Baseline Model

A baseline model is the simplest thing that could work, built first and kept as the reference every later model has to beat. It is usually not a model in any interesting sense: always predict the majority class, always predict the previous value, a dozen keyword rules, or a plain search over the same data. The point of building it is that a score has no meaning on its own. Ninety-four percent accuracy sounds like a result until the baseline that always answers no scores ninety-three, at which point the real contribution of the model is one point, and the interesting question becomes whether that point is worth the cost of running, monitoring and explaining it. Baselines also make the cheaper decision visible. In a surprising number of tasks a rule reaches most of the achievable quality, and the remaining gap is closed by something a small team cannot maintain. Knowing that early changes what gets built, and it is much harder to discover after months have gone into a system that nobody wants to abandon. The baseline is equally useful as an instrument. A model that fails to beat it usually indicates a problem with the data or the framing rather than a weak model, and that failure is a faster diagnosis than tuning. Kept in the evaluation harness and scored on every run, the baseline also detects silent breakage: when both the model and the baseline drop together, the data pipeline moved, and when only the baseline drops, the evaluation set changed. Record the baseline alongside every reported result, because a number without its reference point cannot be judged by anybody reading it later.