Intermittent demand occurs when many periods have no demand at all, with occasional transactions appearing irregularly. It is common in spare parts, maintenance inventory, industrial components, and slow-moving products. Because both the timing and size of demand are uncertain, methods designed for steady sales often produce misleading results.
Good intermittent demand forecasting does not promise to predict the exact next order. It estimates the probability and range of demand, then connects that uncertainty to stocking, service, and network decisions.
What Makes Demand Intermittent?
Low volume alone is not enough. A product can sell one unit every week and remain regular. Intermittent demand includes repeated zero periods and irregular intervals between transactions. It becomes lumpy when the nonzero orders also vary substantially in size.
Spare parts are a classic example because demand is often triggered by equipment failure. Planned maintenance may create a more predictable component, while unplanned failures remain uncertain. Asset age, installed-base size, geography, operating conditions, and reliability can all influence future requirements.
Why Traditional Forecasts Struggle
A simple moving average or exponential smoothing model can spread sparse demand evenly across time. An average of 0.4 units per month does not mean demand will arrive smoothly; it may mean a small probability of needing several units at once.
Long runs of zero demand are informative, but they do not prove that future risk is zero. Likewise, one large order may be an anomaly, a project, or a genuine change. Automatically treating it as a trend can create excess inventory.
Point forecasts hide these distinctions. In intermittent demand forecasting, probabilistic forecasts estimate the likelihood of different demand quantities over the relevant lead time. That gives planners the information needed to compare a stocking cost with the service consequence of being short.
Methods for Intermittent Demand Forecasting
Croston's method was an important advance because it estimated demand size separately from the interval between demands. Later approaches, including the Syntetos-Boylan adjustment and Teunter-Syntetos-Babai method, addressed known biases and the risk of obsolescence.
No single method is best for every SKU. In intermittent demand forecasting, model choice should reflect intermittency, lumpiness, lifecycle, and the decision being supported. Modern AI can evaluate multiple models, learn across related items, incorporate additional signals, and update classifications as behavior changes.
Historical demand is only one input. Useful signals may include:
Installed-base size, location, and age
Reliability and failure-rate data
Planned maintenance schedules and bills of material
Product supersessions and substitution options
Known projects, campaigns, and one-time orders
Supplier lead times and discontinuation notices
Known future demand should be separated from the statistical baseline. Planned maintenance and confirmed orders should not be blended blindly with unpredictable failure demand.
Connect Forecasting to Inventory Decisions
Intermittent demand forecasting only matters when it informs better inventory decisions.
Inventory policy should consider the probability of demand during lead time, order size, replenishment variability, item value, criticality, and target service.
Not every spare part needs the same service level. Safety-critical or operationally critical components may justify very high availability. Important but deferrable items can accept a lower target, while noncritical or substitutable items may be centralized or supplied to order.
The service-level curve reveals the inventory cost of incremental availability. Moving from 90% to 95% service may be economical for one part and extremely expensive for another. Optimization should therefore operate across the portfolio rather than maximizing every item independently.
Location and Network Matter
Demand may look intermittent at each local warehouse while appearing more stable across the network. Centralizing rare-demand inventory can pool risk, but response-time commitments may require selected local stock.
Multi-echelon inventory optimization complements probabilistic forecasting by deciding how much inventory to hold and where. Planning each warehouse independently usually duplicates protection. Network-level optimization balances central, regional, local, and technician inventory against service and cost.
Forecast at the level where the decision is made, while using hierarchical information where appropriate. Local stocking decisions need local demand and lead-time context, but related products, regions, or assets can provide useful information when individual history is sparse.
Lifecycle, Supersession, and Repairables
Intermittent behavior changes through introduction, growth, maturity, decline, and end of life. New spare parts have little history and may be forecast using analogous products, installed-base plans, and engineering information. Declining items need structural-change detection so outdated history does not sustain unnecessary stock.
Supersession must be modeled correctly. Forecasting predecessor and successor parts independently can split demand history and strand inventory. Substitution options should be reflected in both the forecast and stocking decision.
Repairable parts need different logic because repair turnaround acts as a replenishment lead time. The buffer must consider failures, repair yield, turnaround variability, and the number of items circulating through the repair loop.
Supplier discontinuation creates last-time-buy decisions. Scenario analysis should compare possible future demand, remaining installed-base obligations, substitution, repair, and the cost of leftover stock.
Measure Outcomes That Matter
Traditional percentage-error metrics behave poorly when actual demand is zero. Evaluation should combine intermittent-demand-appropriate measures with probabilistic calibration, bias, service attainment, emergency shipments, inventory investment, and excess by demand class.
A forecast can be statistically imperfect and still create a good stocking decision. Conversely, a strong aggregate accuracy score can hide shortages on critical parts and excess on slow movers. Business outcomes are the final test.
Exception-based planning keeps the workload manageable. Models should continuously recalculate routine items, while planners focus on material changes such as a new contract, unusual order, supplier disruption, lifecycle event, or high-impact risk.
A Practical Process
Clean transaction history and separate returns, corrections, and nonrecurring events.
Classify demand as regular, variable, intermittent, lumpy, new, declining, or dormant.
Select and compare suitable statistical or AI models.
Add installed-base, maintenance, reliability, lifecycle, and commercial signals.
Produce probability distributions over the relevant lead time.
Define service targets based on criticality and business value.
Optimize inventory at each SKU-location and across the network.
Automate replenishment and prioritize exceptions.
Measure service, bias, emergency actions, inventory, and excess.
Reforecast and reoptimize as conditions change.
Improve Intermittent Demand Forecasting with Smarter Inventory Decisions
ToolsGroup combines probabilistic forecasting with inventory and multi-echelon optimization, helping organizations translate sparse demand into better service, lower excess, and more focused planner decisions.
