Why Your Forecasts Fail for Sporadic Items and What Croston’s Method Fixes

March 6, 2026

If you have ever tried forecasting a product that barely sells, you already know the frustration. Most weeks, nothing. Then suddenly, three units go out. Then silence again.

Traditional forecasting methods do not handle this well. They expect some kind of regular pattern. Even a messy one. But intermittent demand does not behave. It is uneven, unpredictable, and honestly a bit annoying to work with.

You end up with forecasts that feel wrong.

Either the system predicts tiny fractions every week, which never actually happen, or it overreacts to a random spike. Neither helps when you are trying to decide how much stock to keep.

So what counts as intermittent demand?

A simple way to think about it is this. If your demand has lots of zero periods, you are dealing with intermittent demand.

Not just low demand. That is different.

Intermittent demand means gaps. Weeks or months with no sales at all, followed by occasional bursts. This is common with spare parts, specialized components, or niche products.

For example, imagine you manage inventory for replacement machine parts. One specific valve might sell 5 units in a single week, then nothing for the next two months. That is not unusual. It is just how the demand behaves.

Trying to smooth that out using standard methods creates noise instead of insight.

Why standard forecasting struggles

Most basic models, like simple exponential smoothing, assume demand shows up every period. Even if the numbers are small, they expect continuity.

That assumption breaks here.

When you feed intermittent data into these models, the zeros pull the forecast down. Then a sudden sale pushes it up again. The model keeps chasing these changes but never really captures the pattern.

It is like averaging silence with sudden noise and hoping it becomes meaningful.

You end up with forecasts that suggest selling 0.3 units per week. Which sounds precise, but is not useful in practice.

How Croston’s method looks at it differently

Croston’s method takes a step back and splits the problem into two parts.

Instead of treating demand as one messy series, it separates:

how often demand happens
how big the demand is when it does happen

This small shift makes a big difference.

First, it looks at the size of demand, but only when a sale actually occurs. So it ignores all the zero periods for this part.

Then, it tracks the intervals between those non-zero demands. How many periods passed before the next sale showed up?

Each of these is smoothed separately. Then combined to produce a forecast.

It sounds a bit technical, but the idea is simple. Do not mix zeros and actual demand into one average. Treat them as different signals.

A quick example

Let’s go back to that spare valve.

Say demand looks like this over several weeks:

0, 0, 4, 0, 0, 0, 2, 0, 5

A traditional method tries to average all of that together. Croston’s method ignores the zeros when estimating demand size, so it focuses on 4, 2, and 5.

Then it looks at the gaps. Three weeks before the first demand, then four weeks, then two.

Now you have two clearer pieces of information. When demand happens, it is usually a few units. And it happens every few weeks, not every week.

That is much closer to reality.

What about safety stock?

This is where things often get messy in real life.

With intermittent items, uncertainty is higher. Not just in quantity, but in timing. You do not know when the next order will come, only that it will.

So safety stock cannot be treated the same way as fast-moving items.

In practice, many teams lean a bit conservative here. They hold slightly more stock than the forecast alone suggests, especially if stockouts are costly.

At the same time, overstocking slow movers ties up cash and space. That tension never really goes away.

Some planners adjust safety stock based on service level targets, but also add a layer of judgment. If a part is critical, they keep extra. If it is rarely used and easy to source, they take more risk.

It is not perfectly scientific. And that is okay.

Where Croston’s actually helps

Croston’s method does not magically fix everything. It will not predict exactly when the next order arrives.

But it does something more useful. It gives you a forecast that respects how the demand really behaves.

Less noise. Fewer misleading averages. A clearer sense of both size and timing.

And when you are dealing with sporadic items, that alone can make planning feel a bit less like guesswork.