Forecasting purchases to limit stockouts

Industrial parts distributor

The distributor, which has 90 employees, generated $15 million in sales with $2 million in inventory. It had a management system, but buyers adjusted their forecasts in separate spreadsheets; stockouts and surpluses coexisted.

Approach

1. 01

   Frame

   The director aimed for fewer stockouts without increasing average inventory. The pilot covered 300 items with reliable history, and each order remained approved by a buyer.

2. 02

   Investigate

   I matched two years of sales, stock levels, supplier lead times and lost orders, checked substitutions, and distinguished a sale missed for lack of stock from genuinely weak demand.

3. 03

   Explain

   Two diagnoses were compared: inadequate forecasts or unstable supplier lead times. The analysis showed that some stockouts called for a change of supplier rather than a better model.

4. 04

   Build

   The options were the status quo, adjusted thresholds, a simple statistical forecast and an AI-enhanced model, each compared at the same inventory cost and the same service level.

5. 05

   Test

   After validation on periods separate from those used for training, the pilot ran for twelve weeks. The threshold was 20% fewer stockouts than the simple model, at constant stock. The model reached 22% on the selected product families; for stable items, AI brought no advantage, and they remain managed by rules.

6. 06

   Decide

   The director set aside $45,000, including $10,000 for strategy, and deployed the model only on the families where the advantage was confirmed. Unreliable suppliers are handled separately.

7. 07

   Steer

   Tracking measures availability, average inventory, emergency purchases and orders actually recovered. A product family reverts to simple rules if the advantage disappears for two cycles.