EN
Insight

Spare parts planning: why traditional forecasts fail with intermittent demand

Spare parts sit in the warehouse for years, and yet when equipment goes down, the one part you need is missing. The forecasting method is often to blame: averages and smoothing do not fit demand that occurs only now and then. With suitable methods, properly measured forecast accuracy, and safety stock based on failure consequences, you can manage inventory and downtime risk more precisely.

Martin Hörnig10 min read

In brief
  • Intermittent demand means many periods without consumption, with occasional demands in between, sometimes of similar size, sometimes varying widely. Smoothing and moving averages are systematically off here, usually too high right after a demand.
  • First separate out the predictable demand from maintenance and overhauls. For the rest, methods based on Croston fit, as do probabilistic forecasts, which show how likely each quantity is.
  • MAPE is undefined in periods without demand. Measure with KPIs that can handle zero values, and ultimately by availability and inventory.
  • The consequences of a failure determine how much safety stock a part needs. Only a comparison on your own data shows whether an AI model does better than Croston.

How to recognize intermittent demand

A fast mover is needed almost every week, in similar quantities. A typical spare part, by contrast, has no demand for months, then a demand for two units, then nothing again for a long time. You can describe this pattern by the average interval between two demands and the variation in demand quantities.

PatternCharacteristicsWhat usually fits
SmoothDemand in almost every period, similar quantitiesTraditional forecast models, such as those in the SAP material master
ErraticDemand in almost every period, widely varying quantitiesTraditional models or Croston variants; plan outliers such as large orders separately
IntermittentLong gaps, similar quantitiesCroston’s method and its refinements
LumpyLong gaps, widely varying quantitiesCroston variants or probabilistic forecasts, safety stock based on the consequences of failure

Syntetos, Boylan, and Croston defined the boundaries between the patterns in 2005. Gaps count as long when the average interval between two demands exceeds 1.32 periods. Quantities vary widely when their squared coefficient of variation, a measure of their spread, exceeds 0.49.

The pattern depends on the time bucket. A part needed every month can be intermittent on a weekly basis. So analyze demand in the time buckets you plan in.

Why averages and smoothing miss the mark

An example with assumed values: a part is needed three times in twelve months, once three units, once one, and once two. On average, that is half a unit per month. This quantity never occurs, and it says nothing about whether the stock will cover the next failure.

Exponential smoothing, such as the constant model in the SAP material master, jumps up after every demand and declines slowly in the months without demand. But orders are usually placed right after a demand, when stock has fallen. That is exactly when the forecast is too high, and too much is ordered.

Croston described this effect in 1972 and developed the method that bears his name to address it. If SAP’s automatic model selection finds no regular pattern, it picks the constant model, which is exactly this smoothing.

Safety stock suffers too. In automatic reorder point planning, SAP calculates it from service level, replenishment lead time, and the forecast’s mean absolute deviation.

The safety factor follows the normal distribution and assumes deviations that scatter symmetrically around the mean. This assumption does not fit intermittent demand, which is usually zero and only rarely much higher.

Separate out predictable demand first

Not every spare parts requirement is random. Parts for maintenance plans, inspections, and planned shutdowns are often known weeks in advance. In SAP, they enter deterministic MRP through the reservations of maintenance orders.

In reorder point planning, such reservations only count if the MRP type includes external requirements, in the standard system V1 or V2. Configuration determines which reservations are included. Otherwise, materials planning only notices the planned demand at goods issue.

For materials planned with deterministic MRP, SAP records unplanned consumption alongside total consumption, that is, withdrawals without a reservation or beyond the reserved quantity. In the configuration of the MRP type, you define whether the forecast uses total consumption or only unplanned consumption. That way, you only forecast what is truly unplanned.

Methods for intermittent demand

Croston’s method separates the size of a demand from the interval to the next one. It smooths both values separately and only after an actual demand; the forecast per period is the size divided by the interval. For the example above, that means two units roughly every four months.

Mathematically, it is still half a unit per month. The difference lies in the behavior over time: the forecast does not automatically jump up after a demand; it stays flat in the empty months. And planners see what to prepare for, namely a rare demand of two units.

MethodIdeaStrengthLimitation
CrostonSmooth demand size and interval separatelyNo jump after every demandSlightly overestimates demand and stays unchanged during long gaps
Croston with the Syntetos-Boylan correctionCroston with a correction factorLargely removes the overestimation and is easy to implementRemains a single number with no information on spread
TSB, named after Teunter, Syntetos, and BabaiProbability of a demand instead of the interval, updated every periodDeclines when a part is no longer needed, which reveals phase-out partsAlso reacts to long gaps that occur by chance
Probabilistic forecastProbability of each demand quantity during the replenishment lead time, from a distribution or from your own historyDirectly provides the stock level for a target service levelNeeds enough history or well-founded assumptions

What ultimately counts for materials planning is which stock level lasts through the replenishment lead time, and with what probability. Probabilistic forecasts answer that directly.

Simplified, with the example above and a replenishment lead time of one month: in three out of four months, there is no demand; otherwise, there is one demand of one, two, or three units. One unit in stock then covers demand in just over eight out of ten months, two units in just over nine out of ten. The criticality of the part decides which rate is enough.

In SAP, you store the result, for example, as the reorder point in manual reorder point planning and update it regularly via mass maintenance.

Measure forecast accuracy when demand is mostly zero

MAPE, the mean absolute percentage error, is a common KPI for forecast accuracy. It divides each error by the actual demand. In periods without demand, that means dividing by zero, so MAPE is undefined there.

If you leave these periods out, MAPE only rates the months with demand. A forecast that builds up stock in the empty months then goes unnoticed. For parts with regular demand, MAPE remains a good, easy-to-understand KPI.

For intermittent demand, other measures are better suited, such as MASE, which Rob Hyndman proposed for this purpose in 2006. MASE and WAPE have a limitation too: with many zeros, a forecast of zero often scores well, although it never supplies a part. That is why bias and simulation belong in the evaluation.

KPIWhat it showsIn periods without demand
WAPE, the weighted MAPESum of errors relative to the sum of demandComputable as long as there is some demand in the evaluation window; favors forecasts near zero
MASE, the mean absolute scaled errorError relative to a simple benchmark forecast; below 1 means more accurate than the benchmarkComputable, but favors forecasts near zero
Bias, the systematic errorWhether the forecast is too high or too low on averageComputable; too high means inventory buildup
CoverageWhether a stock level meant to suffice in nine out of ten cases actually did so that often in hindsightComputable, for probabilistic forecasts
SimulationInventory and service level that the forecast would have producedYardstick for the decision

What matters is what the forecast does in the warehouse. So run each method on your own history and compare the inventory and service level it would have produced.

Base safety stock on the consequences of failure

A uniform service level for all spare parts is misleading. A part whose absence stops a line needs different protection than one with a workaround. Maintenance and production define criticality, and materials planning puts it into practice.

Service level can mean two things: the probability of no stockout during the replenishment lead time (cycle service level), or the share of quantity delivered immediately from stock (fill rate). SAP describes the material master field as a share of quantity but takes the safety factor from the normal distribution, as for a probability. Decide which measure applies, and measure exactly that one.

Consequence of a shortageReplenishment lead timeStocking
Equipment down, safety or environment affectedLongStock it, even without consumption in the history; check condition and shelf life regularly
Equipment downShortSmall stock or agreed express delivery, for example through an outline agreement
Reduced output, workaround possibleLongStock based on forecast and service level, with a lower target
No impact on operationsShortDo not stock; procure when needed

Where several plants use the same parts, a pooled stock can cover all of them if transport is fast enough. For the same protection, it usually needs fewer parts than many separate stocks.

For parts with regular consumption, on the other hand, the levers lie more in lot sizes and safety stocks.

When AI forecasts pay off

AI models can use drivers that a consumption history does not contain, such as equipment run times and age, maintenance history, or planned shutdowns. With intermittent demand, however, their data basis is thin too.

In a project on AI-based demand forecasting for spare parts, we served as project lead and product owner. We directed forecasting, simulation, and the measurement of forecast accuracy from the business side and signed off on the results.

Our standard for such projects: an AI model goes live only when it performs measurably better than simpler methods on the same history. It is measured by inventory and service level, not by forecast error alone. In a pilot project, you should insist on these points.

  • Comparison with today’s SAP settings and with Croston variants on the same history
  • KPIs that can handle periods without demand, plus bias
  • Simulation of inventory and service level, broken down by demand pattern and criticality
  • Explanations that planners can follow
  • A defined path for the results into the SAP system, for example as a reorder point
  • Clear ownership in operations, that is, who monitors forecast accuracy and who retrains the model

Frequently asked questions

Can SAP forecast intermittent demand using Croston’s method?

Not with the traditional forecast models in the material master. There, you choose from constant, trend, and seasonal models, moving averages, and second-order exponential smoothing.

Planning tools such as SAP Integrated Business Planning (IBP) offer Croston variants. Their results, such as reorder point or safety stock, reach the material master through an interface or mass maintenance.

What should we do with parts that have not been needed for years?

First check whether the equipment the part is intended for is still running. If not, the part is a candidate for sale or scrapping; if it is, criticality decides. Methods such as TSB make such parts visible early, because their forecast declines during long gaps.

How much history does a spare parts forecast need?

Enough for several demands per part; for spare parts, that often means several years. If there is no history, as with new equipment, comparable parts help, as do manufacturer data on wear and the number of installed parts.

Does this also apply to spare parts we sell to customers?

Yes, the methods are the same. The service level is then set by the availability you promise your customers instead of the consequences of failure in your own plant. Additional drivers are the number of units in the field and the phase-out of a product line.

How we help

We segment your spare parts by demand pattern and measure the accuracy of today’s forecast. Maintenance and production define the consequences of failure, and we translate them into reorder points and safety stocks. The first step is the analysis of inventory and MRP parameters.

If you are planning an AI forecasting project, we take on project leadership and product ownership on your side. The checkpoints above then become the criteria for sign-off. Both are part of our supply chain planning and inventory optimization service.

Initial call: 30 minutes, free of charge.

Tell us where things get stuck in materials planning and the warehouse. We will tell you where we would start and whether we are the right fit.

Initial call Call