Intermittent demand: Croston and SBA for rare sizes and accessories
From a series of sales with many zero periods estimates the expected demand per period by the Croston method or its SBA correction, sets the demand class (smooth, erratic, intermittent, lumpy), compares the accuracy with simple smoothing and shows the expected demand over the lead time.
At a glance
- What question does it answer?
- What demand to expect per week for an item that sells rarely and why that is better than ordinary smoothing.
- What data is needed?
- Sales by week with zeros (zeros while the item was in stock).
- How to read the result?
- Example on the default values (26 weeks, 7 with sales): ADI 3.71, demand is intermittent; expected demand 0.49 units a week, over the lead time 2 units, the probability of at least one sale 75%. RMSE 0.84 against 0.89 for simple smoothing. Example on the default values: ADI: 3.71; CV²: 0.19; Demand class: intermittent; Average demand size: 1.7. The figures are illustrative: enter your own to see your result.
Why and how to reason
Rare demand (a size, an accessory, an expensive item) sells not every period: zeros alternate with single sales. Ordinary smoothing behaves badly: after a sale the forecast jumps, then slowly falls and is wrong all the time. Croston splits the task in two: how many units in one sale (z) and after how many periods on average a sale occurs (p). Each estimate is updated only in periods with demand. The forecast per period equals z ÷ p, and in the SBA variant it is multiplied by (1 − α ÷ 2) to remove overestimation. By default the series has an interval between sales of 3.7 periods, an average size of 1.7 units, a forecast of 0.49 units a week. Over a lead time of 4 weeks expected demand is 1.95 units, and the probability of at least one sale 75%. The demand class is set by two numbers: ADI above 1.32 and CV² below 0.49 gives “intermittent”.
When to use it
Use it for sizes at the ends of the run, accessories, spare parts and anything that sells rarely and unevenly. Do not use it for regular demand, seasonal series and novelties without history. Assumptions: demand has no trend and seasonality, the series is long enough (dozens of periods), zero periods are real, not due to absence of goods. The method also suits computing safety stock for sizes, where the normal distribution clearly does not work.
How to read the result
First look at the demand class and ADI: if ADI is close to one, sales are not intermittent, and ordinary smoothing will do. “Expected demand over the lead time” (1.95 units) and the probability of at least one sale (75%) help decide whether to hold a unit in stock: if the probability is high and the unit price low, hold one, if low, buy to order. Compare the method’s RMSE with simple smoothing and the mean (a gain of 4.6%): a small gain means the method is not critical, and you can choose the simpler. Set safety stock for rare demand by probability, not by the normal distribution.
Common mistakes
Typical trap: using the method for regular demand, counting weeks with no stock as “zero demand”, taking α too large (the forecast jumps), judging the method by error on zero periods (it is always small). For a stock decision look at the probability of shortage, not the average error.
What data is needed?
- Sales by week with zeros (zeros while the item was in stock).
- The lead time in weeks.
- The method and the smoothing: SBA and 5–20% for rare items.
Inputs
| Field | Unit | Value in the example | Meaning |
|---|---|---|---|
| Method | SBA: corrected for over-forecasting | Classic Croston slightly over-forecasts demand; SBA multiplies the forecast by 1 − α ÷ 2 and is usually more accurate on rare items. | |
| Smoothing (α) | % | 15 | For rare items low values are taken, usually 5–20%: one random demand then does not swing the forecast. The example value is illustrative. |
| Lead time | periods | 4 | How many periods (weeks) before an order arrives; the expected demand and the probability of at least one sale are counted from it. The example value is illustrative. |
| ADI limit | periods | 1.32 | An interval between demands above the limit means “intermittent”. 1.32 is customary as a starting point; set your own. |
| CV² limit | 0.49 | The squared coefficient of variation of the demand size above the limit means “uneven size”. 0.49 is customary as a starting point; set your own. |
How it is calculated
| Step | Formula |
|---|---|
| Size and interval | in periods with demand: z = z + α × (demand − z), p = p + α × (interval − p) |
| Forecast | Croston: z ÷ p; SBA: (1 − α ÷ 2) × z ÷ p |
| ADI | periods ÷ periods with demand |
| CV² | (standard deviation ÷ mean)² over non-zero demand, divisor n − 1 |
| Probability of demand | 1 − (1 − 1 ÷ p)^lead time |
Worked numbers
Default values from the app; in the interactive version you enter your own.
- ADI
- 3.71Periods per demand: the number of periods ÷ the number of periods with sales.
- CV²
- 0.19The square of the ratio of the standard deviation (divisor n − 1) to the mean over non-zero sales.
- Demand class
- intermittent
- Average demand size
- 1.7
- Expected demand per period
- 0.49
- Expected demand over the lead time (4 per.)
- 2
- Probability of at least one sale over the lead time
- 75.2 %
- RMSE of the chosen method
- 0.84The root of the mean squared one-step error; on zeros WAPE misleads, so RMSE is used.
- RMSE of simple smoothing
- 0.89
- RMSE of the series mean
- 0.87
- Gain against simple smoothing
- 4.6 %
- Method bias
- 27.3 %
- Sold
- Forecast
Data as a table
| Period | Sold | Forecast |
|---|---|---|
| 1 | 0 | — |
| 2 | 0 | — |
| 3 | 2 | — |
| 4 | 0 | 0.617 |
| 5 | 0 | 0.617 |
| 6 | 0 | 0.617 |
| 7 | 1 | 0.617 |
| 8 | 0 | 0.543 |
| 9 | 0 | 0.543 |
| 10 | 3 | 0.543 |
| 11 | 0 | 0.598 |
| 12 | 0 | 0.598 |
| 13 | 0 | 0.598 |
| 14 | 0 | 0.598 |
| 15 | 1 | 0.598 |
| 16 | 0 | 0.507 |
| 17 | 2 | 0.507 |
| 18 | 0 | 0.546 |
| 19 | 0 | 0.546 |
| 20 | 0 | 0.546 |
| 21 | 0 | 0.546 |
| 22 | 1 | 0.546 |
| 23 | 0 | 0.468 |
| 24 | 0 | 0.468 |
| 25 | 2 | 0.468 |
| 26 | 0 | 0.488 |
Periods: demand and one-step forecast
| Period | Sold | Forecast |
|---|---|---|
| Week 1 | 0 | — |
| Week 2 | 0 | — |
| Week 3 | 2 | — |
| Week 4 | 0 | 0.62 |
| Week 5 | 0 | 0.62 |
| Week 6 | 0 | 0.62 |
| Week 7 | 1 | 0.62 |
| Week 8 | 0 | 0.54 |
| Week 9 | 0 | 0.54 |
| Week 10 | 3 | 0.54 |
| Week 11 | 0 | 0.6 |
| Week 12 | 0 | 0.6 |
| Week 13 | 0 | 0.6 |
| Week 14 | 0 | 0.6 |
| Week 15 | 1 | 0.6 |
| Week 16 | 0 | 0.51 |
| Week 17 | 2 | 0.51 |
| Week 18 | 0 | 0.55 |
| Week 19 | 0 | 0.55 |
| Week 20 | 0 | 0.55 |
| Week 21 | 0 | 0.55 |
| Week 22 | 1 | 0.55 |
| Week 23 | 0 | 0.47 |
| Week 24 | 0 | 0.47 |
| Week 25 | 2 | 0.47 |
| Week 26 | 0 | 0.49 |
Class “intermittent”: ADI 3.71 against the limit 1.32, CV² 0.19 against 0.49. Croston or SBA suit such demand; ordinary methods and WAPE on zeros distort the picture.
The forecast shows the average demand and not safety stock: the stock needs the spread of demand over the lead time. For footwear count by size, for accessories by article, for children’s with the season.
Nuances and limits
- Counting WAPE on series with zeros: the error is huge for any method.
- Taking zeros for periods with no stock on the shelf: that is not demand.
- Treating the average demand as safety stock: the stock is needed for the spread.