Latest / MBA Training AI / SR 11-7: why gradient boosting models fail validation

Transcript

SR 11-7: why gradient boosting models fail validation · MBA Training AI

• 0:00
  Host: Leaders' insights. Five minutes on SR11, seven still bites, and gradient boosting just
• 0:10
  Host: made the wound worse. If a bank can't explain why its model denied someone a mortgage,
• 0:16
  Host: should regulators just shut the model down?
• 0:18
  Expert: That's exactly the fight happening right now. SR11-7. The Federal Reserve's model
• 0:24
  Expert: risk guidance from 2011 doesn't literally say shut it down. But examiners in 2026 are
• 0:30
  Expert: treating unexplainable models as unvalidated, and an unvalidated model can't go into production.
• 0:37
  Expert: So functionally, yes.
• 0:38
  Host: 15-year-old guidance. Why is it suddenly biting harder this year?
• 0:43
  Expert: Because the models changed and the rules didn't. SR11-7 was written when everyone scored
• 0:49
  Expert: credit with logistic regression, a simple statistical method where you can read off exactly
• 0:54
  Expert: how much each factor moves the outcome. Now banks are shipping gradient boosting, which
• 0:59
  Expert: stacks hundreds of little decision trees on top of each other, and the interpretability
• 1:04
  Expert: just evaporates.
• 1:06
  Host: Give me the moment this became a headline.
• 1:08
  Expert: The trigger was examiners flagging Foundation model-powered scoring, big pre-trained AI
• 1:13
  Expert: exams repurposed for credit decisions in supervisory exams earlier this year. Google's
• 1:19
  Expert: AI team has claimed their explainability tooling can attribute a score down to individual
• 1:24
  Expert: features with high fidelity, but they sell that tooling, so I'd cross-check that against
• 1:29
  Expert: a validation team that doesn't have a product to move.
• 1:33
  Host: Let me throw three things people believe at you. First, SHAP values solve the explainability
• 1:37
  Host: problem. True, half true, or wrong?
• 1:40
  Expert: Half true, and the dangerous half. SHAP, a method that assigns each input a contribution
• 1:46
  Expert: to the final score, gives you a number for every feature. The problem is, those numbers
• 1:51
  Expert: describe correlation inside the model, not the causal story a regulator wants. You can
• 1:57
  Expert: hand an examiner a beautiful SHAP chart, and they'll ask, so would approving this
• 2:02
  Expert: person have been the right call? And SHAP has nothing to say about that.
• 2:06
  Host: So it explains the math, not the decision.
• 2:09
  Expert: Right, it tells you what the model did, not whether what it did was defensible. Under SR11,
• 2:15
  Expert: 7, that gap is where validations die.
• 2:17
  Host: Second belief, gradient boosting is inherently more accurate, so the explainability tradeoff
• 2:23
  Host: is worth it. Where does that land?
• 2:25
  Expert: Mostly wrong for credit, actually. On tabular data, rows and columns, income,
• 2:31
  Expert: balances, payment history, gradient boosting often beats logistic regression by a few
• 2:37
  Expert: points of predictive lift. But a few points on a portfolio where you already have decades
• 2:42
  Expert: of clean data rarely justifies losing the ability to answer a fair lending complaint.
• 2:48
  Expert: OpenAI has published numbers suggesting large models close reasoning gaps fast. Though again,
• 2:55
  Expert: they're selling the models. So treat the benchmark like marketing until someone
• 2:59
  Expert: independent reruns it. You're saying the accuracy gain is real but small,
• 3:03
  Host: and the governance cost is large.
• 3:05
  Expert: And the governance cost is fixed, not one time. Every model refresh, every retrain,
• 3:11
  Expert: you revalidate. A logistic model you validate once and barely touch.
• 3:16
  Expert: A boosted model drifts, so you're paying that tax quarterly.
• 3:19
  Host: Third, banks can just bolt on a simpler explainer model next to the complex one
• 3:25
  Host: and satisfy the regulator.
• 3:26
  Expert: Wrong. And examiners caught on to this one specifically. The trick is called a surrogate.
• 3:32
  Expert: You train a simple, readable model to mimic the complex one's outputs,
• 3:36
  Expert: then show the regulator the readable version. The issue is the surrogate can agree with the
• 3:42
  Expert: real model 95% of the time and disagree exactly on the edge cases, the denials,
• 3:48
  Expert: where the legal risk lives. You've built an explanation that's most wrong precisely
• 3:53
  Host: when it matters. So the honest answer to my opening question?
• 3:56
  Expert: If you can't reconstruct the reason for an individual adverse decision,
• 4:00
  Expert: not the average, the individual, you shouldn't have that model in front of consumers.
• 4:05
  Expert: That's not a regulator being difficult. That's the actual law behind adverse action notices,
• 4:11
  Expert: the letter you're owed explaining why you got denied.
• 4:14
  Host: Leave the risk officer listening with one thing to do, Mundy.
• 4:17
  Expert: Pull ten recent denials from your most complex model and try to write the
• 4:21
  Expert: adverse action reason for each by hand from the model's own outputs.
• 4:26
  Expert: If you can't do it for even one, that model isn't validated no matter what your documentation says.
• 4:32
  Expert: And it's better you find that out before the examiner does.
• 4:35
  Host: What we read for this one, TechCrunch AI, ours Technica AI, the decoder, KD Nuggets, Google AI,
• 4:43
  Host: vendor, AI Lab, OpenAI, vendor, AI Lab. Done for today.
• 4:49
  Host: There's a new AI piece every morning at MBA-training.com.