How much will we sell in the coming weeks, so we order the right stock?

How it works

The shop. An online coffee shop that sold €6.0 million in the last 52 weeks, about €501,000 a month, up 12.8% on the 52 weeks before. It is the same shop as the market basket demo, with its five categories and their prices. Its weekly sales run for three years, from 2 October 2023 to 27 September 2026.

The demand. Each category grows by a share a year and has a season: a gift peak in mid-December, a drop in the last week of December when parcels stop, more sales in the cold months, most of all for tea and coffee, and a quiet January for the gear. On top sit a slow wander in the level (marketing, competitors, the weather), one-off weeks such as a newsletter that took off or a week out of stock, and ordinary weekly noise.

The forecast. Holt-Winters exponential smoothing: a level, a trend and an index for each of the 52 weeks of the year, multiplied together. Each week, the level moves toward the week’s sales by a weight, and the trend and the indices can move by weights of their own. The weights are picked from a grid of 550 on the first two years, by how well each set forecasts one week ahead. In every category of this shop, the level moves 5% of the way each week, and the trend and the season hold as the first two years set them.

The test. The last 52 weeks are held out: the weights never see them. Each held-out week is forecast as many weeks before it as you set, and set against last year’s amounts, the units sold in the same week a year earlier. At 13 weeks ahead, the forecast’s weekly error runs from 6.0% to 9.3% across the categories, and last year’s amounts’ from 7.6% to 18.2%. The forecast’s is lower in every category, however many weeks ahead. Last year’s amounts carry last year’s level and its one-off weeks into this year; the forecast averages two years of season and follows the level.

The money. Both ways of ordering are held to the same service level: each gets the smallest buffer on top of its numbers that left enough stock in the share of the 52 weeks you set. Then each week’s stock left over is counted at cost, 60% of the price, and each week’s sales missed at the price, added up over the 52 weeks. A higher service level leaves more stock over and misses less. Across the shop, at every setting the controls reach, ordering to the forecast leaves less stock over and never misses more in sales, because it misses by less.

The range and the stock to order. The 80% range around each week ahead comes from the same 52 weeks: how far the forecasts made that many weeks ahead missed, from the 6th to the 47th of the 52 misses, so 8 weeks in 10 sat inside it. The stock to order is the forecast times the buffer for your service level at that distance ahead; at 90% it is the top of the range.

What this shows, and what it does not

What it does not. No dated events: Black Friday, a promotion or a price change moves sales on a date that a 52-week season does not hold, so a forecast on your data takes them as inputs. Each week’s stock is counted on its own: what is left at the end of a week counts as left over, not as sold the week after, so for stock that keeps, the left-over figure is cash tied up, not a loss. No supplier lead times, minimum orders or pack sizes, and categories rather than products.

On your own data. The model is fitted to your order history, product by product, with your promotions calendar. Each week the forecast is set against what sold, and the misses size the next buffer. Anything you act on, such as the peak-season buy, carries the analytics director’s sign-off.