Noblle

Intelligence

Forecasting demand with limited history

Approaches that work when order history is short, lumpy and concentrated in a few customers, as it is in most SME factories.

By Noblle · · 7 min read

Short answer

When order history is short and lumpy, forecast at the level where there is a signal, usually by customer or product family rather than by individual part, combine simple statistical baselines with what sales knows about upcoming work, and report a range instead of a single number. Accuracy improves as real orders accumulate.

Key takeaways

  • Do not forecast every part. Aggregate to the level where patterns appear, then allocate down.
  • In customer-concentrated businesses, one customer's plans can matter more than any model.
  • Simple methods with sensible baselines often beat complex ones on small data.
  • A range with a stated confidence is more useful, and more honest, than a single number.

Why small-factory data is difficult

Forecasting textbooks assume years of steady demand across many customers. A typical SME manufacturer has a few years of history at best, with a handful of customers accounting for most of the volume, and orders that arrive in lumps tied to projects, tenders or the customer's own cycle.

A model fitted to each part's history will mostly fit noise. The result is a forecast that looks precise and is wrong, which damages trust in the whole exercise.

Forecast at the right level

Demand for a specific part code in a specific month is rarely predictable. Demand for a product family, a material, or a customer's total orders over a quarter often is.

  • Forecast by customer or product family first, where there is enough volume to see a pattern.
  • Translate to materials and capacity, which is what you actually need to plan, rather than to every finished part.
  • Allocate down to individual parts only where a part has consistent, repeated demand.

Combine statistics with what you know

The best signal in a concentrated business often sits in someone's head: a customer mentioned a big project, a tender is due, or a seasonal peak is approaching. A good process captures that knowledge explicitly instead of ignoring it.

  • Start with a simple baseline, such as a moving average or a seasonal pattern if you have at least a couple of cycles.
  • Add known events: confirmed pipeline, tenders, planned shutdowns and price changes.
  • Weight pipeline by realistic win rates drawn from your own quote history.
  • Review forecast versus actual every month and adjust. The error tells you where to trust the model.

Use ranges and decide what to do with them

A forecast of 120 units with a likely range of 80 to 160 supports better decisions than a bare 120. Use the low end to commit to long-lead materials only where you are confident, and the high end to decide where to hold flexible capacity or a supplier option.

The aim is not perfect prediction. It is to make fewer panic purchases, avoid obvious stockouts, and give planning an informed starting point that gets sharper as order history grows.

Frequently asked questions

How much history do I need to forecast demand?
Seasonal patterns need at least two full cycles. With less, rely on simple baselines, aggregation and known pipeline, and treat the output as a range.
What is the best forecasting method for small datasets?
Simple methods, such as moving averages, seasonal naive and aggregated forecasts, often outperform complex models when data is short or lumpy.
Should I forecast every part number?
Usually not. Forecast at customer or family level where signal exists, then derive material and capacity needs, and only forecast individual parts that have steady repeat demand.

Related in the platform

  • ML Forecaster: Local, factory-specific demand forecasting.
  • Supply Chain: Supplier, procurement and inbound visibility in one place.
  • MRP: Materials, scheduling and capacity that reflect the real floor.

See how this works in your factory

We configure the walkthrough to your own process before the call.