Causal Inference in Lending

Every model in this track so far answers one kind of question: given what I observe, what will happen? But almost every decision a lender actually makes is a different kind of question: if I change something — the rate, the limit, the cutoff, the collections script — what will happen then? The first is prediction; the second is intervention. They look similar, they use the same data, and they have different answers. The final module of this track is about the gap between them, because that gap is where models that validated beautifully produce policies that fail.

The parable: does the interest rate cause default?

Pull your portfolio data and run the obvious analysis: borrowers paying higher rates default more. Massively more. Does raising rates cause defaults — should you cut pricing to cut losses?

You already know what’s wrong: you set those rates, based on risk. Risky borrowers got high rates and defaulted more — risk is a confounder, a common cause of both the “treatment” (rate) and the outcome (default). The observed association mixes the causal effect of price (real: higher payments do strain borrowers — and adverse selection means price changes shift who accepts) with the risk signal flowing through your own pricing policy. Decompose it yourself:

The experiment to run first: drag the true causal effect to zero. The naive comparison still reports high-rate borrowers defaulting at multiples of the low-rate group — a strong, stable, completely non-causal association, manufactured by your own pricing. Any model consuming this data will learn it; any decision-maker reading the model causally will act on it; and the action will not produce the predicted result. Then set risk-based pricing to zero and notice the naive estimate snap to the truth — that is what randomization buys, and why it’s the gold standard: it severs the arrow from risk to treatment by force.

Why lending data is a confounding machine

The rate example isn’t an unlucky case — it’s the structure of the domain. Credit data is generated by the lender’s own decisions, so the treatment you want to evaluate is almost never assigned independently of risk:

• Selection into the book: you only observe outcomes for approved applicants — reject inference is a causal-inference problem that the industry named before the field existed. Every model trained on booked loans inherits this (Module 6’s feedback loop).
• Treatment follows risk everywhere: limits, pricing, collections intensity, restructuring offers — all assigned by risk. The sicker patients get more medicine; medicine then “correlates” with death.
• Policy defines the data: change the cutoff and you change which relationships are observable at all. A model estimated under one policy describes that policy’s world — use it to evaluate a different policy and you’re extrapolating into data you deliberately never collected. This is why “the model says approving segment X is safe” can be false precisely because the old policy never approved segment X.

The toolkit, in order of preference

1. Randomize — you’re closer than you think. The gold standard is cheaper in lending than almost any industry, because decisions are centralized, digital, and repeated at scale. Randomized pricing tests, randomized limit increases, randomized collections treatments — and the two randomized designs this track has already endorsed without the vocabulary: champion/challenger is an A/B test of decision policies, and the “approve a random sliver of rejects” fix for reject inference is a randomized experiment purchasing unbiased data with a known default budget. The classic of the genre — Karlan and Zinman’s field experiments randomizing offer rates — found real but modest causal price effects, far smaller than naive portfolio correlations suggest. Exactly the widget’s lesson.

2. Adjust for confounders — with humility. When you can’t randomize, control: regression with risk controls, stratification, matching, or propensity scores (model the probability of receiving treatment given observables, then compare treated and untreated borrowers at the same propensity — collapsing many confounders into one balancing score). The widget’s “adjusted” bar recovers the truth this way — but only because its confounder is fully observed. The honest caveat every adjusted estimate carries: adjustment handles the confounders you measured. Unobserved confounding — the underwriter’s judgment that never hit a field, the borrower’s job insecurity — stays in your estimate, unmarked. Sensitivity analysis (“how strong would a hidden confounder need to be to kill this result?”) is the professional habit.

3. Hunt natural experiments. Sometimes the world randomizes for you, and lending is rich in these: cutoff discontinuities — applicants at score 649 vs 651 are statistically identical people who received different decisions, so comparing outcomes just around the threshold estimates the policy’s causal effect (regression discontinuity — your own historical cutoffs are a lab you already paid for); policy changes — a limit-increase program rolled out in region A but not B enables difference-in-differences, comparing the change in A against the change in B so that stable differences cancel; instrumental variables — some assignment quirk (randomized underwriter assignment, where underwriters differ in leniency) that shifts treatment but plausibly doesn’t otherwise affect outcomes.

The classic application that brings it together is the credit line increase problem: does raising a limit cause more default, or do risky borrowers just get flagged by the increase-request? Observationally hopeless (requests signal stress); solved properly by randomized increases or discontinuities in the increase-eligibility score. The notebook implements exactly this, all three ways, on data where you know the truth.

Prediction and causation, correctly divided

The resolution isn’t that causal methods replace predictive ones — it’s a division of labor:

A predictive model is a superb targeting device inside a causally evaluated policy: use the experiment to learn what the treatment does, use the model to decide who gets it. The failure mode is skipping the first step.

The track, closed

Eight modules: the traps in the raw data, the features worth building, the dynamics of financial series, honest metrics, honest validation, staying honest in production, reasoning under scarce data, and — today — the difference between predicting the world and changing it. The arc has a single spine: financial data is generated by processes that fight your assumptions — markets that price away signal, policies that select the sample, cycles that break stationarity, and decisions that entangle treatment with risk. Data science in finance is the discipline of modeling anyway, with the failure modes named in advance. Paired with the credit-models track, you now have both halves: the economics of what the numbers mean, and the statistics of how to compute them without fooling yourself.

Where this connects

• Reject inference (credit Module 2) and the production feedback loop (Module 6) are both selection problems — this module is their general theory.
• The “model trained under one policy breaks under another” warning is the microeconomic Lucas critique, and it’s the deep reason Module 5’s backtests can’t validate a policy change — only an experiment can.
• Attribution-vs-causation is the governance edge of credit Module 12; fair-lending analysis increasingly demands the causal version of the question, not the correlational one.

Try it yourself

The notebook below builds the whole argument on simulated lending data where the truth is known: the rate-confounding parable with naive, regression-adjusted, and propensity-matched estimates; an unobserved-confounder sensitivity sweep showing how each estimator degrades; a regression-discontinuity analysis at a score cutoff (your free natural experiment); and a difference-in-differences evaluation of a staged limit-increase rollout — plus the randomized version of each, as the benchmark the observational methods are trying to earn. The closing exercise is the credit-line-increase memo: given all four analyses, write the recommendation, with the uncertainty you actually have.

Get new posts by email

One email per new article. No spam, no upsells, unsubscribe anytime.