Demand Forecasting Wiki

Safety Stock and Reorder Point for Intermittent Demand

By MLAIA Data Science Ltd. · Published 10 November 2026 · Updated 10 November 2026

For intermittent demand, set the reorder point from the actual distribution of demand over the lead time, not from the normal-distribution formula. In the worked example below, a lumpy part with monthly demand 0, 0, 20, 0, 1, 0, 0, 50, 0, 0, 4, 0 and a two-month lead time, the classic formula z·σ·√L gives a reorder point of 48 units for a 95% cycle service level, but the probability that lead-time demand stays at or below 48 is only 84%. The empirical distribution puts the 95% reorder point at 50 units, which means safety stock of 37.5 units above the expected lead-time demand of 12.5.

What the reorder point really is

In a continuous-review policy you order when the inventory position falls to the reorder point (ROP), and the order arrives one lead time later. The ROP protects against demand during that lead time. A cycle service level (CSL) of 95% means: the probability that lead-time demand does not exceed the ROP is 95%. So the ROP is, by definition, the 95th percentile of the lead-time-demand distribution, and safety stock is the gap between that percentile and the expected lead-time demand:

ROP = QCSL(lead-time demand)   SS = ROP − E[lead-time demand]

The familiar textbook formula, SS = z·σd·√L, is what this becomes when lead-time demand is normally distributed. The formula is not wrong; the assumption behind it is wrong for intermittent items.

Why the normal formula misfires on intermittent demand

Lead-time demand for a slow mover is most often zero, sometimes small, and occasionally large. Its distribution is discrete, has a spike at zero and a long right tail. A normal curve fitted to the same mean and standard deviation is symmetric, continuous, and puts probability on negative demand. Its 95th percentile can land almost anywhere relative to the real one, and the discrepancy grows with the variability of the order sizes (CV²): for intermittent items with stable sizes the two often roughly agree, while for lumpy items they can differ badly in either direction.

Worked example: a lumpy part, two-month lead time, 95% target

The history. Twelve months of demand: 0, 0, 20, 0, 1, 0, 0, 50, 0, 0, 4, 0. Total 75 units; mean 6.25 per month; demand in 4 of 12 months; sizes 20, 1, 50, 4. ADI = 3.0 and CV² = 1.07, so the item is lumpy.

Step 1 — the classic formula. The sample standard deviation of the twelve monthly values is 14.92 (the sum of squared deviations from 6.25 is 2,448.25; divided by 11 that is 222.57, whose square root is 14.92). With L = 2 and z = 1.645 for 95%:

SS = 1.645 · 14.92 · √2 = 34.7   ROP = 6.25 · 2 + 34.7 = 47.2 → 48 units

This is what the safety stock calculator returns when the series is pasted in. Keep the number: we are about to test it.

Step 2 — a model of one period. Treat each month as a coin toss: demand occurs with probability 4 ÷ 12 = 1/3, and when it occurs the size is drawn from the observed sizes, each equally likely: 1, 4, 20 or 50. This is a compound Bernoulli model, the simplest member of the family that bootstrapping (Willemain, Smart & Schwarz, 2004) builds by resampling.

Step 3 — compound over the lead time. Over two months the number of demand events is 0, 1 or 2 with probabilities (2/3)² = 0.444, 2·(1/3)·(2/3) = 0.444 and (1/3)² = 0.111. With one event, each of the four sizes has probability 0.444 ÷ 4 = 0.111. With two events, each of the 16 ordered pairs of sizes has probability 0.111 ÷ 16 = 0.0069; for example lead-time demand of 24 arises from (4, 20) and (20, 4), probability 0.0139. Collecting every outcome:

Lead-time demand (LTD) over two months under the compound model, with its probability and cumulative probability.
LTDHow it arisesProbabilityP(LTD ≤ this)
0no demand event0.4440.444
1one event of 10.1110.556
21 + 10.0070.563
4one event of 40.1110.674
51 + 4, 4 + 10.0140.688
84 + 40.0070.694
20one event of 200.1110.806
211 + 20, 20 + 10.0140.819
244 + 20, 20 + 40.0140.833
4020 + 200.0070.840
50one event of 500.1110.951
511 + 50, 50 + 10.0140.965
544 + 50, 50 + 40.0140.979
7020 + 50, 50 + 200.0140.993
10050 + 500.0071.000

Step 4 — read off the quantile. The first LTD level at which the cumulative probability reaches 0.95 is 50, so ROP = 50. Expected lead-time demand is 6.25 · 2 = 12.5, so safety stock is 50 − 12.5 = 37.5 units. For a 99% target the ROP would be 70.

Step 5 — test the formula's answer. At the formula's ROP of 48, the table shows P(LTD ≤ 48) = 0.840. A planner who set 48 believing it bought 95% protection would actually stock out in about one cycle in six, because 48 sits just below the 50-unit order that carries 11% of the probability on its own. Two more units closes the gap from 84% to 95%; the formula could not see that, because a bell curve has no lump at 50.

A less lumpy comparison. Run the same two steps on the intermittent series 0, 5, 0, 0, 8, 0, 12, 0, 0, 6 (CV² = 0.12, lead time 2): the normal formula gives 6.2 + 1.645·4.38·√2 = 16.4 → 17, and the compound model also reaches 95% at 17. With stable sizes the two agree; it is size variability that pulls them apart.

Other ways to get the distribution

Cycle service level is not fill rate

Everything above targets the cycle service level: the chance of no stockout in a cycle, regardless of its size. Fill rate, the share of demanded units shipped from stock, is usually what the business means by "service". The lead-time-demand table shows why they diverge on this part. Look at a single lead-time window with 50 units on hand and no replenishment arriving inside it: the 5% of windows that run short are the ones with demand of 51, 54, 70 or 100, with shortfalls of 1, 4, 20 and 50 units, so the expected shortfall is 0.014·1 + 0.014·4 + 0.014·20 + 0.007·50 ≈ 0.69 units, about 5.5% of the 12.5 units expected — single-window coverage of roughly 94.4%. With 24 units on hand the window is covered 83% of the time, but the expected shortfall rises to about 5.0 units and the coverage falls to roughly 60%, because the big orders are the ones that go unserved. These are coverage figures for one window, not the fill rate of the continuous-review policy: that fill rate is 1 − (expected shortfall per cycle ÷ order quantity Q), so it also depends on how much is ordered each time, which this example does not fix. The lesson carries over regardless: on a lumpy item, two reorder points with similar cycle service levels can leave very different numbers of units unserved. Size the buffer against the measure you are accountable for.

A practical recipe

  1. Classify the item. Smooth and mildly erratic items can use the z-score formula; intermittent and lumpy items should not.
  2. Forecast the demand rate with SBA or TSB; use it to sanity-check the mean of whatever distribution you build.
  3. Build the lead-time-demand distribution from the history: the compound model above for a quick answer, bootstrapping or a fitted discrete distribution for a production system.
  4. Set ROP at the target quantile and report safety stock as ROP minus expected lead-time demand, so the buffer is visible.
  5. Recompute on every forecast refresh and whenever the supplier lead time changes, and watch the ADI: a lengthening interval is the first sign the item is heading for excess stock.

Frequently asked questions

Can I use the normal safety stock formula for spare parts?

Only for the minority of parts whose demand is smooth. For intermittent or lumpy parts the formula z·σ·√L returns the 95th percentile of a bell curve that does not describe the demand: in the worked example it gives a reorder point of 48 units, at which the actual probability of covering lead-time demand is 84%, not 95%. Build the lead-time-demand distribution from the history instead, or use a discrete distribution such as the Poisson or negative binomial.

How do I calculate safety stock for intermittent demand in Excel?

Compute the probability of demand per period (periods with demand ÷ total periods) and list the observed non-zero sizes. For a lead time of L periods, enumerate how many demand events can occur (0 to L, binomial probabilities) and, for each count, the sums of sizes with their probabilities; sort the resulting lead-time-demand values and accumulate the probabilities. The reorder point is the first value whose cumulative probability reaches your service level, and safety stock is that value minus the mean demand per period times L. For a long lead time, resampling (bootstrapping) is easier than enumeration.

Why is the fill rate different from the cycle service level for slow movers?

Cycle service level counts stockout cycles; fill rate counts unserved units. On an item whose orders range from 1 to 50 units, the cycles that stock out are the ones with the largest orders, so each of them loses many units. Two reorder points with the same cycle service level can therefore have very different fill rates, and lowering the reorder point hurts fill rate faster than it hurts cycle service level. Decide which measure you are managing to before sizing the buffer.

References

See also

Try it on your own items