What Demand Forecasting Is and How to Do It Step by Step

What demand forecasting is

Demand forecasting is the estimate of how much your customers will buy of each product in each future period, made from sales history and from what the business knows is going to happen. It is a number per item and month (or week), with a horizon of several months ahead, and it is recomputed every cycle.

Three clarifications that avoid misunderstandings:

• It is demand, not sales. If you ran out of product last month, you sold less than you were asked for. Demand includes what you could not serve.
• It is an estimate with error, not a promise. A good forecast is not the one that hits, it is the one that misses little, in a known way and without bias.
• It is the input to planning, not the result. With the forecast you decide how much to buy, produce and store; without it, those decisions are made by eye.

What it is for (and what happens without it)

The forecast feeds four decisions: purchasing (how much to order and when), production (what to make and in which sequence), inventory (how much safety stock to hold per item) and finance (sales budget and working capital). In an S&OP process it is the starting point of the monthly cycle.

Without a formal forecast, each area makes its own: purchasing repeats last month\'s order, sales promises what it wants to sell and finance budgets what it needs. The result is familiar: stock-outs on what moves, overstock on what does not, and meetings where people argue about which number is right.

Step 1: the data

A forecast is worth what its data is worth. The minimum:

• Sales history by item, customer (or channel) and month, ideally 24 to 36 months. In units, not only in money: money mixes price and quantity.
• Item master with the hierarchies you use: family, category, brand, and substitution relationships (which item replaced which).
• Periods with stock-outs, to correct sales in those months and not teach the model that demand fell.
• Known events: past promotions, closures, shortages, anomalous months. They are flagged so the model does not read them as seasonality.

Before modeling, clean: duplicate items, inconsistent units, returns recorded as negative sales, codes that changed. It is the step nobody wants to do and the one that adds the most accuracy.

Step 2: level and horizon

Level is the combination of dimensions at which you forecast: item by customer, item total, family by region. The finer the level, the noisier the series and the worse the forecast per series; the more aggregated, the more accurate but less useful for deciding per item.

The usual practice is to forecast at the level where the series is stable and break it down with the historical weight of each item and customer. An aggregation analysis tests the combinations and says which one gives the least error after breaking down; there is no need to guess.

Horizon is how many periods ahead you need. It is set by the longest lead time in your chain plus the planning cycle: if the supplier takes three months and you plan monthly, you need at least four or five months of horizon with reasonable accuracy. Beyond that, the forecast serves budgeting, not purchasing.

Step 3: classify demand

Not all items are forecast the same way. Before choosing a method, each series is classified by its pattern:

Pattern	What it looks like	Which method fits
Smooth	Sells every month, similar quantities, maybe with seasonality	Exponential smoothing, Holt-Winters, ARIMA
Erratic	Sells every month, very variable quantities	Simple smoothing, aggregation to family, robust averages
Intermittent	Many zero months, similar quantities when it sells	Croston, SBA
Lumpy	Many zero months and very variable quantities	SBA, aggregation, service-based inventory policy

Classification is also crossed with maturity: a launch has no history and is forecast with profiles of similar items; a declining item needs a method that notices it is dying. Intermittent demand has its own article: why the forecast fails and what to use.

Step 4: choose the method per item

Here is the most important mindset change: you do not choose a method for the company, you choose one per item. The way to do it is called a model tournament or validation-based selection:

1. The last months of the history (say, six) are set aside and hidden.
2. Each candidate method is fitted on the rest of the history and forecasts those six months.
3. Each forecast is compared with what actually happened and the error is computed.
4. For that item, the method with the lowest error wins. It is refitted on the full history and used to forecast the future.

The usual candidates are the moving average, exponential smoothing in its variants, Holt-Winters, ARIMA, Croston and SBA for intermittent items, and machine learning models when there are many items or external variables. Running the tournament by hand for 500 items is unfeasible; that is what tools automate.

Step 5: measure the error

The forecast is always measured, every cycle, against actual demand, with at least two indicators:

• WAPE (weighted absolute percentage error): sum of absolute errors divided by sum of demand. It says how far off you are, weighted by volume.
• Bias: sum of (forecast minus actual) divided by sum of demand. It says whether you always miss in the same direction. Persistent positive bias is excess inventory; negative, stock-outs.

They are measured by item, by family and in total, and with the lag used by whoever consumes the plan (the forecast made a month ago, three months ago). The formulas and their traps are in MAPE, MAD and bias. And something that seems obvious and almost nobody does: compare against a naive method (the average of the last three months). If your forecast does not beat it, the effort is not adding value.

Step 6: adjust with the business and close the cycle

The model does not know that next month there is a promotion, that a new customer came in or that a competitor closed. That knowledge is contributed by sales and marketing in an adjustment step, with two rules: adjust at the level where the information exists (the promotion of a family, not of each item) and record the adjustment apart from the statistical forecast, with author and reason.

That way, at cycle close, you can measure whether the adjustment improved or worsened the result. In many teams, discovering that adjustments make the forecast worse is the first big improvement. Then the approved plan is frozen: it is the snapshot of that cycle, and next month it is compared against actuals and the process starts again.

Full example

A hardware products distributor with 1,200 SKUs and 40 wholesale customers wants to forecast the next six months.

1. Data: 36 months of sales by item and customer exported from the ERP. 14 months with stock-outs on 60 items are corrected (sales are replaced by estimated demand) and two months of closure for physical inventory are flagged.
2. Level: the aggregation analysis shows that item by customer is too noisy (most series are intermittent) and that item total, broken down to customer with the weight of the last 12 months, gives 30 % less error after breakdown.
3. Classification: 640 smooth items, 180 erratic, 310 intermittent, 70 lumpy. Almost a third of the catalog needs intermittent methods.
4. Tournament: with six months hidden, Holt-Winters and exponential smoothing win the smooth items; SBA wins most of the intermittent ones; lumpy items are forecast by family and broken down.
5. Measurement: WAPE of 24 % at item level and 11 % at family level, against 38 % and 19 % for the previous method (three-month average). Total bias of +2 %, with no trend.
6. Adjustment and close: sales raises the power tools family 15 % for an October promotion; the adjustment is recorded. The plan is approved and frozen; in November it will be measured whether the promotion sold as expected.

The example figures are illustrative, but the proportions (a third intermittent, the improvement from choosing the level, the drop in error from an average to a tournament) are what we frequently see in distribution catalogs.

Common mistakes

• Forecasting in money. It mixes price and quantity; a price increase looks like demand.
• Using sales without correcting stock-outs. The model learns that demand fell when what was missing was product.
• One method for everything. The moving average that works for the star item destroys the intermittent one.
• Not hiding data to validate. A model that "fits well" on the full past can forecast terribly; only validation on hidden data reveals it.
• Adjustments that overwrite the model. Without a separate record, nobody knows who was right.
• Not measuring, or measuring with average MAPE. Error is computed every cycle with WAPE and bias, against a naive method, and published.
• Updating without closing cycles. If March\'s forecast is lost when recalculating in April, you will never know how wrong you were.