The Role of Machine Learning in Modern Demand Planning

Machine learning has been both oversold and undersold in demand planning. Oversold by vendors who imply it replaces the planner, and undersold by planning teams who tried it on poor data, got worse forecasts than their own judgement, and concluded the technique does not work.

## What machine learning is genuinely good at The strongest use cases are narrow and repetitive. Machine learning is excellent at finding patterns across many related series at once, at detecting the effect of price and promotion from history, at handling intermittency and lumpy demand where classical statistical methods struggle, and at scaling a decent forecast across thousands of items where no human could review them individually. It is also good at continuously re-evaluating itself as new actuals arrive.

## What it is not good at It cannot know about the promotion that has not been agreed yet, the competitor launching next month, or the customer who is about to consolidate suppliers. It has no view of intent. It also inherits every flaw in the history it learns from: if past promotions were badly executed, or if stockouts truncated the sales history, the model will faithfully reproduce those distortions unless they are corrected first.

## Where it fits in a planning process The productive arrangement is a layered one. A statistical or machine learning engine produces the baseline at the level where it has enough signal, typically item and location. Planners review the exceptions rather than every line, adding the intelligence the model cannot have: customer plans, distribution gains and losses, new listings, known supply constraints. The result is a forecast that is both scalable and informed, and the planner's time goes where the judgement actually changes the number.

## The data you actually need Before any modelling, three things have to be true. Sales history must be cleaned of stockouts, so that zero sales due to no stock are not read as zero demand. Product and customer hierarchies must be stable enough to aggregate sensibly. And the promotional calendar must be available as structured data rather than as notes in a folder. Businesses routinely skip this and then blame the algorithm.

## Getting started sensibly Pilot on a bounded, well-behaved category, measure against the current forecast rather than against perfection, and hold back a period of history to test on honestly. If the model does not beat the existing process on that test, it will not beat it in production either. Then extend slowly, and keep the planner in the loop throughout.

Explore machine learning and AI in sofco, see how it supports demand planning and collaborative forecasting, or book a demo.