Back to Blog
Data by OrandiMachine LearningScikit-LearnLogistic RegressionBanking Risk AssessmentModel TrainingModel Retraining

First Model, God Model, Bullshit Model: My 3-Month Journey Training and Retraining a Scikit-Learn Logistic Regression for African Banking Risk Assessment

9 min read  · 1,680 wordsBy Orandi Felix

Lesson: ROC-AUC on a train-test split is not a deployment health check. Shadow scoring is.

Lesson: Monitor labels, not just features. If your target definition changes, the model must start from scratch.

If I had stopped at v2 I would have been fired after the label change. If I had stayed on v3 I would have been buried in joblib files. The only winning move is treating the model as software, not a asset.

Share this article: