Two supplier behaviours do more damage to a distributor's service than demand ever does: shipping less than was ordered, and shipping late. Both are visible in the receipts the business already has. Both are usually summarised by an average, and the average is the wrong number.
Fill rate is a distribution
A supplier with an 85% average fill might ship 100% most of the time and 40% occasionally, or 85% every time. The first pattern is far more dangerous: the occasional 40% is the week the customer is short. Measure, from goods receipts against purchase orders, the accepted ratio on every order, and look at the whole distribution: the average, the spread, and the probability of a partial delivery below the level that would cause a stockout. Protection should scale with the spread and the partial probability, not with the mean.
Lead time has a tail
Import lead times are typically right-skewed: most shipments arrive near the quoted date, a few arrive much later. The standard deviation understates the tail. What matters for protection is the lead time at the required quantile (the 90th percentile for a 90% target, roughly) and the probability that an order is late by more than one review period.
A subtlety: at any moment, some orders are still open. Ignoring them biases the estimate toward the quick arrivals; treating them as if they had arrived fabricates data. The right tool is the Kaplan-Meier estimator, which uses open orders as right-censored observations: it knows the order has taken at least this long, and no more. It also gives the conditional estimate that a planner actually needs: given that this order is already 20 days old and still open, when will it arrive?
Why variability costs inventory
Hopp and Spearman's variability laws are the compact statement: inventory needed to hit a service level rises with the variability of both demand and replenishment, and the effects compound. A supplier whose lead time doubles in variance while its mean stays the same forces more stock, not because more is being sold but because when it arrives is less certain. The same is true of fill variance.
What a planner can do about it
Order earlier for the tail rather than more for the mean; split orders between suppliers whose failures are independent; move part of the volume to a more reliable source when the shortage avoided is worth the price difference; expedite only when the economics say so. And treat the supplier's own record as the evidence: a source whose observed tail has widened gets more protection; one whose record is clean does not, however alarming its quote.
Measuring it honestly
Use the operator's receipts, not the supplier's promises. Keep censored orders as censored. Report the number of records the estimate rests on: a distribution built from four deliveries is a guess with error bars, and the plan should say so.