Guide
The safety stock formula
Safety stock is not a comfort blanket. It is a priced decision about how often you are willing to run out — and the formula tells you exactly what that decision costs.
Why any buffer at all
Order exactly lead time demand and you will stock out roughly half the time. Not occasionally — half the time. Demand exceeds its own average in about half of all periods, by definition. Safety stock is what buys you down from a 50% stockout rate to something a business can live with.
How far down is a decision, and it has a price. That is the whole content of the formula.
The formula
The square root of a sum of two variances. That construction encodes an assumption worth stating plainly: demand risk and supply risk are treated as independent. Your supplier being late is not correlated with your customers buying more. That is usually true, and where it is not — a market-wide shortage hits supply and demand together — the formula will understate the buffer you need.
Service level and the Z-score
Z is the number of standard deviations of cover you are buying, drawn from the normal distribution:
| Service level | Z | Meaning | Relative buffer |
|---|---|---|---|
| 50% | 0.00 | No buffer. Stock out half the time | 0 |
| 80% | 0.84 | Stock out 1 cycle in 5 | 0.51× |
| 90% | 1.28 | Stock out 1 cycle in 10 | 0.78× |
| 95% | 1.65 | Stock out 1 cycle in 20 | 1.00× |
| 98% | 2.05 | Stock out 1 cycle in 50 | 1.25× |
| 99% | 2.33 | Stock out 1 cycle in 100 | 1.41× |
| 99.9% | 3.09 | Stock out 1 cycle in 1,000 | 1.88× |
The right-hand column is the one to read before choosing. Moving from 95% to 99.9% removes about one stockout in twenty cycles and costs 88% more buffer — permanently, on every unit, forever. Note also that the relationship is not linear: the first 45 percentage points of service cost you 1.0×, and the last 4.9 cost you another 0.88×.
Reading the variance split
The two terms under the root are separately meaningful, and comparing them is the most actionable output of the whole exercise.
If supply risk dominates, your buffer exists to insure against your supplier. The fix is commercial, not analytical: tighten the lead time, agree a delivery window, or add a second source. Cutting σL in half typically cuts the total buffer by a third or more.
If demand risk dominates, the buffer is the price of genuinely unpredictable customers. Better forecasting, promotion planning or shorter lead times will help; supplier conversations will not.
The calculator prints this split as a percentage, because most businesses guess it wrong — and usually guess demand when the answer is supply.
Why "two weeks of cover" fails twice
The flat rule is popular because it needs no data. It is also wrong in both directions simultaneously:
- On a steady, reliably supplied product, two weeks is far more than the risk warrants. That is cash on a shelf earning nothing, incurring 20–30% a year in carrying cost.
- On a volatile product with an unreliable supplier, two weeks is not enough. You stock out anyway, having paid for a buffer that did not cover the actual risk.
So you over-invest where you are safe and under-invest where you are exposed. A flat rule is not a conservative choice; it is an uninformed one, and it costs money at both ends.
Choosing service levels per class
Service level should not be uniform, because the cost of a stockout is not uniform. A reasonable policy:
| Class | Typical service level | Reasoning |
|---|---|---|
| A — high value or critical | 98–99% | A stockout loses the customer, not just the sale |
| B — ordinary lines | 95% | Standard balance of cost and risk |
| C — slow tail | 85–90% | Carrying cost outweighs the occasional wait |
| Perishable | 90–95% | Higher buffers become write-offs, not insurance |
| Single-source, long lead time | 98%+ | Recovery from a stockout is measured in weeks |
ABC analysis is how you assign the classes. Doing this well is typically worth more than any refinement of the formula itself.
In practice
- Pull daily sales for a representative period and take
STDEV.Pfor σd. - Pull your recorded delivery times and take
STDEV.Pfor σL. If you have never recorded them, start now — this single dataset is worth more than most forecasting effort. - Assign a service level by ABC class.
- Compute, add lead time demand to get the reorder point, and load it as each product's threshold.
- Review quarterly, and immediately on a supplier change.
If you record nothing else, record delivery dates. Almost every business can estimate demand variability from sales history it already has. Almost none can estimate lead-time variability, because nobody wrote down when deliveries actually arrived — and that is usually the larger term.