A retailer's churn dataset has 8,000 customers, of whom 1,200 churned. A vendor demo reports a model with 86% accuracy on a held-out sample and calls it "highly accurate".
What accuracy does the do-nothing baseline (predict "no churn" for everyone) achieve, and what does that say about the vendor's model?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
Non-churners are 6,800 of 8,000 customers, so always predicting "no churn" scores 6,800 / 8,000 = 85%. The vendor's 86% is one percentage point above doing nothing. Accuracy without a baseline is a number without meaning; on imbalanced data it is usually flattering.
15% is the churn rate. The majority-class baseline predicts the majority class, which is right 85% of the time.
A coin flip is the baseline only when classes are balanced. With 85/15 classes the naive baseline is 85%, not 50%.
The baseline is 85%, not 86%. The model does add something — about one point — which is the honest way to describe it.