Data Department built a probabilistic spare parts forecasting system for a large automotive distribution group that manages more than 100,000 active parts across several brands and sales channels. It was a 12-month project, and the system was built for daily forecast updates. It improved forecast accuracy by about 30% over the legacy method and lowered inventory holding cost by about 10%.
The decision
Every day, planners decide how much of each part to hold, and where. Hold too little and a vehicle waits in an authorised service bay. Hold too much and capital sits on shelves for years.
The group sold parts through authorised service, dealers and direct channels. Its legacy method produced a single number per part: one point forecast. A point forecast says nothing about how uncertain it is. A planner looking at "12 units next month" could not tell whether the real range was 10 to 14 or 0 to 40, so safety stock was set by habit rather than by risk.
The question we were asked to answer was not "can you forecast better?" It was "how much stock should we hold for each part, given what we do not know?"
What we built
We replaced point forecasts with probabilistic forecasts: a prediction interval for every part, so planners see both the expected demand and how wide the uncertainty is.
One model does not fit 100,000 parts. We segmented the catalogue and used a different approach for each behaviour:
| Part segment | Model | Why |
|---|---|---|
| Frequent movers | Gradient boosting with seasonality | Enough history to learn patterns directly |
| Sporadic or intermittent parts | Hierarchical Bayesian models using substitution data | Sparse history; borrows strength from related and substitute parts |
| High-value parts | Conservative models tuned for availability | A stock-out here is expensive, so the model leans towards availability |
The models used order history, the vehicle park by region, part compatibility and substitution, seasonality and weather, and sales across all channels.
The system was built on AWS (Amazon Web Services) using SageMaker, in containers. Forecasts were designed to update daily and feed directly into the group's ERP (enterprise resource planning) and planning tools, so planners could work in the systems they already use.
Results
- Engagement: 12-month project.
- ~30% better forecast accuracy than the legacy method.
- ~10% lower inventory holding cost.
- Interpretable confidence bands for planners, so stock levels can be set by the risk the business is willing to accept for each part.
What made it work
We started from the stocking decision, not the model. Planners needed to know how much uncertainty surrounded each forecast. A more accurate single number would not have changed how they set safety stock. Prediction intervals did.
Segmentation mattered more than any single algorithm. Intermittent parts behave nothing like fast movers. Treating them separately, and using substitution data to fill gaps in sparse history, is where much of the gain came from. We explain the reasoning in more detail in our article on probabilistic forecasting for spare parts.
Different parts carry different risks. For high-value parts we deliberately chose conservative models. A small loss in theoretical accuracy was worth more availability where a missing part is costly.
Integration was part of the scope from day one. Daily updates inside the ERP and planning tools were designed to bring the forecasts to planners where they work, rather than leave them in a report.
This type of work is the core of our demand forecasting and inventory optimisation service, and a common starting point for distribution companies.
Frequently asked questions
What is a probabilistic forecast?
A probabilistic forecast gives a range of likely outcomes with probabilities, not a single number. For spare parts, this lets planners set stock levels based on how much stock-out risk they accept for each part.
How do you forecast parts that sell only a few times a year?
For sporadic and intermittent parts we used hierarchical Bayesian models. They borrow information from related and substitute parts, so a part with little history still gets a sensible forecast and an honest uncertainty range.
Did the system replace the planners?
No. The system was designed to give planners daily forecasts with confidence bands inside their existing ERP and planning tools. Planners make the stocking decisions, with better information.