What Would Have Happened...?

Risk decisions are about specific exposures: "Did the storm cause this loss, or was the structure already compromised?" "Would this claim have been filed without the policy change?" "Was the underwriting decision defensible for this applicant?"

These are counterfactual questions — they require Rung 3 on Pearl's Ladder of Causation. Rung 2 gets you closer: causal models that predict what happens on average when you intervene. But averages don't defend individual decisions. And most risk teams aren't even at Rung 2 — they're at Rung 1, where the problems are more fundamental.

Rung 1: The Correlation Trap

The trap is wrong answers (confusing correlation with causation).

Most risk models — GLMs, gradient-boosted trees, even deep learning — operate on Rung 1. They find patterns in historical data and project them forward. This works until it doesn't:

Spurious pricing. The model sees that coastal properties have higher claims and raises premiums accordingly. But it can't distinguish storm surge damage from aging infrastructure that happens to cluster near coasts. It optimizes for the pattern, not the mechanism — and misprices the entire book.

Confounded attribution. Claims involving attorneys cost more. Rung 1 concludes: attorneys increase severity. But the causal arrow may run the other way — high-severity claims attract attorneys. Without causal structure, the model conflates consequence with cause, and the reserve strategy follows the wrong signal.

Simpson's paradox⁴ in development. Aggregate data shows that a book's reserves are developing favorably. But segment by litigation status and the trend reverses in every subgroup. Rung 1 sees the aggregate. Rung 2 would see the subgroups. Rung 3 would tell you what this claim would have developed to without litigation.

The common thread: more data makes these problems worse, not better. A larger dataset reinforces the spurious pattern. The only escape is causal structure — and that requires climbing the ladder.

⁴ Simpson's paradox occurs when a trend that appears in aggregate data reverses when the data is segmented by a confounding variable. It's one of the clearest demonstrations of why correlation without causal structure misleads.

Rung 2: Right Answer, Wrong Person

The trap is right answers to the wrong question (population averages applied to individual decisions).

Rung 2 fixes the causal confusion — it correctly estimates the effect of interventions. But it answers for the average policyholder, not the one sitting in front of you:

Mitigation that doesn't apply. A causal model shows that roof reinforcement reduces hurricane claims by 30% on average. The insurer offers a premium discount. But for this property — built on fill soil with a cracked foundation — the roof was never the vulnerability. The discount is wasted; the risk is unchanged.

Reserves that mask outliers. Average development factors say this book will settle within projections. But this claim — with its unique combination of jurisdiction, injury type, and legal representation — is not average. The population estimate absorbs the outlier. The individual reserve is wrong.

Pricing that ignores context. The average effect of a rating variable is statistically significant and directionally correct. But for this applicant — with their specific combination of exposure characteristics — the variable's effect is negligible or reversed. The price is defensible on average, but indefensible for this individual.

Rung 2 gets the mechanism right but treats everyone the same. Rung 3 treats everyone as themselves.

Rung 3: Right Answer, Right Person

The breakthrough is the right answer for the right person (individual counterfactual reasoning).

Rung 3 uses abduction to infer each individual's unique characteristics — the exogenous variables (U) — and then reasons² about what would have happened to this person in particular, specifically:

Claim-level attribution. Not "storms increase losses on average" but "this storm caused $47,000 of this $62,000 claim — the remaining $15,000 was pre-existing deterioration." The adjuster, the reinsurer, and the regulator all get a defensible number.

Individualized reserves. Not average development factors applied to every open claim, but a specific projection for this claim given its unique combination of injury, jurisdiction, counsel, and treatment history. The reserve reflects the individual, not the portfolio.

Defensible pricing. Not "this rating variable is significant on average" but "for this applicant, with their specific exposure profile, the premium would have been $X if the variable had been different." That's the answer a regulator is looking for — and the answer Rung 2 cannot provide.

² We say reasoning deliberately. SCMs don't pattern-match — they perform logical inference over causal structures: premises in, conclusions out. That's what makes them AI in the classical sense, not just statistics with more data.

A Structural Causal Model differs from a causal Bayesian network in one critical way: it includes exogenous variables (U) — the unobserved factors that make each individual unique. In risk: construction quality, maintenance history, local soil conditions, occupant behavior — everything the data doesn't capture but the outcome depends on.

With U variables in place, the model supports a three-step counterfactual process:

The Three-Step Counterfactual Process

This three-step process is what makes Rung 3 possible. Without abduction, you can't fix U. Without graph surgery, you can't intervene cleanly. Without the structural equations, you can't propagate the intervention through the causal mechanism. In risk modeling, this is the difference between "storms increase losses on average" and "this storm caused this loss to this property."

The SCM Inference Engine

Pearl's Structural Causal Model isn't just a diagram — it's a schematic for a computational engine. It takes three inputs and produces three outputs, connecting assumptions to data to answers with mathematical precision.

The formal definition is a triplet: ⟨U, V, F⟩ — exogenous variables (what you can't see), endogenous variables (what you can), and structural equations (how they connect). This is what separates a causal model from a statistical one: the equations represent mechanisms, not correlations.

Three Levels of Query

The Rung 3 expression reads: "the probability that Y would be y had X been x, given that we actually observed X to be x' and Y to be y'." As Pearl illustrates: "the probability that Joe's salary would be y had he finished college, given that his actual salary is y' and he had only two years of college." In risk terms: the probability that this claim would have been y had mitigation been x, given that the actual claim was y' and actual mitigation level was x'.