Demand Forecasting in Manufacturing: ELM vs. ARIMA — What to Choose for SMEs?

Demand forecasting is one of those problems that seems solved — until the plan diverges from reality. Seasonality, one-off orders, irregular customer rhythms: classic methods based on Excel and an experienced planner's intuition fail precisely when you need them most.

A study published in LogForum (DOI: 10.17270/J.LOG.2022.637) tests two algorithmic approaches under conditions typical of manufacturing SMEs: ARIMA (AutoRegressive Integrated Moving Average) — a statistical standard known for decades — and ELM (Extreme Learning Machine), a type of single-hidden-layer neural network that trains in a fraction of the time required by conventional deep networks.

What is the difference between ARIMA and ELM?

ARIMA is a time series model — it analyses sales history and captures autocorrelation patterns: what follows from what over time. It requires data stationarity (often requiring differencing) and has limited ability to capture sudden structural demand shifts.

ELM is a different class of model. Instead of relying on statistical assumptions, it learns from examples. Input layer weights are randomly assigned and frozen — only the output is optimised analytically (not by gradient descent). This gives:

• very short training time (seconds, not hours),
• ability to include external variables (price lists, campaigns, calendar seasonality),
• better handling of non-linear relationships in data.

Study results: when does ELM win?

The authors compared both approaches using real data from several manufacturing SMEs. The key finding: for regular, seasonal demand, ARIMA and ELM are comparably accurate. However, for irregular demand — spiky, with outliers, with short history — ELM achieves a significantly lower forecast error (measured by MAPE).

For companies with a product portfolio of varying sales character, the practical conclusion is simple: there is no single best model for the entire catalogue. Segment products and match the method to the demand profile.

What does this mean for the production planner?

The goal is not to replace the planner with an algorithm. The goal is to give the planner a better starting point than "last year times 1.05".

The algorithm provides a baseline forecast. The planner adds context: we know customer X is losing a major buyer, so we lower the forecast by 20%. We know a distributor campaign is running — we raise it for 6 weeks. This hybrid process — algorithm plus expert corrections — is what the most effective manufacturing companies practice.

Three prerequisites before ELM is worth implementing

1. Sales history in one place. ELM needs data. Minimum 18–24 months of history at SKU or product group level, in a format suitable for automated processing (not a pivot table with manual corrections).
2. A defined planning cycle. The algorithmic forecast must be updated regularly and must enter the process — not be produced once and forgotten.
3. Someone accountable for the result. The most common cause of failure in "AI in planning" projects is not the algorithm — it is the absence of a process owner who takes ownership of the forecast and bears responsibility for purchasing and production decisions.

A practical first step: data audit

Before investing in a tool (whether an ERP with an ML module or separate software), run a simple test: can your team pull two years of sales history broken down by SKU, with dates and quantities, within one hour? If not — that is the right starting point, not an algorithm.

Most companies we work with have the data — scattered across spreadsheets, systems, and planners' heads. Consolidating data and standardising the planning process delivers more value than the best algorithm applied to poor data.