When not to use machine learning

A client asked for demand forecasting across 200 SKUs. A four-week moving average beat the first gradient-boosted model by 3% on MAPE and cost nothing to run.

Models earn their maintenance cost when the relationship is genuinely non-linear, the data is plentiful, and the decision repeats often enough to matter. Two out of three is usually not enough.

We now require a baseline in every proposal. If the baseline wins, the project ends early and the client keeps the budget.