Baseline before you boost

Steps

1. 1Predict the prior/majority class; record its metric as floor
2. 2Fit one linear/logistic model with sane regularization
3. 3Add one gradient-boosted model only if it beats baseline materially
4. 4Compare on identical splits and preprocessing pipelines
5. 5Quantify lift in business terms, not just AUC deltas
6. 6Deploy the simplest model within epsilon of the best

Common Pitfalls

• ▲Deep nets on tabular data losing to logistic regression
• ▲Metrics improving while business outcomes don't

Commands

$ npx skills add aniruddhaadak80/skills --skill ml-modeling-baseline-first-modeling

$ npx skills add aniruddhaadak80/skills --skill ml-modeling-baseline-first-modeling -g

Tags