Demand forecasting: practical methods for industry and retail

Almost 8 out of every 100 items that should be on sale in Brazilian retail are missing because someone ordered too little or the order arrived late. Abrappe, the Brazilian loss-prevention association, calls this commercial stockout, and its loss survey published in 2026 put it at 7.81%. That isn’t theft or breakage. It’s demand forecasting done badly, or done by gut feeling.

The odd part is that the same company losing sales to empty shelves usually has cash tied up in products that don’t move. Both problems share a root: purchasing decisions are made by looking at last month and the buyer’s instinct, with no reference number anyone can challenge.

The good news is that most companies don’t need a sophisticated model to break that cycle. They need the right method for the data they already have.

Demand forecasting is the estimate of how much of each product will sell in a future period, built from sales history and the factors that move it (seasonality, price, promotions). Methods range from a simple moving average in a spreadsheet to machine learning models. It is the basis for purchasing, production and safety stock decisions.

Why demand forecasts are always wrong (and why that isn’t the problem)

Every forecast misses. A manager who expects the exact number will abandon the method by month three. The useful question is different: how much does it miss, in which direction, and what does the miss cost?

Being 10% high on a cheap, fast-moving item costs little. Being 10% low on the item that brings customers into the store costs a lot, because a lost sale rarely comes back: the customer buys from a competitor. That’s why the forecast doesn’t end with a number. It ends with an inventory policy that absorbs the error, which we cover further down.

In practice, the biggest gain doesn’t come from switching methods. It comes from having any method at all, written down, with its error measured every month. That alone separates companies that learn from companies that repeat the same guess.

What are the demand forecasting methods?

Methods fall into two groups. Qualitative methods rely on structured judgment: sales team estimates, expert panels (the Delphi method), customer surveys. They’re for product launches, new markets or any situation without history. Quantitative methods use past sales data, and that’s where day-to-day purchasing lives.

The most practical way to choose is to think in maturity steps. Each step needs a bit more data and discipline, and delivers a bit more accuracy.

Method	When to use	Where it breaks
Simple moving average	Stable items, little history, first step	Lags behind trends and ignores seasonality
Exponential smoothing	Many SKUs, monthly or weekly updates	On its own, it also lags when sales are rising
Seasonal index (Holt-Winters)	Retail with strong dates, industry with harvest cycles	Needs at least 2 years to measure the pattern
Causal regression	Promotions, price and weather move sales a lot	Requires logging the causes, which almost nobody does
Machine learning	Thousands of SKU x store combinations, integrated data	Dirty data produces wrong forecasts that look precise

How to forecast demand in practice: a worked example

Picture a neighborhood grocery store with this history for cases of bottled water: January 820, February 760, March 900, April 940, May 880, June 1,010. How much should it order for July?

3-month moving average

Add the last three months and divide by three: (940 + 880 + 1,010) / 3 = 943 cases. Simple, easy to explain in a meeting, and already better than repeating June. The flaw shows when sales have a direction: the average always trails the trend.

Exponential smoothing (α = 0.3)

Here, next month’s forecast is 30% of the latest actual sales plus 70% of the previous forecast. Starting with February’s forecast equal to January’s sales (820), the sequence runs: March 802, April 831, May 864, June 869, and July 911 cases (0.3 × 1,010 + 0.7 × 869).

Notice the two methods give different numbers for the same series. Neither is "right". A higher α reacts faster to change and to noise; a lower one is steadier and slower. The honest way to choose is to run both over past months and see which missed less, which is exactly what the next section shows.

Adjusting for seasonality

If two years of history show December sells, on average, 40% above a typical month, December’s seasonal index is 1.4. A baseline forecast of 1,050 units becomes 1,050 × 1.4 = 1,470 units. It’s the adjustment that prevents the most stockouts on peak dates, and our guide on preparing an online store for Black Friday with data is worth a read if you live through that spike.

How to measure forecast error

Two metrics cover almost everything. MAPE (mean absolute percentage error) tells you how big the error is. Bias tells you which way you keep missing.

Month	Forecast	Actual	Absolute error %
1	1,100	1,000	10.0%
2	1,050	1,200	12.5%
3	1,000	900	11.1%
4	1,150	1,100	4.5%

MAPE is the average of the errors: 9.5%. Bias adds up forecast minus actual: +100 units over 4,200 sold, or +2.4%. A persistent positive bias means the buyer pads the number "just in case", and that cushion turns into dead stock. In our experience, bias is the first indicator that changes behavior, because it reveals a habit, not chance.

Be careful with MAPE on slow movers: missing by 2 units when actual sales were 4 gives a 50% error. For the long tail, evaluate by product family, not by SKU.

From forecast to purchase order

A forecast alone doesn’t tell you when to buy. The bridge is the reorder point, which adds expected consumption during the supplier’s lead time to a safety stock that covers the error.

In the example: forecast demand of 30 units per day, error deviation of 8 units per day, a 9-day lead time and a target of covering 95% of cycles without a stockout (z = 1.65). Safety stock is 1.65 × 8 × √9 ≈ 40 units. The reorder point is 30 × 9 + 40 = 310. When stock drops below 310, the order goes out.

Note that forecast error feeds straight into the math. Cutting the deviation from 8 to 5 lowers safety stock from 40 to 25 units without touching the service level. That’s why measuring error isn’t red tape: it’s cash that stops sitting on the shelf.

This formula assumes reasonably stable demand and a reliable supplier lead time. If the supplier is often late, lead-time variability has to enter the calculation too, and the buffer grows considerably.

Demand forecasting in manufacturing vs retail: what changes

The methods are the same; the cut is different. In manufacturing, forecasts are usually made by product family over a horizon of months, because they feed the production plan and long-lead raw material purchases. The ritual that ties it together is S&OP, the monthly meeting where sales, operations and finance agree on a single number.

In retail, granularity is SKU by store and the horizon is days or weeks. The typical problem is the empty shelf: Neogrid measured an 11.9% stockout rate in Brazilian supermarkets in September 2025, according to its stockout report. Part of that is forecasting and part is in-store replenishment, and separating the two causes already tells you where to act. Our post on data analysis in retail covers more uses of this kind of data.

A mistake common to both: forecasting from sales during periods when the item was out of stock. If a product sat empty for ten days, recorded sales understate real demand. Flag those periods and correct them before calculating, or the model learns to under-order and the stockout perpetuates itself.

When AI earns its place in demand forecasting

Machine learning beats classic methods when many variables interact (promotions, competitor prices, weather, regional holidays) across thousands of product and store combinations. At that point, tuning index by index in a spreadsheet stops being viable. We explain how it works in sales forecasting with AI and, more broadly, in predictive analytics.

Now the trade-off: an AI model fed duplicated product records, sales without stockout flags and unlogged promotions produces a wrong forecast that looks very precise. For most small and mid-sized businesses, the best-return path is to clean up the data, run smoothing with seasonality, measure error for three months, and only then test a more sophisticated model against that baseline. If the AI can’t beat the simple method on historical data, it isn’t worth the cost.

Frequently asked questions

What is the difference between demand forecasting and sales forecasting?

A sales forecast estimates what the company will bill, often already factoring in targets and stock constraints. A demand forecast estimates how much the market wants to buy, even if you don’t have the product. The difference shows up precisely during stockouts, when demand existed and sales didn’t.

How do I forecast demand in Excel?

Lay out sales by product and month in columns, compute the moving average with AVERAGE over the last three cells and drag the formula down. For exponential smoothing, use =0.3*previous_sales+0.7*previous_forecast. Recent Excel versions also include FORECAST.ETS, which already accounts for seasonality.

What are the qualitative demand forecasting methods?

The main ones are sales force opinion, the Delphi method (rounds of expert consultation until views converge), customer intent surveys and analogy with similar products. They suit launches and markets with no history, and work best when combined with data as soon as the first sales come in.

Which method should I use with little history?

With less than six months of data, use a short moving average or an analogy with a similar product, and review weekly. Seasonality should only be estimated with at least two complete cycles, meaning two years for annual patterns.

What is an acceptable forecast error?

It depends on the item’s turnover and value. For fast movers, a MAPE between 10% and 20% is usually workable; for long-tail items, per-SKU error will be high and it’s better to evaluate by family. More important than the absolute number is a falling error trend and a bias close to zero.

Buyers don’t need a crystal ball

They need a reference number, a measured error and a rule that turns both into an order. Companies that do this with a moving average and discipline buy better than companies that hired a sophisticated model and never checked the results. The bottleneck is almost always getting sales, inventory and promotions in one place, and that’s where Sherlok helps: you connect your sources, ask in plain language which product is about to run out or which item is piling up, and get an analysis ready for this week’s purchasing decision.