How to Forecast Demand for Industrial MRO Supplies
A machine is down. The spare part you need has an OEM lead time of 20 weeks. Intermittent, failure-driven consumption makes standard forecasting methods unreliable, so you need a framework built on criticality and failure data.
Most procurement teams fail because they treat MRO forecasting like finished-goods forecasting. We see the consequences daily: buyers scrambling for obsolete components while their SAP systems show healthy stock levels of parts they will never use again.
Why Standard Forecasts Fail for MRO Spare Parts
Extrapolating historical consumption fails to account for the intermittent, failure-driven nature of industrial parts. A part that fails once every 18 months does not follow a normal distribution curve.
Traditional high-frequency transaction models assume steady, repeatable purchases. MRO demand is different. As Verusen's analysis of MRO demand trends notes, demand is intermittent and failure-driven, rendering those consumer-based models ineffective. Predictive maintenance events, not consumption patterns, should drive your forecasts. Without this shift, you carry either excessive safety stock or face stock-out risk on mission-critical components.
Start With Criticality, Not Consumption History
Prioritizing inventory based on asset criticality is the foundational step. A $50 sensor that stops a production line is more important than a $5,000 drive with a redundant backup. Classify every part by the operational impact of its failure before assigning any forecasting model.
This contradicts the instinct to focus on high-value items. If a component is critical, its forecast should favor availability over inventory turnover.
Step 1: Classify Your Inventory With ABC-XYZ Analysis
The ABC-XYZ classification is the backbone of any serious MRO forecasting framework. It combines the value of consumption (ABC) with the variability of demand (XYZ).
ABC Analysis ranks items by annual consumption value. XYZ Analysis ranks items by demand variability. Their intersection tells you exactly how to manage each SKU.
| Class | Demand Pattern | Forecasting Approach | Replenishment Strategy |
| AX | High value, steady demand | Statistical forecasting, low safety stock | Regular reorder points |
| AY | High value, moderate variability | Time-series with smoothing | Higher safety stock, monitor trends |
| AZ | High value, intermittent demand | Failure-driven or ML models | Criticality-based stock levels |
| BX | Medium value, steady demand | Simple moving average | Standard reorder points |
| CY | Low value, variable demand | Manual review | Batch ordering |
| CZ | Low value, erratic demand | No forecast needed | Reorder on failure or use Request Board |
This framework helps you avoid applying one model to every SKU. For CZ items, forecasting is a waste of time. For AZ items, it is a matter of operational survival.
Step 2: Model Failure-Driven Demand Signals
Once you have classified your inventory, replace consumption-based forecasts with failure-driven ones. The trigger for replenishment should be a maintenance event, not a periodic stock check.
The goal is to link mean time between failures (MTBF) data to your replenishment triggers. If a servo drive has a known MTBF of 5 years and you have 10 in operation, you should plan for roughly 2 failures per year. This stochastic modeling approach is far more accurate than looking at what you purchased last year. Research on forecasting methods and models for maintenance inventory confirms that tailored forecast support systems improve maintenance inventory management outcomes.
The mathematics of failure-driven demand
For components with a constant failure rate, the Poisson distribution is the standard model. The probability of observing exactly k failures in a given period is:
P(k) = (λ^k * e^-λ) / k!
Where λ is the expected number of failures over that period. If you have 10 drives with an MTBF of 5 years, your expected failure rate is 2 per year. The probability of experiencing 3 or more failures in a single year is roughly 32%. That is not a tail risk; it is a near-certainty over a 3-year horizon. Your safety stock calculation must cover this variance, not just the average.
A common mistake is ignoring lead time variability. A part with a 20-week lead time requires different stock levels than one with a 2-week lead time.
The lead-time demand formula
The standard calculation for reorder point under uncertainty is:
Reorder Point = (Expected daily demand × Lead time in days) + Safety stock
For failure-driven parts, expected daily demand is your annual failure rate divided by 365. Safety stock is a function of the standard deviation of demand during lead time and your desired service level. For a 95% service level, the z-score is 1.65. For 99%, it is 2.33.
Consider a Siemens 6SE7021-8EB87-0AA1 drive with a 20-week lead time from the OEM. Your plant runs 12 of these drives, each with a documented MTBF of 6 years. Your expected failure rate is 2 per year. Lead time is 140 days. Expected demand during lead time is 0.77 units. If the standard deviation of demand during lead time is 0.9, your safety stock for a 95% service level is 1.65 × 0.9 = 1.5 units. Round up: you need a reorder point of 2 units on hand. If you only stock 1, you have a roughly 30% chance of a stock-out before the replacement arrives.
Forecasting during supply-chain volatility
The failure-rate math above assumes a stable lead time. When global logistics are disrupted, a 20-week lead time can stretch to 30 or 40 weeks without warning. Your forecast must include a lead-time inflation factor based on current supplier performance.
If your OEM quoted 20 weeks but delivered in 26 for the last three orders, your planning lead time is 26 weeks, not 20. Recalculate your reorder point with the actual lead time.
For parts already flagged as obsolete or end-of-life, the lead-time question becomes moot. Your forecast should trigger a sourcing decision before the last unit is consumed. When your model identifies a critical drive with a 26-week effective lead time and only 1 unit in stock, the forecast has done its job. The next step is finding a verified alternative source.
Step 3: Use Machine Learning for Intermittent Demand
Standard statistical models break down on intermittent demand. Machine learning algorithms, specifically Support Vector Machines (SVM), have proven effective in addressing these sparse demand patterns. A case study on SVM for [MRO inventory(https://automa.net/blog/best-strategies-managing-mro-inventory) demand forecasting | researchgate.net] demonstrates how these models handle the erratic data that defeats traditional methods.
For most plants, this data lives in the CMMS, the ERP, and the maintenance logs.
The human-in-the-loop gap no algorithm closes
Every ML model deployed in European plants shares the same blind spot: it ignores the tacit knowledge of maintenance technicians who have worked on the line for 15 years. They know that the KUKA robot on line 3 fails more often in August when the ambient temperature rises.
A pure ML approach treats these insights as noise. A human-in-the-loop approach treats them as features.
A practical workflow for combining ML output with technician judgment
Set up a monthly review meeting with the maintenance team, not just procurement. Review the ML forecast for the next 90 days for your top 20 critical parts. Ask one question: does this match what you see on the floor?
The VFD on the packaging line has been making a high-pitched whine for two weeks. The spare part you stocked for the CNC spindle is the wrong revision; the machine was retrofitted last year. These are demand signals no historical dataset contains because they are happening now.
Capture these adjustments in the CMMS as structured notes. Over time, you build a feedback loop: the ML model learns from the technician's corrections, and the technician sees their input reflected in better stock availability.
The data you need to make ML work
Before you invest in any ML tool, audit your data quality. The minimum viable dataset for a single part includes:
- Failure date and time
- Asset identifier and location
- Part number and manufacturer
- Failure mode (from your CMMS codes)
- Maintenance action taken (repair vs. replace)
- Lead time from order to receipt
- Technician notes (free text)
Most plants have this data scattered across the CMMS, the ERP, and Excel spreadsheets. The ML model is only as good as your ability to join these tables. Start by standardizing your failure codes and requiring the maintenance team to select from a dropdown rather than typing free text.
When ML is the wrong tool
For CZ-classified parts with near-zero consumption, ML adds nothing. The demand is so sparse that no algorithm can find a pattern. The correct forecast is simple: reorder on failure, or use a marketplace request board when the failure occurs.
For AZ parts with a genuine failure history, ML earns its keep. But if your MTBF data is unreliable because maintenance events were not logged for two years, the model will produce confident nonsense.
Step 4: Apply BOM Analysis to Maintenance Procurement
BOM analysis for maintenance procurement helps you understand which components are shared across multiple assets. When a single part number appears in 15 different machine BOMs, its criticality is higher than its individual failure rate suggests.
You cannot forecast demand for a part that no longer exists in your supplier's catalog, but you can forecast the risk of needing it. That risk should trigger a decision: find a verified alternative, stock the last available units, or plan a retrofit.
Oliver Wyman's aviation MRO survey highlights that supply chain strain is increasing across industries, with spare parts demand forecasting becoming a cost management priority. For obsolete parts, the forecast horizon is not a quarter. It is the remaining asset lifecycle.
Step 5: Managing Obsolete Automation Component Inventory
Obsolete inventory requires a different decision framework than active stock. The goal is to minimize both the cost of carrying dead stock and the cost of a machine down with no replacement available.
For components in active production assets, the question is whether to stock the last buy or seek a verified alternative. For components already replaced, surplus stock should be monetized rather than written off. When your forecast identifies a part as high-risk due to obsolescence, the solution is not a better statistical model. It is a verified source for the physical part.
The Forecasting Framework in Practice
The complete framework is straightforward: classify by criticality, model by failure data, apply ML to intermittent demand, and use BOM analysis to catch shared and obsolete components. The output is a set of replenishment triggers and sourcing strategies tailored to each inventory class.
| Step | Action | Output |
| 1 | ABC-XYZ classification | Inventory segmentation |
| 2 | Failure-driven demand modeling | MTBF-based replenishment triggers |
| 3 | ML for intermittent demand | Improved forecast accuracy on critical spares |
| 4 | BOM analysis | Shared component and obsolescence risk map |
| 5 | Obsolete inventory strategy | Last-buy vs. verified alternative decisions |
The numbers show why this matters. Global aviation MRO spend is projected to reach $119 billion by 2034 according to IATA's MRO expenditure forecast, and the global aircraft fleet is expected to grow 37% by 2036 per Oliver Wyman's fleet growth analysis. The pressure on spare parts availability is only increasing.
The hard part is not the forecasting method. When your analysis flags an obsolete drive with a 20-week lead time, you do not need better data. You need a verified source for the physical part, today. That is where AutomaSEARCH connects your forecast to real, in-stock inventory across a verified network of suppliers.
Frequently Asked Questions
How do you distinguish between independent and dependent demand for MRO parts?
Independent demand for MRO parts comes from unpredictable events like machine breakdowns or preventive maintenance schedules. Dependent demand is tied to a specific production plan, like a filter changed every time a machine runs a batch. For forecasting, treat them separately: use failure data and maintenance schedules for independent demand, and link dependent demand directly to your production forecast.
What are the primary challenges in forecasting demand for discontinued PLC and drive components?
Discontinued components have no historical consumption data and an unpredictable failure rate. The main challenges are long lead times (often 20 weeks or more), the risk of stock-outs on a critical machine, and the high carrying cost of a part that may never be used. This is where managing obsolete inventory requires a different approach than standard forecasting.
How does machine criticality influence MRO inventory forecasting?
Machine criticality determines the cost of a stock-out. For a critical machine, the cost of downtime far exceeds the carrying cost of a spare part. Your forecast must prioritize these items with higher safety stock and faster replenishment triggers. For non-critical machines, you can accept a higher stock-out risk and keep lower inventory levels.