Research that moves the field.
Annotated for practitioners.

The standard assumption in causal discovery is that we know which variables were intervened on during data collection. This paper relaxes that assumption, developing identification results for causal structure even when the intervention targets are unknown — which is the realistic condition in most observational enterprise data, where the “interventions” are business decisions made without experimental control.

Why it matters: Most enterprise datasets are generated by a mix of deliberate actions and natural variation whose causal origins are ambiguous. This work extends the conditions under which structure learning produces reliable graphs — relevant to the Structure Learning question of when algorithmic discovery complements expert encoding.

Develops algorithms for combining observational and experimental data sources when each source alone is insufficient to identify the causal effect of interest. Directly addresses the most common practical situation: an organization has historical observational data and limited experimental evidence from pilots or A/B tests, and wants to use both.

Why it matters: The identification assumptions on this site’s case studies assume either purely observational data or a clean causal graph. This work provides the formal basis for the mixed case — which is the realistic one when a risk program has both historical claims data and a handful of controlled trials.

A systematic evaluation of LLM performance on causal reasoning tasks — pairwise causal discovery, counterfactual reasoning, and intervention prediction — across benchmarks. LLMs perform well on pairwise direction tasks (where they can draw on world knowledge) and poorly on counterfactual and interventional queries (where computation is required). The gap between the two widens as queries move up Pearl’s Ladder.

Why it matters: Provides empirical grounding for the architectural claim on this site: LLMs are useful for Rung 1 pattern recognition and graph structure suggestion; they cannot substitute for an Structural Causal Model (SCM) on Rung 2 and Rung 3 queries. The failure mode is not random — it is systematic and predictable from the Ladder.

Extends score-based structure learning to additive noise models without assuming a specific noise distribution. In practice, enterprise risk data rarely follows Gaussian noise assumptions — loss distributions are heavy-tailed, claim counts are over-dispersed, credit defaults cluster. This work removes a significant restriction on when algorithmic structure learning produces valid graphs.

Why it matters: The gap between structure-learning theory and practice has long included the noise distribution assumption. This narrows it — relevant to any domain where the data-generating process is non-Gaussian, which is most risk domains.

Formalizes actual causality — the “but for” question in legal and regulatory contexts — within multi-agent settings where decisions are made in parallel with partial information. Extends Halpern-Pearl actual causality to the case where responsibility is distributed across agents who cannot observe each other’s actions.

Why it matters: Directly relevant to the Insurance Attribution and Criminal Causation cases, where the legal question is exactly this: given that multiple parties acted, which ones bear responsibility for the outcome, and by how much?

The mathematical foundation. Chapters 1–3 (the do-calculus and identification theory) and Chapter 7 (counterfactuals) are the core. The rest is important but can be read in any order once those are clear.

The accessible introduction for statisticians. Shorter and more direct than Causality. The right starting point for a quantitative practitioner who wants the formal argument without the full technical apparatus.

The applied reference for Bayesian networks as decision tools. Covers belief propagation, CPT specification, influence diagrams, and LIMIDS. The book closest to what a practitioner building the models on this site actually needs.

The ML-oriented treatment. Covers structural causal models, identifiability, and structure learning from a machine learning perspective. Freely available online. Particularly useful for practitioners coming from an ML background who want to understand where the causal framework sits relative to what they already know.

Formalizes creative abduction — the inference that posits a new unobserved cause to explain correlated dispositions — as inference over a Bayesian network with a latent common-cause node. Provides three necessary probabilistic conditions and grounds the move in Reichenbach’s Common Cause Principle. The formal bridge between what senior experts do when they recognize a hidden driver and what a Bayesian network with a latent variable represents.

The standard graduate reference for observational-statistical QRM — copulas, extreme value theory, coherent risk measures, multivariate GARCH. The most complete account of how sophisticated the classical apparatus becomes before it meets the rung-3 wall. Read it to understand exactly what statistical machinery a modern risk function already has, and what questions that machinery still cannot answer.

The practitioner companion to the academic elicitation literature. Hubbard’s lasting contribution is operational: he turned the calibration-training finding from Tversky & Kahneman, Spetzler, and O’Hagan into a workshop module any risk function can run. Read it for the discipline that makes elicited probabilities defensible — and for the case that anything an organization cares about can be measured well enough to support a decision.