Yes, indirectly and measurably. Safety stock is sized from forecast error, so reducing the error reduces the buffer needed for the same service level. This is the clearest financial case for investing in the forecasting process.
Warehouse Wise is a free, independent library of guides on inventory management, warehousing and fulfilment, written for the people actually doing the work — new warehouse and inventory staff, small and growing retail and ecommerce businesses, and anyone who holds stock and wants to understand it properly. Every guide explains the term, shows a worked example with real numbers where it helps, and stops. We are independent: we do not sell software or equipment, we take no commission on referrals, and our guides are not ranked by who pays us.