We measure our models against the dumbest possible baseline

Before we train anything, we build the dumbest model that could possibly work, and we make the real model beat it. The baseline is deliberately trivial: predict the majority class every time, or carry yesterday's value forward, or apply a three-line rule a junior analyst could write on a whiteboard. It costs an afternoon. Its job is not to be good - its job is to be the floor that everything else has to clear before it earns a place in production.

The reason we insist on it is that the floor is often higher than people expect. On an imbalanced classification problem where 92% of cases are one class, a model reporting 91% accuracy sounds respectable until you notice that always guessing the majority class scores 92% and needs no GPU, no training run, and no monitoring. We have watched genuinely sophisticated models - gradient-boosted ensembles, fine-tuned networks - fail to beat a heuristic on the first pass more often than is comfortable to admit. Without the baseline in front of you, that failure is invisible; the metric looks fine in isolation, and you ship something that adds cost and risk for no lift.

The baseline also answers a question that comes before model selection: does this problem need machine learning at all. If a simple rule gets you 95% of the value, the honest engineering answer may be to ship the rule, write it down, and spend the model budget elsewhere. ML carries a standing cost - data pipelines, drift, retraining, the monitoring to catch it when it quietly degrades - and that cost is only worth paying when the lift over a plain rule is real and large enough to matter to the business.

When a model does clear the baseline, the baseline tells us by how much, and that number sets expectations honestly. Two points of lift and twenty points of lift are different products with different operational stories, and the client should hear which one they are buying before launch, not after. The gap is the actual value of the model, stated in the same units the business already understands.

So our rule is narrow and we hold to it: ship only the part that beats the baseline. Sometimes that is the full model. Sometimes it is the model on one segment where it clearly wins and the rule everywhere else. And sometimes it is the rule alone, which is a perfectly good outcome - we would rather hand you three lines you can read and run than a model that performs no better and asks for a maintenance budget on top.