Demand forecasting: methods, accuracy, and how to choose

What demand forecasting actually is

A demand forecast is a number, or a series of numbers, describing what you expect to sell over a future period. Its only job is to make a better purchasing decision: how much to order, when, and from whom. A forecast that nobody acts on is a spreadsheet decoration.

Two things make a forecast useful rather than decorative. It has to be produced per product, because products behave differently, and it has to be measurable, because otherwise you cannot tell progress from luck.

The main methods, honestly compared

• Moving average. Averages the last n periods. Simple, transparent, and slow to react to a trend or a step change.
• Simple exponential smoothing. A weighted average that leans on recent data. Good for stable demand, and the starting point for most spreadsheet forecasts.
• Holt and Holt-Winters. Adds trend and seasonality. Strong on growing product lines and genuine seasonal peaks, and the method most businesses should be using but are not.
• Croston and its variants. Designed for intermittent demand: products that sell nothing for weeks and then sell a batch. Standard smoothing handles these badly, which is why slow movers so often look unforecastable.
• ARIMA and regression models. Use the structure of the series and any outside variables. Powerful, and easy to misconfigure in untrained hands.
• Machine learning. Gradient boosting and similar methods can capture complex patterns, particularly with promotional or pricing signals, at the cost of explainability.
• Naive and seasonal naive. Repeat the last period, or the same period last year. Unfashionable, hard to beat on some seasonal lines, and a useful benchmark that stops fancy methods claiming false victories.

The trap: one method for the whole catalogue

Most tools pick one method, or a small handful, and apply it everywhere. That is a comfortable engineering decision and a poor commercial one. A distributor with 4,000 lines has fast movers, slow movers, seasonal lines, new lines and lines killed by a supplier change. One method cannot be right for all of them, and the damage is invisible because nobody reports per-product accuracy.

How to measure whether a forecast is any good

Two numbers matter. Error tells you how far off you typically are. Bias tells you which direction you are wrong in, which is the more expensive of the two because it repeats.

• MAPE expresses average error as a percentage. Easy to read, though it misbehaves when actuals are near zero.
• WMAPE weights error by volume, so a fast mover's mistake counts more than a slow mover's. Usually the fairer business measure.
• Bias is the average of (forecast minus actual). A persistent positive bias means you are chronic over-forecasters and buying too much; negative means you are stockout-prone.
• Back-testing means hiding the most recent history, forecasting it, and comparing the forecast to what actually happened. Without it, every accuracy claim is marketing.

Why we run a tournament instead of picking a method

Optimal Chain makes 18 forecasting methods compete for every single product. Each one is tested against that product's own history with the recent past hidden, so the winner is the method that actually predicted this product, not the method that looks good in a brochure. You see the winner and the scores, which means you can defend the number to a finance director.

It also changes how seasonal products are treated. Traditional classification calls a seasonal line erratic because its sales wobble, then drops it to a crude fallback. Graded on predictability instead, a genuine seasonal pattern is one of the easiest things to forecast. The wobble was the signal, not the problem.

A practical way to choose, whatever tool you use

• Start with a benchmark. Seasonal naive, meaning the same period last year, is harder to beat than most vendors admit. If a method cannot beat it, it is not adding value.
• Measure per product, not in aggregate. A headline accuracy figure hides the 800 lines being forecast badly. Ask for the per-product view.
• Watch for stockout-contaminated history. Weeks with no sales because you had nothing to sell look like zero demand. Feed that in and the forecast learns to under-buy.
• Recalculate regularly. Products move from growing to declining. A method that won last year may not win this quarter, and the tournament re-runs every time you refresh the data.
• Connect the forecast to a decision. Accuracy is a means, not an end. The test is whether the forecast leads to better orders, less cash tied up and fewer stockouts.

How it compares, honestly

Frequently asked questions

What is demand forecasting?

Demand forecasting is predicting what you will sell over a future period, so you can decide what to buy, how much and when. A good forecast is produced per product and measured against what actually happened, because otherwise there is no way to tell whether it is helping.

Which demand forecasting method is most accurate?

There is no single answer, and that is the important point. A method that wins on a stable best-seller can lose badly on an intermittent spare part. The reliable approach is to test several methods against each product's own history and use the winner for that product, rather than imposing one method across a catalogue.

What is a good MAPE?

It depends on the product. Fast-moving, stable lines can sit in single digits; lumpy or intermittent lines are often in the twenties or thirties and still be forecastable enough to plan with. Benchmark against your own history rather than a generic figure, and watch bias as closely as error.

How much sales history do I need?

Two years or more is ideal, because it captures a full seasonal cycle and one repeat. Twelve months still works well, and it is enough to beat a spreadsheet. Under a year, seasonal methods have little to learn from, so expect wider error bars.

Why test 18 methods rather than one good one?

Because inventory is not uniform. Fast movers, slow movers, seasonal lines, lumpy demand and brand-new lines fail in different ways, and each responds to a different method. Testing all of them per product costs almost nothing and removes the guesswork about which one to trust.