Reader question: how to count safety stock for rare sizes when demand is not normal
An answer to the question of how to set stock for a size that sells a few units a month and does not look like the “bell” of a normal distribution at all: why the ordinary formula with a service level coefficient under-protects, how to count stock by the Poisson and the negative binomial distributions and what service level you really get. The result: the stock for the protection period by size and an honest estimate of the probability of running out of a size.
Why this document
For fast sizes the standard safety stock formula with a normal distribution works. For rare sizes demand is discrete: 0, 1, 2 units, and the normal approximation gives errors and fractional “stocks”. The planner, the buyer and the category manager read this document when deciding how many extreme sizes to hold. It shows how to count stock by a distribution that matches the nature of demand and how to check that stock by Poisson does not mislead if demand fluctuates more strongly than Poisson. Without this either one unit is put everywhere “just in case”, or exactly the size that was asked for is missing.
At a glance
- What is this document and why is it needed?
- An answer to the question of how to set stock for a size that sells a few units a month and does not look like the “bell” of a normal distribution at all: why the ordinary formula with a service level coefficient under-protects, how to count stock by the Poisson and the negative binomial distributions and what service level you really get. The result: the stock for the protection period by size and an honest estimate of the probability of running out of a size.
- Who owns the document and how often is it rebuilt?
- The planner and the analyst count, the buyer approves the stock by size; recalculated once a season and when the lead time or demand changes.
- What mistakes are the most common?
- Brand: zeros for weeks without stock are counted → demand is understated → take the weeks in stock. Retail: the normal formula is applied to an average of 1–2 units → fractional stocks and a wrong service level → count by a distribution. Brand: Poisson is taken without checking the variance → during promotions and spikes the stock protects worse than declared → compare the variance with the mean.
For whom
- Brand
- A brand decides how many extreme sizes to hold in the warehouse for reorders and partners and what share of the series to buy. The main things: the protection period equals the lead time plus the review period, and the target service level by size is chosen by the price of being out of stock.
- Retail
- A retailer decides whether to hold an extreme size in every store or centrally. The main things: the display minimum by size, replenishment from the warehouse and a redistribution rule between stores.
What it consists of
Why the normal formula under-protects
- The normal approximation fits when the average demand over the protection period is large (tens of units); at an average of 1–3 units the distribution is skewed to the right, with zeros, and the result of the formula “mean + z × standard deviation” is not a whole number and reflects the probability poorly.
What to count
- The average over the protection period: weekly demand × the number of weeks (the lead time plus the review period). The Poisson distribution: variance equals the mean, fits independent rare sales; the negative binomial: variance above the mean, fits when demand fluctuates more strongly (promotions, spikes, batches).
- The stock for the protection period: the smallest integer s at which the probability of demand not exceeding s reaches the target service level; the safety part: s minus the mean.
- The actual service level: if the stock is counted by Poisson while demand really has a larger spread, the real service level is lower than the declared one; it is counted by the negative binomial distribution and compared with the target.
What each side does
- Brand: counts stock by size for the warehouse and reorders, takes minimum batches and the size run into account; supports extreme sizes with a reorder and not with a series.
- Brand: checks the spread of demand by size (the variance to the mean) and chooses the distribution.
- Retail: determines the display minimum and holds extreme sizes centrally, replenishing the store on request.
- Retail: monitors the share of requests when a size is out and revises the service level.
Where the data comes from
| Data source | Fields |
|---|---|
| Sales history by size | units by week, weeks in stock, variance to mean |
| Lead time and review | lead time in weeks, the period between orders, minimum batches |
| Price of a missing size | margin per unit, the share of lost sales, holding cost |
How to build it step by step
- Step 1: Brand: collect weekly sales by size only for the weeks when the size was in stock.
- Step 2: Count the average weekly demand and the variance; if the variance is noticeably above the mean take the negative binomial.
- Step 3: Retail: determine the protection period: the lead time plus the period between orders.
- Step 4: Count the average over the protection period and find the smallest stock s at which the probability reaches the target level.
- Step 5: Brand: compare with the stock by the normal formula and check what service level the Poisson stock gives at a higher spread (see the example).
- Step 6: Choose the target level by size based on the price of being out: higher for fast sizes, lower for extreme ones.
- Step 7: Retail: for extreme sizes decide whether to hold them in the warehouse instead of stores; leave the store stock at the display minimum.
- Step 8: Recalculate once a season and when the lead time changes; record the actual service level.
Formulas
| Measure | Formulas | Note |
|---|---|---|
| The average over the protection period | μ = λ × T | λ: average weekly demand, T: weeks in the protection period (the lead time plus the review period). |
| Stock by Poisson | s = min{k : Σ_{i≤k} e^{−μ} μ^i / i! ≥ α} | α: the target service level; the stock is an integer. The safety part: s − μ. |
| Negative binomial | variance = φ × μ, r = μ / (φ − 1), q = (φ − 1) / φ | φ > 1: how many times the variance exceeds the mean; the probability P(k) is counted by recursion over k. |
Formulas in notation
- Stock by distribution
s = min{k : F(k; μ) ≥ α}, F(k; μ) = Σ_{i=0}^{k} e^{−μ}·μ^i / i!
F: the Poisson distribution function; α: the target service level; μ = λ·T. For a higher spread the negative binomial with r = μ/(φ − 1) and q = (φ − 1)/φ is taken.
Worked example
A protection period of 6 weeks, a target service level of 95%, six sizes with a conditional weekly demand; the stock by Poisson, by the normal formula and by a negative binomial with a variance twice the mean are compared
| Size | Average per period, units | Stock by Poisson | Stock by the normal formula | Stock at a spread ×2 | Actual service of the Poisson stock at a spread ×2, % |
|---|---|---|---|---|---|
| XXS | 0.60 | 2 | 2 | 3 | 93.7 |
| XS | 1.50 | 4 | 4 | 5 | 93.6 |
| S | 3.60 | 7 | 7 | 9 | 91.3 |
| M | 7.20 | 12 | 12 | 14 | 90.8 |
| L | 4.80 | 9 | 9 | 11 | 92.0 |
| XL | 1.80 | 4 | 5 | 6 | 91.0 |
The demand is conditional. The normal formula gives comparable numbers for an average of about three units and often diverges at small values; the Poisson stock at a real spread twice as high gives a service level below the declared 95%: so look at the variance and not only at the mean.
Figures
- Poisson
- Spread ×2
- Target
Data as a table
| Stock, units | Poisson | Spread ×2 | Target |
|---|---|---|---|
| 0 | 16.5 | 28.7 | 95 |
| 1 | 46.3 | 54.6 | 95 |
| 2 | 73.1 | 72.7 | 95 |
| 3 | 89.1 | 84.1 | 95 |
| 4 | 96.4 | 91 | 95 |
| 5 | 99 | 95 | 95 |
| 6 | 99.7 | 97.2 | 95 |
| 7 | 99.9 | 98.5 | 95 |
| 8 | 100 | 99.2 | 95 |
The probability of not running out of a size depending on stock (an average of 1.8 units). At the same average the negative binomial distribution requires more stock for the same service level.
- Poisson
- Normal formula
- Spread ×2
Data as a table
| Measure | Poisson | Normal formula | Spread ×2 |
|---|---|---|---|
| XXS | 2 | 2 | 3 |
| XS | 4 | 4 | 5 |
| S | 7 | 7 | 9 |
| M | 12 | 12 | 14 |
| L | 9 | 9 | 11 |
| XL | 4 | 5 | 6 |
Stock by size in three ways. For extreme sizes the stock by the formula and by the distribution differs by a unit, and that decides whether the size will be on the shelf.
Specifics by category
- Apparel
- Apparel: extreme sizes (XXS, XXL) sell rarely, but their absence hits the range hard; hold them centrally and replenish by reorder and not by series.
- Footwear
- Footwear: there are many sizes and the neighboring one does not substitute; rare sizes and pairs in small stores are counted by Poisson, and large sizes by a distribution with a higher spread because of peaks; stock is costlier because of the price of a pair.
- Accessories
- Accessories: there is almost no sizing (belts and gloves have it), rarity is created by a long tail of SKUs; the formula is the same but the SKU becomes the “size”.
- Kids
- Kids: sizes change quickly, demand by size depends on age and season and the remainder of an extreme size becomes obsolete; it is better to reduce the service level at the edges and replenish more often.
- Men
- Men: the size run is more predictable, demand in the middle of the run is steadier, extreme sizes are rare; the normal formula fits the middle, a distribution the edges.
- Women
- Women: the sizes are wider and demand depends more on the model and promotions, the spread is above Poisson; check the variance and use the negative binomial for promotional models.
Common mistakes
- Brand: zeros for weeks without stock are counted → demand is understated → take the weeks in stock.
- Retail: the normal formula is applied to an average of 1–2 units → fractional stocks and a wrong service level → count by a distribution.
- Brand: Poisson is taken without checking the variance → during promotions and spikes the stock protects worse than declared → compare the variance with the mean.
- Retail: one service level for all sizes → capital is frozen in extreme sizes → set the level by the price of being out.
- The protection period is counted only by the lead time → the period between orders is forgotten → add the review period.
Russian specifics
- Brand: the legislation does not set rules for calculating safety stock; it is a management method, the parameters are determined by the company’s data and goals.
- Retail: light-industry goods under marking are received with codes; when redistributing sizes between stores the operation is recorded in the system under the operator’s rules. Rule: Decree 1956 of 31.12.2019: Government Decree of 31.12.2019 No. 1956 (marking rules for light-industry goods).
- Brand: the absence of the needed size does not violate the law, but under contracts with retail chains fines for short supply are possible; check the contract.
- Retail: the lead times and the time of marking are reflected in the protection period; the numbers of the example are conditional.