Service part inventory models can be distinguished by their intended application on single part/machine/facility, or on servicing an installed base of equipment. Besides typical uncertainties in part failure, the install base problem is compounded with the heterogeneity of the part age in equipment life cycle and scarcity of data toward the end of life phase. This paper presents an empirical study of the final order problem after the sale of a product is discontinued but there is an installed base to be serviced. Two fundamental issues are addressed: (1) to build forecast models on sales data directly or on failure probability indirectly, and (2) to use few but recent data or many but dated data. This paper first shows that regression on derived failure probabilities yields more accurate forecasts than regression on part sales data. The effect of data age and quantity on forecast is next investigated. The recency of historical data is shown to be more informative than the quantity of input data in demand forecast. Finally, an installed-base forecast model is constructed which show an improvement of 16.0% in absolute forecast errors than an existing method used in practice by a case study firm.