SEFR enables fast SQL-only classification in BigQuery

Hot take: this is less about inventing a better classifier and more about matching the math to the warehouse engine.

• –The real win is operational: one query, no persisted model object, and no separate ML pipeline to manage.
• –SEFR's closed-form training maps cleanly to aggregations and joins, so BigQuery can spread the work across many slots.
• –Logistic regression still has the better ranking quality, so SEFR reads best as a fast baseline or ELT-native option, not a universal replacement.
• –It is strongest where auditability and simplicity matter more than squeezing out the last few AUC points, and weakest on nonlinear or multiclass problems.