What is a baseline model?
A baseline model is the simplest model that can produce a prediction for a problem, built first so that every later model has something to beat. For a classification task, the baseline often predicts the most common class for every record. For a forecast, it might repeat last month’s value or a moving average. A rule of thumb the business already uses also counts as a baseline.
The baseline sets the floor. If a gradient-boosted model with two hundred features cannot beat “predict that every invoice gets paid on time,” the project has learned something important before spending more money.
Why a baseline model matters
Model scores mean nothing without a reference point. A collections agency might build a model that predicts which accounts will pay within 60 days and report that it is right nine times in ten. If nine in ten accounts already pay within 60 days, the model has added nothing over the null error rate. The baseline exposes that immediately.
Baselines also protect budgets. A team can build one in an afternoon, which makes it the natural first deliverable of a proof of concept. Comparing every candidate against the same baseline keeps model evaluation honest and makes the gain from each added feature visible.
What to watch for
- Ask any vendor which baseline its reported results were compared against.
- Use the business’s current process as a baseline, because that is what a new model actually replaces.
- Keep the baseline running alongside the production model so that a decline in quality is easy to spot.