An inventory forecast is an estimate of how many units a product will sell in the coming weeks and months. It is used to decide when to order and how much. The forecast will never be exact, but it can be good enough to get it right on the products that matter most.
Start with the history
Before choosing a method, the sales history has to be reliable. Three things distort it more often than anything else:
- Stockout days. When the product is out of stock, it sells nothing, and the period looks like weak sales. Read the guide to stockout days.
- Promotions. A week with 30% off can sell as much as a whole month, and those sales do not come back on their own.
- One-off orders. A B2B customer who buys 200 units once is not part of the regular sales.
It is usually a good idea to clean out these three things before calculating further. Otherwise even a good method will be skewed.
Four common methods
| Method | How it calculates | Suits |
|---|---|---|
| Moving average | The average of e.g. the last 3 months | Stable products without seasonality |
| Exponential smoothing | A weighted average where the most recent months weigh the most | Products where the level shifts slowly |
| Seasonal profile | A level multiplied by an index per month | Products with a fixed season, e.g. Christmas or summer |
| Croston/SBA | Calculates how often the product sells, and how much when it does | Products that sell rarely and unevenly |
Moving average
This is the method most spreadsheets use. If a product has sold 40, 52 and 46 units over the last three months, the forecast is 46 units per month. It is easy to explain, but it reacts slowly and does not see seasonality.
Exponential smoothing
Here the most recent month gets the most weight, and older months weigh less and less. The weight is controlled by a parameter (alpha) between 0 and 1. With alpha 0.3 and a previous forecast of 46 units, the next forecast after sales of 60 units is:
0.3 × 60 + 0.7 × 46 = 50.2 units
A high alpha follows changes quickly but also reacts to noise. A low alpha is steadier but lags behind.
Seasonal profile
If a product sells twice as much in December as in an average month, the forecast for December should also be twice as high. This is handled with a seasonal index per month, which is multiplied onto the product’s level. The approach is described in the guide to seasonality.
Croston and SBA for products that sell rarely
A spare part that sells 0, 0, 3, 0, 0, 1 units per week fits poorly with an average, because the average never matches an actual week. Croston splits sales into two numbers: the size of a sale when it happens, and how many periods pass between sales. The forecast per period is size divided by interval.
Croston overestimates slightly, so the Syntetos-Boylan variant (SBA) is often used instead, which multiplies the result by 1 − alpha/2. With alpha 0.1, 5% is deducted.
Choose a method per product
A method that suits the entire range is rare. A practical way to choose is to test: let each method calculate the last three months as if they had not happened yet, and compare with actual sales. The method that comes closest for the product is used from then on. How to measure this is described in the guide to forecast accuracy.
It is also worth looking at how much history there is. With less than a year of sales you cannot see the product’s own seasonality, and then it is probably better to borrow the seasonality from the product’s category than to guess.
From forecast to order
The forecast alone does not say what to order. It has to be combined with lead time, safety stock, incoming goods and the supplier’s minimum order. That is the subject of the guides on reorder point, safety stock and order quantity.
The forecast in lesslater
lesslater cleans stockout days and promotion days out of the history and uses the product’s own seasonality, or the category’s if the history is short. If sales grow or fall from year to year, this is taken into account, so a trend is not read as seasonality. Four methods are tested per product, and the seasonal profile is used unless another method was much more accurate over the last three months. Below each number it shows how it was calculated.