regression
Predict, or explain?
Here is a single dataset and two irreconcilable mandates handed to you at once. The bank wants the most accurate possible predictions of who will default — it does not care why, only that the ranking is right. The regulator wants a defensible account of which factors actually cause default — accuracy matters less than the causal story holding. These are not the same task, and the model that wins one can lose the other: a black box may predict beautifully while explaining nothing; a clean causal model may forecast worse. Show precisely why the best answer differs by mandate, and refuse to pretend one model serves both. Get this wrong and someone acts on a prediction as if it were a cause — intervening on exactly the wrong lever.
Who this problem belongs to
The two figures whose methods fit it best, out of 58 in contention.
This problem is Breiman's 2001 paper 'Statistical Modeling: The Two Cultures' rendered as a scenario. He lived both mandates: thirteen years as a freelance consultant building predictive models for clients who, like the bank, cared only that the ranking was right, then a Berkeley professorship among colleagues who wanted data-generating stories. CART (1984, with Friedman, Olshen, and Stone), bagging (1996), and random forests (2001) are the bank's toolkit, and his variable-importance measures are an honest admission that black boxes explain poorly. Crucially, his paper argues exactly what the problem demands: the algorithmic and data-modeling cultures answer different questions, and forcing one model to serve both corrupts each. He would refuse the pretense on principle, with three decades of consulting scars as evidence.
Pearl owns the regulator's half of the mandate outright. His Bayesian networks (1980s) began as prediction machinery, but by 'Causality' (2000) he had built the do-calculus precisely to formalize the distinction this problem turns on: P(default | income) is not P(default | do(income)), and acting on the first as if it were the second is intervening on the wrong lever — his phrase almost verbatim. He can state, in graph-theoretic terms, exactly which covariate adjustments license a causal reading of the bank's data and which do not. His limitation is symmetric: he offers little for maximizing predictive ranking accuracy, and his framework requires a defensible causal graph the bank's messy observational file may not support. But no one shows 'why the mandates differ' more precisely.
Fought here
In the mind map
The same ideas, as concepts rather than history — in John's ML knowledge map.
58 figures are scored on this problem. Draw it in a battle to see where you land.