Forecasting and prediction
Forecasting spare parts demand for NZ service firms
Set reorder decisions for rarely used parts using stockouts, supplier lead times and service commitments. Test the forecast against real replenishment choices.

A maintenance business can sell no pump seals for weeks, then need several in one morning. Buying a demand forecast by average monthly sales misses two practical questions. How long does replacement stock take to arrive, and what happens to a customer if the workshop has none? Build the buying brief around the reorder decision for each part.
Does zero mean nobody needed the part?
Keep issues from stock, customer requests, substitutions and unavailable items as separate events. If the shelf was empty, a zero in the sales file describes what the workshop supplied. It gives little evidence of the demand it could have met. Ask staff how they record a request that goes straight to another dealer.
Look at a depot serving a mix of rural customers and local contractors. A planned pump overhaul may require a known list of parts next month. An urgent seal failure arrives without warning. Reserve confirmed job requirements separately and estimate the remaining uncertain demand. Otherwise a model can count the same impending overhaul in the reservation list and the forecast.
What decision does the lead time create?
Draw the timeline from order placement to stock available for use. Include the supplier dispatch date, freight, receiving checks and any local transfer. Start with the number of units likely to be needed across that period. Compare it with on-hand stock plus confirmed inbound deliveries, less existing commitments. Make each quantity visible in the suggested order.
Suppose a technician can see four seals on hand, two reserved and three on a purchase order due after the next service visit. Only two unreserved seals are available for that visit. An interface that simply adds the purchase order to the shelf tally would overstate cover. Let the buyer adjust an uncertain delivery date and see which jobs are affected.
Where does a forecasting trial earn its keep?
Replay orders in time sequence, using supplier lead times and job knowledge available on each historical ordering date. Compare the proposed reorder rule with the current minimum-and-maximum stock rule. Count jobs delayed, urgent freight and leftover units for each approach. Show the parts that caused the worst outcomes so the buyer can check causes against workshop records.
Hyndman and Athanasopoulos describe benchmark forecasts and evaluation against later observations in Forecasting: Principles and Practice. For intermittent stock, also test the actual purchasing rule. A small numerical improvement has little value if minimum pack sizes leave the order unchanged. Ask the supplier for an exception list covering newly introduced parts, substitutes and parts that a supplier has discontinued. Those cases need a purchasing decision alongside the forecast.
When evaluating seldom-used parts, show unit errors and days a serviceable part was unavailable. Percentage error is undefined when actual demand is zero, and a percentage calculated after dropping all those periods hides the many weeks the workshop carried the stock. A short list of expensive stoppages often gives the buyer a more concrete basis for a decision.