baseline
/ BAYS-line /
A baseline is a simple, no-frills reference point you build first, so you can tell whether your fancy model is actually worth anything. It answers a question every project must ask before celebrating: compared to what? A weather forecaster who claims to predict rain correctly 70% of the time sounds impressive — until you learn that simply guessing "no rain every day" would have been right 75% of the time in that dry climate. That dumb guess is the baseline, and it just beat the fancy model.
Baselines come in degrees of simplicity. The simplest just predicts the most common answer every time, or the average value, ignoring all the features entirely. A slightly richer one uses a single obvious rule or a basic, well-understood method. Either way, the point is the same: a baseline is cheap to build, easy to understand, and gives you an honest yardstick. A new model that can't clearly beat the baseline is, for all its sophistication, not doing its job.
Baselines guard against a very human failure: being dazzled by a number with nothing to compare it to. "95% accurate" means nothing on its own — if 95% of emails aren't spam, a model that blindly labels everything "not spam" also scores 95% while catching zero spam. Establishing the baseline first keeps you honest, focuses effort on problems where the simple approach truly falls short, and often reveals that the simple approach was good enough all along.
Before building a churn predictor, an analyst notes that 88% of customers stay. The trivial baseline — "predict everyone stays" — is therefore 88% accurate. Now any real model must clear 88% to have earned its place; a model scoring 86% is worse than doing nothing.
Always ask "compared to what?" — a score is meaningless without a baseline.
Build the baseline before the complex model, not after. It is tempting to skip this and admire a big number, but a baseline is the cheapest insurance against fooling yourself — and sometimes it quietly wins.