The Intermittent Demand Problem

If you have ever tried to forecast demand for semiconductor equipment spare parts, you know the special frustration I am about to describe. You are staring at 24 months of demand history for a critical component, and it looks like this: 0, 0, 0, 0, 3, 0, 0, 0, 0, 0, 0, 12, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0.

Mostly zeros. Occasional spikes. No discernible pattern. And yet, when demand does appear, it is urgent — because somewhere a $50 million fabrication tool is down, and every hour of downtime costs $100,000 or more in lost production.

This is the intermittent demand problem, and it is one of the hardest forecasting challenges in supply chain analytics. Get it wrong in one direction and you are sitting on millions of dollars in excess inventory. Get it wrong in the other direction and you are explaining to a customer why their fab line is down for a week while you expedite a part from the other side of the world.

Why Traditional Methods Fail

Let me start with what does not work, because I have tried all of it.

Moving averages produce forecasts dominated by zeros. A 12-month moving average of the demand series above gives you roughly 1.4 units per month — which is never correct. Demand is either zero or significantly higher. The average is a number that never actually occurs.

Exponential smoothing has the same problem, just with recency weighting. It reacts sluggishly to demand events and takes months to decay back after a spike. The forecast lags reality in both directions.

Standard regression struggles because the relationship between predictors and demand is not continuous. You are not trying to predict "how much" so much as "will there be any, and if so, how much" — a fundamentally two-part question.

The core issue is that intermittent demand violates the assumptions underlying most forecasting methods. These methods assume demand is a continuous, reasonably smooth signal with some noise. Spare parts demand is mostly silence with occasional shouts.

Croston's Method: Better, But Not Enough

Croston's method was designed specifically for intermittent demand, and it is a significant improvement over naive approaches. The insight is simple but powerful: decompose the forecast into two separate models.

Demand interval — how often does demand occur? Model the time between non-zero demand events.
Demand size — when demand occurs, how large is it? Model the magnitude of non-zero demands.

The forecast is then: expected demand size divided by expected demand interval.

This works well for moderately intermittent patterns. It avoids the "averaging zeros with spikes" problem by treating the two components separately. And it is easy to implement and explain, which matters in operational settings.

But Croston's method breaks down in several important ways for semiconductor spare parts:

It assumes stationarity. The demand interval and size are assumed to be stable over time. But in semiconductor, the installed base is constantly changing — new machines deployed, old machines retired, utilization patterns shifting with market cycles.
It ignores causal factors. Croston treats demand as a purely time-series phenomenon. It does not consider why demand occurs — machine age, utilization intensity, preventive maintenance schedules, technology node transitions.
It struggles with lumpy demand. When demand sizes vary dramatically (2 units one month, 50 the next), the demand size model produces unreliable estimates.

The ML Approach: Features Matter More Than Algorithms

Here is what I have learned from applying machine learning to this problem: the installed base is your best predictor. Knowing which machines are running, their age, and their utilization tells you more than years of demand history.

Think about it from first principles. A spare part fails because a machine is running and the component has degraded. The probability of failure depends on:

Machine age and vintage. Older machines fail more frequently. Different vintages have different failure profiles.
Utilization intensity. A tool running 24/7 wears parts faster than one running 8 hours a day.
Installed base size. More machines in the field means more potential demand events.
Technology node. Different process nodes stress different components differently.
Maintenance history. Machines with recent preventive maintenance have different failure probabilities than those that are overdue.

When you build ML models with these features — rather than just historical demand — the prediction accuracy improves dramatically. You are modeling the demand generation process, not just the demand signal.

The specific algorithms matter less than you might think. I have seen good results with gradient-boosted trees (XGBoost, LightGBM), random forests, and even well-tuned logistic regression for the demand occurrence component. The magic is in the features, not the model architecture.

The Ensemble Approach

The approach I have found most effective in practice is an ensemble that combines statistical and ML methods:

Croston's method provides a baseline forecast grounded in demand history. It captures the inherent demand rhythm for each SKU.
An ML model trained on installed base features provides a causal forecast. It captures the structural drivers of demand.
A combining model learns the optimal weighting of these two forecasts for different SKU profiles. High-volume SKUs lean toward Croston. Low-volume SKUs with strong installed base signals lean toward ML.

This ensemble consistently outperforms either method alone. Croston anchors the forecast in observed patterns. The ML model adjusts for known changes in the demand generation environment. The combination captures both backward-looking and forward-looking information.

Safety Stock: Where Forecasting Meets Finance

Forecasting accuracy matters because it directly drives safety stock decisions, and in semiconductor spare parts, the financial stakes are enormous.

Too much safety stock means millions of dollars tied up in inventory that may never be used. Spare parts can become obsolete when machines are retired or upgraded. Capital locked in excess inventory is capital not invested in growth.

Too little safety stock means stockouts. And a stockout on a critical spare part does not mean a back-order with a minor delay. It means a fabrication tool is down, a production line is stopped, and the meter is running at six figures per hour. One significant stockout can cost more than a year of carrying excess inventory.

The optimal safety stock level depends on forecast accuracy — specifically, on the forecast error distribution. Better forecasts produce tighter error distributions, which enable lower safety stocks at the same service level. This is why investing in forecasting capability for intermittent demand has such high ROI: every percentage point of forecast accuracy improvement translates directly into inventory reduction without service level degradation.

Practical Recommendations

If you are tackling intermittent demand forecasting in a spare parts context, here is where I would start:

Build your installed base dataset first. Before touching any forecasting model, invest in knowing what equipment is in the field, where it is, how old it is, and how hard it is running. This data is your single biggest lever.
Implement Croston as your baseline. It is fast to deploy, easy to explain, and better than whatever moving average approach you are currently using.
Layer ML on top, do not replace. Use ML models to adjust the Croston baseline using installed base features. The ensemble will outperform either method alone.
Segment your SKUs. Not every part needs the same forecasting approach. High-volume parts work fine with statistical methods. Focus your ML investment on the high-value, highly intermittent parts where the cost of error is greatest.
Measure what matters. Forecast accuracy metrics like MAPE are meaningless for intermittent demand (you cannot divide by zero demand). Use metrics designed for intermittent series: scaled MAE, hit rate (did you predict non-zero when demand was non-zero), and inventory-weighted bias.

The intermittent demand problem will never be fully solved — the inherent randomness of component failures ensures that. But the gap between what most organizations do today and what is achievable with the right methods and data is enormous. Closing that gap is some of the highest-value analytics work in supply chain.