Demand Forecast Calculator

The arithmetic

The simplest useful forecast is a moving average of recent periods. Exponential smoothing improves on it by weighting recent data more heavily: forecast = α × last actual + (1 − α) × last forecast, where α between 0.1 and 0.3 is typical.

Adding a trend and a seasonal index gives the standard decomposition: forecast = base level × trend factor × seasonal index. Most practical forecasting for a product business is some version of that, and elaborate methods rarely beat it by much.

How that looks in practice

Recent weekly sales of 38, 44, 41, 40, 43 give a five-week average of 41.2. With α = 0.2 and a previous forecast of 40, exponential smoothing gives 0.2 × 43 + 0.8 × 40 = 40.6.

Now apply a seasonal index. If November historically runs at 1.6× the annual average and December at 2.1×, the November forecast is 40.6 × 1.6 = 65 a week and December is 85.

At a three-week lead time, the November stock has to be ordered in October against a forecast of 195 units for the lead time rather than the 122 that trailing demand would suggest, a 60% difference driven entirely by the seasonal index.

Where this breaks down

History that includes stockouts understates demand for exactly the products that sell best. Sales during an out-of-stock period are zero, the forecast learns from it, and the next order is smaller, a loop that quietly shrinks the best sellers.

Promotions distort it in the other direction. A week with a 30% discount is not a normal week, and leaving it in the history inflates the baseline for every period that follows.

Making it useful

Clean the history before forecasting from it. Mark stockout periods and estimate what would have sold; mark promotional periods and either exclude them or model them separately. That single step improves most forecasts more than any change of method.

Then measure the forecast error and use it, rather than only the forecast. Mean absolute percentage error tells you how much to trust the number, and the error is what sizes the safety stock.

Forecast error is the more useful output

Every forecast is wrong, and the practical question is by how much. Tracking absolute error over time gives a distribution, and that distribution is exactly what the safety stock calculation needs; it is a better input than a theoretical standard deviation of demand.

It also identifies where forecasting effort is worth spending. Products with low error need no attention; products with high error and high value are where a human looking at the number changes the outcome, and they are usually a small fraction of the catalogue.

The other use is honesty about horizon. Forecast error grows with lead time, so a supplier at twelve weeks requires a forecast that is materially less reliable than one at three, which means the buffer has to be larger for two separate reasons, and the case for shortening lead time is stronger than the pipeline stock alone suggests.