How do you forecast demand for slow-moving spare parts?

Not with a moving average. Classic methods assume a regular demand signal, and spare parts demand is intermittent: long stretches of zero, then a sudden requirement. Averaging those zeros gives a forecast of 0.3 units a month, which is both mathematically defensible and operationally useless.

Three approaches work better, and they combine.

Methods built for intermittent demand, such as Croston's method and its variants, split the problem in two: they forecast the size of a demand when it occurs and the interval between occurrences separately, instead of smoothing both into one meaningless number.

Probabilistic forecasting produces a distribution of possible demand rather than a single point. That is what lets you set stock against an explicit service-level target: covering 99% of scenarios on a critical part, 85% on a non-critical one. Saint-Gobain Sekurit AGR improved SKU-level forecast accuracy by 15 points on a portfolio of more than 10,000 references by moving to this kind of model.

Maintenance data as a demand signal is the third lever. Preventive and condition-based maintenance plans tell you which parts will be needed and roughly when. In most organizations that information exists but stays in the computerized maintenance management system (CMMS). Connecting it to inventory planning turns part of an unpredictable demand into a scheduled one.