Safety stock in fashion: why classic formulas break and what replaces them

Classic safety stock formulas break in fashion because they assume demand that is steady, independent from week to week and roughly normally distributed. Fashion demand is seasonal, driven by campaigns and weather, often sparse at size and store level, and tied to a short selling window. A buffer calculated with the textbook formula will therefore be too high in some cases and too low in others.

This explainer sets out what the standard formula assumes, where it fails in fashion, and which approaches planners use instead, including probabilistic forecasts and newsvendor logic.

What is safety stock and how is it normally calculated?

Safety stock is extra inventory held as a buffer against stockouts. It absorbs variation in demand, delays in supplier deliveries and manufacturing yield shortfalls. Holding too much raises holding costs and write-down risk, while holding too little causes lost sales, so the aim is a balance.

The common formula combines a z-value for the target service level (about 1.65 for 95%, for example), the mean and standard deviation of lead time, and the mean and standard deviation of demand per period. The reorder point is the average demand during lead time plus safety stock. The approach is widely taught and works well for stable, continuously replenished items.

Why do classic formulas break in fashion?

A reference overview of the formula lists its assumptions and limitations. They map closely onto fashion problems.

The source itself warns that no universal safety stock formula exists, and that applying one without checking its assumptions can cause serious problems. The last row of the table is an addition of ours rather than a quotation: it reflects how seasonal buying works, and it is the main reason many fashion planners think in terms of a season quantity rather than a reorder point.

Read also

AI for merchandisers and planners: a practical guide

What replaces the classic formula for seasonal items?

For items that are bought once and lose value after the season, the newsvendor model is the standard reference. It describes a single-period problem with uncertain demand, where unsold units lose all value at the end of the period. The optimal stock level is the demand quantile equal to the underage cost divided by the sum of underage and overage cost. The underage cost is the margin lost on each unit of demand that goes unmet, and the overage cost is the purchase cost of each unit left unsold.

The logic is simple and useful: high-margin items justify a higher quantile of demand, because lost sales cost more than leftover stock, and thin-margin items justify a lower one. The source discusses perishable products such as newspapers and fresh food, and applying it to fashion is an extension of the single-period logic rather than something the source states. In practice fashion has salvage value through markdowns and outlets, so the overage cost is the purchase cost minus what the leftover units eventually fetch.

• Use the newsvendor logic for fashion and trend styles with a single buy.
• Use service-level based safety stock for core, continuously replenished items.
• Differentiate by margin and markdown exposure, not only by sales volume.
• Review buffers for sizes at the ends of the size run, which sell slowly and are the most likely to be left over.

How does AI change safety stock planning?

AI mainly improves the input to the buffer, which is the demand distribution. Rather than a single forecast and a normal approximation, a probabilistic model outputs a range of outcomes by item and location, and the buffer is set from the chosen quantile. This reduces reliance on the normality assumption.

A position paper on learning-based forecasting in fashion, by researchers at the University of Verona and Humatics, shows the scale of the difficulty. It reports weighted MAPE for new product forecasting on the Visuelle dataset improving from about 59% to about 52% as methods moved to multimodal models, and it estimates that a 5% accuracy gain could save over $156,000 in one example. Those figures show that even advanced models leave considerable uncertainty, which is exactly why a buffer is still needed. The paper also argues that classical methods such as ARIMA cannot be fitted for products with no history.

Forecast accuracy is judged on unseen data. The Hyndman and Athanasopoulos textbook recommends holding out a test set and warns that overfitting is as harmful as missing a real pattern. For buffers, the more relevant check is whether stated service levels were achieved: if a 95% target produces stockouts in 15% of cases, the demand distribution is too optimistic.

What should a planner check before trusting a buffer?

1. State the service level target by item group and justify it by margin and markdown exposure.
2. Check lead time data: measured from receipts, not from supplier promises.
3. Check whether demand is censored: sales during stockouts understate real demand.
4. Compare the proposed buffer with a simple rule, such as weeks of cover, on past seasons.
5. Track realised service level and end-of-season leftovers together, because lowering one often raises the other.

Read also

How does AI inventory optimisation reduce deadstock and stockouts in fashion?

What are the limits?

No method removes uncertainty. Better distributions can narrow the buffer for stable lines and justify more for volatile ones, but the cost of a stockout, the markdown cost of leftovers and the service level target are commercial judgements. Planners also need a way to override the model for events it cannot know, such as a sudden supplier problem or a change in marketing plans.