Earn Improvement Over a Baseline

Simple references expose fake progress

A dummy baseline ignores useful feature relationships on purpose. DummyClassifier might predict the most frequent class; DummyRegressor might return the training median. A domain baseline can be stronger, such as the rule operators already use. Any candidate should beat both a trivial reference and the relevant existing process on the metric that matters.

from sklearn.dummy import DummyClassifier
baseline = DummyClassifier(strategy='most_frequent')
baseline.fit(X_train, y_train)
print(baseline.score(X_test, y_test))

Suppose 98% of transactions are legitimate. The most-frequent baseline reaches 98% accuracy while catching no fraud. That immediately reveals accuracy as an unsuitable headline measure. Baselines do more than set a low bar: they expose class imbalance, pipeline bugs, and whether added complexity buys meaningful value.

Analogy: Before judging an elaborate weather model, compare it with tomorrow will match today. If sophistication cannot beat that cheap forecast reliably, it has not earned production cost.

Warning: A model that beats a weak baseline but loses to the current human or rule-based workflow is not an improvement.