Recognising the wrong tool.
WHAT A RULE SOLVES BETTER
Anything with clear, stable, enumerable conditions.
WHY
It is explainable, testable, instantly correctable and free to run.
WHAT TO TRY FIRST
The rule.
WHAT MAKES MACHINE LEARNING UNSUITABLE
No historical data Data that will not resemble future data Requirements for exact correctness Legal or regulatory need for full explanation Too few examples of what you want to predict
WHAT TOO FEW MEANS
It depends on the task, but tens of examples rarely suffice for anything.
WHAT MAKES A PROJECT FAIL LATER
No route to production Nobody to maintain it Predictions nobody acts on Performance that decays with no retraining plan
WHAT TO ASK BEFORE STARTING
What decision changes because of this prediction?
WHY THAT QUESTION
If no decision changes, the model has no value regardless of accuracy.
WHAT TO ASK SECOND
What is the cost of being wrong, in each direction?
WHY
It determines the threshold, the metric, and whether the project is worth doing.
WHAT TO PROPOSE INSTEAD, FREQUENTLY
Better reporting, or a rule.
