Overview
Service parts demand forecasting is the estimation of future requirements for components used to maintain, repair, or replace equipment after its initial sale or installation. Forecasts inform purchasing, stocking, repair capacity, and distribution decisions across manufacturer, dealer, and field-service networks.
Unlike production materials, many service parts are requested irregularly and in small quantities. Demand may depend on the installed equipment base, component reliability, usage, maintenance policy, and contractual response times. Research on intermittent demand forecasting addresses the long intervals between requests that make conventional forecasting difficult. A modern synthesis of those methods is given in Boylan and Syntetos’s Intermittent Demand Forecasting. Forecasts are subsequently translated into inventory actions such as stock replenishment; a forecast is not itself a purchase order or a service-level guarantee.
Etymology and Origins
The term combines service parts, meaning components supplied for maintenance and repair, with forecasting, the practice of estimating future events. The broader word forecast predates modern inventory planning. The exact first documented use of the full phrase “service parts demand forecasting” is not established here; related literature also uses spare parts, replacement parts, and aftermarket parts.
The terminology evolved alongside the recognition that repair networks needed methods distinct from those used for regularly sold goods. The research term intermittent demand describes demand series containing many zero-demand periods and is central to the field, although not every service part has intermittent demand.
Historical Development
Organized spare-parts provisioning grew with industrial machinery and military equipment, where breakdowns could interrupt operations and repair locations needed access to replacement components. A major analytical milestone came in 1968, when Craig C. Sherbrooke published the METRIC model in Operations Research. METRIC (Multi-Echelon Technique for Recoverable Item Control) linked inventory placement and repair processes for repairable items to expected backorders rather than treating each stocking point in isolation.
In 1972, J. D. Croston proposed a forecasting method that separately estimates the size of nonzero demand and the interval between such demands. In 2005, Aris A. Syntetos and John E. Boylan published an analysis of intermittent-demand estimators and a bias-adjusted approach commonly called the Syntetos–Boylan approximation (SBA). Also in 2005, Syntetos, Boylan, and John D. Croston published work on categorizing demand patterns, addressing the selection of forecasting methods for different demand characteristics. In 2011, Ruud Teunter, Syntetos, and Mohamed Zied Babai addressed obsolescence by updating demand probability in periods without demand; that estimator is often called the Teunter–Syntetos–Babai (TSB) method.
Operational practice has moved from manual records and periodic calculations toward computerized inventory systems and specialized forecasting software. Barcode-based identification supports more consistent transaction capture; warehouse barcode tools are one example of the technology used to record movements. Modern models can also incorporate equipment telemetry and statistical or machine-learning methods, but their usefulness depends on data quality and evidence from out-of-sample testing.
Core Concepts
Demand drivers and forecasting horizon
A forecast may be built for a part, location, and time interval, such as units per month at a regional depot. Key drivers include the number and age of assets in service, their operating intensity, failure or replacement rates, preventive maintenance schedules, and whether failed parts are repaired or discarded. A forecast based on observed issues alone may understate actual requirements if customers experienced stockouts or substituted another part.
The planning horizon should cover the decision being made: supplier lead time for replenishment, for example, or a longer period for capacity and end-of-life purchases. Multiechelon repairable-item models illustrate why replenishment and repair flows must be considered across a network, not just at a single shelf.
Intermittent demand
For intermittent items, a useful decomposition is:
Average demand per period ≈ average nonzero demand size ÷ average interval between nonzero demands.
For example, if requests average two units and occur every four months on average, the implied rate is 0.5 units per month. This is a long-run rate, not a prediction that half a unit will be requested in any particular month. Croston’s method smooths the two components separately; variants adjust its bias or change how demand occurrence is updated.
Demand patterns also differ in both the spacing between requests and the variability of nonzero request sizes. Demand-pattern classification can help identify candidate methods, but planners should test those candidates against alternatives rather than assume one method is best for every part.
Forecasting Methods and Data
Common approaches include:
- Simple baselines: historical averages, moving averages, or exponential smoothing. These are useful benchmarks and may suit regularly consumed items.
- Intermittent-demand methods: Croston and related estimators separate demand size from arrival frequency. The Syntetos–Boylan approximation adjusts for bias associated with Croston’s estimator. Later work, summarized in Intermittent Demand Forecasting, compares these estimators with classification rules and inventory implications.
- Probability-based methods: models estimate the distribution of demand over lead time, enabling planners to calculate stock levels for a chosen service objective rather than rely only on an average. A prediction interval expresses uncertainty around a forecast, but its usefulness depends on the assumed model and its calibration.
- Installed-base and reliability models: expected requirements are estimated from assets in use, exposure, failure behavior, maintenance actions, and the quantity of each part required per event.
- Judgment and event-based adjustments: planners incorporate recalls, scheduled overhauls, product launches, and other events that historical transactions may not represent.
Useful inputs include dated demand requests, backorders, cancellations, substitutions, inventory positions, supplier lead times, asset populations, and repair turnaround times. Inventory management systems and service records can supply operational data, but transaction histories should be reconciled before modeling. A shipment date, for example, is not necessarily the date a customer first requested a part.
Illustrative installed-base calculation: suppose 1,000 machines each have one relevant component, and the assumed annual replacement rate is 4%. With no planned overhauls or other adjustments, the implied annual requirement is 40 components. This is an expected total, not a guaranteed demand figure; planners should document the asset count and replacement-rate assumptions and account for uncertainty when setting stock levels, consistent with forecast uncertainty guidance.
From Forecast to Inventory Decisions
Forecasts support decisions about where to hold stock, when to replenish it, and how much protection to carry against uncertain demand and lead time. A basic reorder rule compares the inventory position—on-hand stock plus outstanding replenishments, less backorders—with a reorder point. The latter typically reflects expected demand during lead time plus a buffer or a quantile of the lead-time demand distribution. Standard inventory texts such as Axsäter’s Inventory Control distinguish these policy calculations from the forecast itself.
Illustrative example: a depot expects 0.5 units per month and has a four-month replenishment lead time. Expected lead-time demand is two units. A planner might choose a reorder point above two units to protect against variability, but the appropriate threshold depends on the full demand distribution, lead-time uncertainty, and the cost of a stockout. It cannot be determined reliably from the average alone.
Service objectives should be defined explicitly. A cycle service level measures the probability of avoiding a stockout during a replenishment cycle; a fill rate measures the proportion of requested units supplied immediately. They are different measures, and the stock required to meet one is not generally the stock required to meet the other. For repairable components, multiechelon inventory analysis also considers repair turnaround and the movement of unserviceable items. Once a policy is approved, purchase order management helps administer replenishment; it does not remove the need to validate the forecast or inventory parameters.
Applications, Benefits, and Limitations
Service parts forecasting is used wherever organizations support equipment after deployment. An industrial manufacturer may plan regional stocks for machine repairs; an automotive network may position replacement components among dealers; and an aerospace operator may plan repairable units across bases and maintenance facilities. In each case, demand timing and the consequence of a missing part can differ substantially.
A disciplined forecast can improve purchasing and stocking decisions while making assumptions visible to service, finance, and operations teams. However, low-volume histories contain limited information, and a single major failure or order can distort an estimate. Parts may also be superseded or become obsolete. The Teunter–Syntetos–Babai method explicitly updates demand probability during zero-demand periods, making it relevant when demand may disappear, although it cannot independently identify every engineering change.
Data quality is another constraint. Stockouts can censor observed demand, transfers can be mistaken for consumption, and inconsistent units or duplicate part numbers can fragment history. Cycle counting helps improve the inventory records used by planners, but accurate on-hand balances alone do not establish the quantity of unmet demand.
Best Practices and Performance Measurement
- Segment the portfolio. Group parts by demand pattern, criticality, cost, lead time, repairability, and lifecycle stage. A low-volume safety-critical component may warrant a different policy from a low-cost consumable. Use demand-pattern classification as an input to method selection, not as a substitute for operational judgment.
- Preserve the demand signal. Record request dates and quantities separately from shipments, transfers, and cancellations; identify stockout periods and substitutions.
- Compare methods against a baseline. Test on historical holdout periods using the same forecast horizon as the actual replenishment decision. Review results for both individual parts and meaningful aggregates. Where data permit, use repeated historical test windows and check whether prediction intervals contain outcomes at approximately their stated frequency.
- Measure forecast and inventory outcomes separately. Assess bias and forecast errors alongside fill rate, backorders, downtime exposure, inventory investment, and obsolescence. Percentage errors can be misleading or undefined when actual demand is zero; forecast accuracy guidance explains the limitations of common measures.
- Review exceptions and assumptions. Investigate large forecast changes, supplier disruptions, known maintenance campaigns, supersessions, and end-of-life events. Maintain an audit trail for overrides.
- Connect planning to execution. Reconcile part and location data across maintenance, purchasing, and warehouse systems, and periodically check that replenishment settings reflect the approved inventory policy.