Causal inference · Uplift modeling
Causal Credit Uplift
Risk scoring asks who will default. Uplift asks something else: whose outcome actually changes because you did something, like raising a credit limit. Phase 1 checks the methods on a real email RCT. Phase 2 checks whether the same models can recover known true effects on synthetic credit data.
The four-quadrant frame
Normal classification metrics mash these groups into one ranking. Uplift tries to keep them separate, especially Sleeping Dogs, where treatment can make things worse.
Persuadables
CATE > 0
Only get better because you treated them. These are the people you want.
Sure Things
CATE ≈ 0
Do well either way. Risk models love them. Uplift mostly ignores them.
Lost Causes
CATE ≈ 0
Do poorly either way. Low outcome score, but treatment still barely helps.
Sleeping Dogs
CATE < 0
Treatment makes things worse. Bad people to put on a treated list.
Phase 1
Hillstrom RCT
Real randomized email data. Get Qini / T-learner / causal forest working before moving to synthetic credit.
Teaser: T-learner Qini 7.74
Phase 2
Synthetic credit RCT
Known true CATE. See if the models find Persuadables and keep Sleeping Dogs out of the top decile.
Teaser: CF PEHE 0.0015 · top-decile Sleeping Dogs 0.00%
Why not AUC
Wrong Q
It ranks who looks likely to succeed, not who changes because of treatment.
Ground-truth recovery
0.927 corr
CausalForestDML closely recovered the known treatment effects in Phase 2.
Sleeping Dogs targeted
0.00%
CausalForestDML kept true negative-effect customers out of its top decile.