Baseline

Definition

A baseline is a simple method, such as predicting the last value or the average, against which a new method is compared. A result means something only relative to it: a model that cannot beat a sensible naive baseline has added nothing.

In plain terms

“90% accurate” says little until you know what doing nothing smart would score: if 92% of items belong to one class, always guessing that class beats the model. In forecasting the classic baselines are persistence (“tomorrow equals today”) and seasonal naive (“tomorrow equals the same weekday last week”).

A fair comparison uses strong simple baselines, not weak ones chosen to make progress look large. A well-known paper on classifier research observed that simple methods often capture most of the achievable gain, so apparent progress over weak baselines can be an illusion.

Example

Forecasting daily demand; error is the mean absolute error (lower is better).

Against the first baseline the complex model looks 3% better. Against the stronger seasonal baseline it is 29% worse, and the extra complexity is not justified.

Common misunderstandings

• “90% accuracy is good.” It depends on the baseline; with a 92% majority class it is worse than guessing.
• “A weak baseline shows the model’s value.” It only shows that a weak baseline is weak; use the best simple method you can think of.
• “One baseline is enough.” Check several: last value, average, seasonal repeat, and a plain rule.