Machine Learning for Capacity Planning

The single-number forecast problem

A capacity plan usually starts with a forecast: we will need 14,000 units a month next year, so we need two more machines and five operators. Everyone in the room knows the 14,000 is uncertain. The plan treats it as fact anyway, because a single number is easy to put in a spreadsheet.

Then demand comes in at 11,000 and you have idle machines, or at 17,000 and you are paying weekend overtime and turning work away. The forecast was not necessarily bad. The plan simply ignored how wide the uncertainty was.

What machine learning changes

Machine learning is not magic at predicting next year. What modern forecasting methods do well is quantify uncertainty honestly, which is exactly what capacity planning needs.

• Quantile forecasts that give the 10th, 50th and 90th percentile of demand per month
• Scenario simulation that generates thousands of plausible demand paths, including runs of high months
• Driver-based forecasts that link demand to things you know or can assume, such as sales pipeline, contracted volumes, pricing or market growth
• Workload conversion, where a model learns how much capacity a unit of demand really consumes, including mix effects

That last one is underrated. Ten thousand orders of mostly simple products need very different capacity from ten thousand complex ones. A model trained on actual hours and machine time per order type makes the conversion far more reliable than a single standard rate. If you want the broader picture of forecasting applications, see AI for forecasting and planning.

Choosing capacity by the risk you accept

With a range, capacity becomes a business choice about risk rather than an argument about the forecast.

The right point depends on the relative cost of idle capacity against the cost of a shortfall. If agency staff or subcontracting cover peaks cheaply, plan lean. If a missed delivery loses a major customer, plan higher. The model provides the probabilities; the business chooses.

A worked example: a growing lab

Take a testing laboratory processing around 3,000 samples a week, with contracts from a few large clients plus steady walk-in work. Adding an analyser takes four months from order to validated use, and hiring a qualified analyst takes three.

A quantile forecast built from contract volumes, historical walk-in demand and client pipeline shows the median need fits current capacity until spring. The 80th percentile crosses capacity in February, driven mainly by one client's likely contract expansion. That changes the conversation. Instead of debating whether to buy the analyser, the lab can order it on a timeline that covers the 80th percentile, and arrange overflow with a partner lab for anything above.

The numbers are illustrative. The shape of the decision is the point: capacity, lead time and risk on one page.

Capacity planning with ML versus simpler methods

For service businesses planning staff across a few teams, the approach in capacity planning for service businesses is often enough, and we would not add machine learning for its own sake. The simpler method works when demand is stable and you can flex capacity quickly.

• Use machine learning when capacity lead times are long and expensive
• When demand depends on several drivers that interact
• When product or service mix shifts and changes the workload per unit
• When you have several sites or resources that can share load

Skip it when you have less than a couple of years of history, when a few big contract decisions dominate demand and no model can predict them, or when capacity can be flexed within days.

Where these plans go wrong

1. Treating the median as the plan and filing the range in an appendix
2. Forgetting that ranges widen further out, so year-two numbers deserve less confidence
3. Ignoring correlation: when demand is high for one product, it is often high for others too
4. Missing constraints elsewhere, such as a packing area that caps output regardless of machine count
5. Never checking the forecast ranges against what happened, so nobody knows if they are calibrated

Calibration checks are simple: if you plan on 80th percentile forecasts, actual demand should exceed them roughly one period in five. If it exceeds them one in two, the ranges are too narrow and the plan is riskier than it looks.

How SpiderHunts would build it

At SpiderHunts we tie the model to a specific capacity decision rather than building a general forecasting platform. Which resource, what lead time, what does a shortfall cost? That keeps the work proportionate.

The deliverable is usually a quantile forecast refreshed monthly, a workload conversion model, and a planning view that shows capacity against the demand range with the lead time marked. It connects to existing ERP and planning data and is built by our data science team. Where the decision is strategic and one-off, a well-documented analysis may be all you need, and we will tell you that up front.

Frequently asked questions