A startup pitches a model that predicts store demand with 85% accuracy. Impressive, unless “order the same as last week” already gets 83%. The baseline is that boring comparison point, and it answers the only question that matters to a buyer: how much better is the new thing than what you’d do anyway?
Typical baselines are deliberately simple. For classification, always predict the most common class (scikit-learn’s “dummy” estimators do exactly this). For forecasting, repeat the last value or last season’s value. For a new deep model, a linear or logistic regression. For an LLM feature, the current human process, a keyword search, or last quarter’s prompt and model. A model that can’t clearly beat a baseline is either underfitting, being measured wrongly, or not worth the cost.
Baselines also keep evaluation honest over time. Teams record a baseline score before a change, then compare: did the new prompt, model or retrieval setup move the number, or is the difference noise? On imbalanced datasets the “always say no” baseline already scores very high accuracy, which is a quick way to show why accuracy is the wrong metric there. “Baseline” has many everyday meanings, so this entry only autolinks the phrase “baseline model.”
Sources
- Google Machine Learning Glossary: baseline — Defines a baseline as a model used as a reference point for comparing how well another, typically more complex, model performs, such as logistic regression as a baseline for a deep model; checked October 24, 2026.
- scikit-learn: Dummy estimators — Documents DummyClassifier and DummyRegressor, simple strategies such as predicting the most frequent class or the mean, for use as baselines; checked October 24, 2026.