Spare parts demand forecasting is the art of predicting how many of each service part you'll need, where, and when, so you stock the right quantity in the right place. Get it right and you barely notice it. Get it wrong and you feel it twice: money frozen in parts nobody needs, and a customer waiting on the one part you don't have.
That's the frustrating thing about after-sales inventory. Most teams manage to overspend and still hit stockouts at the same time. It sounds contradictory until you look at where the money actually sits. It's piled up in slow-moving parts in the wrong regions, while the fast movers run dry in the markets that need them. The total inventory value looks fine on a spreadsheet. The service level tells a different story.
Machine learning changes this, but not in the vague "AI makes everything better" way. It changes it because spare parts demand has a specific shape that traditional forecasting was never built to handle. Let's walk through why, and what a proper spares planning and forecasting approach does about it.
Why spare parts demand is so hard to forecast
Finished-goods demand is relatively smooth. You sell roughly so many units a month, seasonality nudges it up and down, and a moving average gets you close. Spare parts don't behave like that at all.
Intermittent and lumpy demand patterns
Most service parts are what planners call intermittent. A given part might see zero demand for months, then three orders in a week because a batch of units hit the same failure point, then nothing again. The average across the year might be one unit a month, but no month ever actually sees "one unit." Averaging that number tells you to hold one in stock, which is simultaneously too many for the quiet months and far too few for the spike.
Classic moving-average and exponential-smoothing methods hate this pattern. They smooth right over the lumps, which is exactly the wrong instinct for parts where the lumps are the whole point. You end up with safety stock set by a formula that assumes demand it never actually sees.
Then layer on the installed base. Demand for a repair part isn't random. It's driven by how many units are in the field, how old they are, and where they are in their failure curve. A part fails more as the installed base ages, then tapers as those units get retired. A forecasting method that only looks at past order history is blind to all of that.
How machine learning improves spares forecasting accuracy
Machine learning models are good at exactly the things that break traditional forecasting. They can learn from many signals at once, not just the order history for a single part.
A well-built ML demand forecasting model looks at the installed base by region and age, historical failure curves for similar parts, seasonality, and the relationships between parts that tend to fail together. From that it produces a forecast at the level you actually plan at, which is the specific part in the specific location, not a national average that hides all the regional variation.
The accuracy gain matters most for the hard cases. For a smooth, high-volume part, honestly, a simple method does fine and ML barely helps. The value shows up in the long tail of intermittent parts, which is where most of your SKUs and most of your trapped capital live. Getting those even a little more right, across thousands of part-location combinations, is what moves fill rate and working capital at the same time.
One caveat worth saying out loud: ML forecasting is not magic and it is not set-and-forget. It's only as good as the data feeding it. If your installed base data is stale or your failure records are inconsistent, the model learns garbage. The forecasting is the easy part. The data discipline underneath it is the real project.
From forecast to plan: SBOM supply planning and inventory allocation
A forecast on its own doesn't stock a warehouse. You need to turn "we'll probably need this many" into "order these, hold them here." That translation runs through the service bill of materials.
SBOM supply planning links each repairable product to the specific parts used to service it. So when your forecast says you'll do a certain number of repairs on a given model, that demand cascades down into part-level requirements automatically. Without an SBOM, planners do this mapping in their heads or in spreadsheets, and it breaks the moment a product has more than a handful of serviceable parts.
From there, allocation decides where the stock physically goes. This is the step teams most often get wrong. They forecast well, then dump inventory into a central hub and let it sit far from where repairs happen. Good inventory allocation logic pushes parts toward the regions and vendors that actually generate the demand, balancing service level against carrying cost location by location.
Distributing parts to service vendors without over-stocking
If you work through third-party service vendors, and most global operations do, the distribution problem gets sharper. Every vendor wants a comfortable buffer of stock because it makes their life easier. Multiply a comfortable buffer across dozens of vendors in dozens of countries and you've quietly built a second full inventory that nobody's tracking.
The fix isn't to starve vendors. It's to plan their stock centrally against real demand signals, replenish based on consumption, and keep visibility of what's sitting where. Our platform plans allocation and vendor distribution across more than 70 countries, matching supply to regional service demand rather than to whoever asked loudest. The goal is boring in the best way: the right vendor has the right part when the repair comes in, and no vendor is sitting on six months of a part that fails once a year.
KPIs that matter: fill rate, stockouts, carrying cost
You can't improve what you don't measure, and spares planning has a few numbers that actually matter.
- Fill rate is the headline: what fraction of demand you satisfy from stock on hand.
- Stockout frequency and duration tell you where service is failing.
- Carrying cost, including the working capital tied up, tells you what your service level is costing you.
- Obsolescence, the parts you'll eventually write off, is the one everyone forgets until year-end.
The trap is optimizing one in isolation. Push fill rate to 99 percent everywhere and your carrying cost explodes. Slash inventory and your stockouts spike. The whole point of good forecasting and allocation is to move the trade-off curve itself, so you get better fill rate and lower cost together, rather than just sliding along the old curve. That's the number to hold a platform accountable to.
Spares planning rarely gets executive attention until a stockout embarrasses someone or a CFO notices the write-offs. But it's one of the highest-leverage things an after-sales operation can fix, because the same money can either sit dead on a shelf or fund the rest of the business. If you want to see where your spares budget is actually going, that's the conversation worth having with our team.
The data you actually need to forecast well
Before any model earns its keep, it needs the right inputs, and this is where most forecasting projects quietly succeed or fail. The order history everyone starts with is necessary but nowhere near sufficient, because past orders only tell you what you happened to stock, not what customers actually needed. If you stocked out, the true demand is invisible in the order data, which trains the model to under-forecast the very parts that already hurt you.
The signals that make a real difference are:
- Installed base by region and age, so the model knows how many units are in the field and where they sit on their failure curve.
- The bill of materials linking products to serviceable parts.
- Failure and repair records, ideally clean enough to see which parts fail together.
- Lead times, because a forecast is useless if you can't act on it in time.
Master data quality underpins all of it. If your part numbers don't reconcile across systems, or the same physical part exists under three codes, the model learns noise. This is the unglamorous groundwork, and it's why we treat data readiness as the first conversation, not an afterthought once the model disappoints.
Balancing service level against working capital
Spares planning is fundamentally a balancing act, and the balance point isn't the same for every part. The instinct to set one target service level across the board, say 95 percent fill rate everywhere, is expensive and wrong. A cheap, non-critical part and a part that gates an expensive repair should not get the same treatment.
Better practice segments parts by criticality and cost. High-criticality parts, where a stockout means an idle machine or a stranded customer, warrant protective stock even when they move slowly, because the cost of being out dwarfs the cost of holding one. Low-criticality, easily-sourced parts can run lean, because a short wait costs little. Getting this segmentation right is where a lot of trapped capital gets freed: you're not carrying less inventory across the board, you're carrying the right inventory, more of what protects service and less of what just sits there. That's the shift from managing an average to managing a portfolio, and it's what actually moves both fill rate and working capital at the same time.
Why forecasting and warranty belong together
There's a quiet advantage to running spares forecasting on the same platform as warranty and after-sales, and it's worth naming because most operations run them separately. Warranty claims are a live signal of what's failing, in which markets, on which products, right now. Fed into forecasting, that signal sharpens demand prediction well ahead of what order history alone would show, because a rise in claims for a part precedes the rise in replenishment orders for it.
When forecasting shares data with warranty, master data, and ERP, the whole after-sales loop tightens. Failures inform forecasts, forecasts inform stocking, stocking supports repairs, and repairs feed back as claims. Run these as disconnected systems and you lose that feedback, forecasting blind to what warranty already knows. Run them together and each makes the other smarter, which is a large part of why integrated after-sales platforms outperform a stack of point tools on both service and cost.
Frequently asked questions
What is spare parts demand forecasting?
It is predicting future demand for service and repair parts so you stock the right quantity in the right location. ML models learn from history, product installed base, and failure patterns to handle irregular demand.
Why is machine learning better than traditional forecasting for spares?
Spare parts often have intermittent, lumpy demand that classic moving-average methods handle poorly. ML models capture seasonality, installed-base signals, and failure curves for more accurate, location-level forecasts.
What is SBOM supply planning?
Service Bill of Materials (SBOM) planning links each repairable product to its serviceable parts, so demand for finished repairs cascades into accurate part-level supply plans.
How does forecasting reduce inventory cost?
Better forecasts let you lower safety stock and avoid both stockouts and dead inventory, improving fill rate while freeing working capital tied up in overstock.
Can it allocate inventory across service vendors and countries?
Yes. Atomquark's platform plans allocation and vendor distribution across 70+ countries, matching supply to regional service demand.
Does this integrate with warranty and ERP systems?
It integrates with warranty data, master data, and ERP so demand signals, repairs, and replenishment stay in sync.