Before investing in a sophisticated prediction model, ask what the simplest possible answer would achieve. A constant baseline replaces learned relationships with a fixed prediction. It gives a selected metric a starting value and tests whether a more complex model earns its development and operating cost. section 8.2 section 8.2.1 section 8.5
Choose a primitive prediction that fits the task:
- For regression, predict an average, median, last available value, or chosen constant. section 8.2
- For classification, predict the majority class or a fixed probability for the positive class. section 8.2
- For ranking, start with an equally simple reference, such as random order or a basic heuristic; it need not be a literal constant. section 8.2
A taxi arrival estimate
Suppose an arrival-time service must avoid both discouraging customers with overestimates and disappointing them with underestimates. “Always predict five minutes” is a constant baseline. Returning a fixed estimate for each district is a step toward a rule-based baseline. In this case, such simple estimates were hard to beat, and a district-based rule served as a production fallback. section 8.3
Keep the baseline after moving on: it is a sanity check for new models and a fallback when prediction fails, exceeds a response-time limit, or lacks history for a new user or item. Compare each proposed upgrade against it on metric improvement, implementation effort, interpretability, and computation time. A stepwise-constant prediction even won a factory time-series competition, so do not assume complexity will win. section 8.2.1 section 8.5
