Baseline

A baseline is a deliberately simple solution built first, so you know what any complex model must beat to justify itself.

Before renovating a shop, a sensible owner asks: what does the shop earn as it is? Without that number, "the renovation brought in ₹2 lakh a month" is an empty claim — maybe the shop earned that anyway. A baseline is the "as it is" number for prediction problems.

Baselines come in grades, and each exposes a different illusion:

predict the most common class        exposes what accuracy is worth on imbalanced data
predict yesterday's value            embarrassingly strong for time series
a couple of hand-written rules       captures the obvious signal
logistic or linear regression        the honest classical reference

The classic humbling: your neural network hits 94% accuracy on churn prediction. Then you notice 93% of customers never churn, so "predict nobody churns" scores 93%. The network has learned almost nothing — a fact invisible without the baseline.

Skipping this step is the most common process mistake in applied ML. A day spent on a logistic-regression baseline tells you how hard the problem is, whether your features carry signal, and whether the deep model is worth its serving costs. Related but different: a benchmark is a shared public test for comparing systems, while a baseline is your project's private floor.

Where to go next#

• Full lesson: Model evaluation
• Related terms: benchmark, accuracy, linear-regression, logistic-regression