Problem Framing

  • Start from a football decision rather than a convenient dataset.
  • Write down what the model is supposed to help with before any code is written.

Experiment Design

  • Build a simple baseline first so later improvements mean something.
  • Choose train and test windows that resemble future use, not just whatever gives the biggest sample.
  • Check early for leakage, unstable samples, and data fields that look informative but would not be available in real use.

Diagnostics And Uncertainty

  • Use measures that match the real task, especially when the output is a probability rather than a hard classification.
  • Check calibration, subgroup behaviour, and where the model is most uncertain.

Iteration Discipline

  • Keep track of rejected ideas, broken assumptions, and failed first passes.
  • Do not present a model as better simply because it is more complex.