Skip to content
Methodology 4 of 8

Causal Inference in Compass

Why correlation is not causation, and what Compass does about it.

Oil prices and airline stock prices are inversely correlated. This much any regression can tell you. But can you buy airline stocks when oil drops and reliably profit? That depends on something no correlation can answer: is the relationship causal, or are both driven by a third factor you have not measured?

If oil drops because of a global recession, airline stocks are going down too — the correlation reverses under intervention. If oil drops because of a Saudi production decision, airlines benefit. The data looks identical in both cases. The causal structure is completely different. And the causal structure is what determines whether your trade works.

This is why Compass builds causal DAGs (directed acyclic graphs) rather than relying on correlations. A DAG is a map of which variables cause which. It makes assumptions explicit. And once you have a DAG, you can do something correlations never allow: you can reason about interventions.

What a DAG captures

A DAG has two kinds of objects: nodes (variables) and directed edges (causal influence). An arrow from A to B means: if you could intervene on A (holding everything else fixed), B would change. The key word is could — a DAG is a hypothesis about what would happen under intervention, not a description of what has happened in historical data.

Four edge semantics the Compass DAG service supports:

  • Direct cause — A directly influences B with no intermediate.
  • Common cause — A and B share a hidden cause C; the system introduces C as an intermediate node and emits two edges C→A and C→B.
  • Mediator — A influences B through C; the system emits A→C and C→B.
  • Aggregator — B aggregates multiple upstream signals (e.g. a prediction market price aggregates many traders).

D-separation: the algebra of independence

Once you have a DAG, d-separation tells you which pairs of variables are conditionally independent given a conditioning set. If two evidence items are d-separated given the hypothesis, they carry independent information. If they're d-connected, they share a common cause and their joint evidential weight is less than the sum of their parts.

Compass runs d-separation on every analysis's DAG and uses the result to scale cell weights in the ACH matrix. Evidence that shares a common cause gets reduced effective weight — still counted, but counted less, reflecting the partial redundancy.

Building the DAG

Two complementary paths:

  • AI-proposed: Given a set of evidence items, Compass asks a causal-reasoning LLM to identify likely relationships — direct cause, common cause, mediator, aggregator — and returns proposed edges the user can accept, override, or extend. Each proposal carries a distributional weight (p05/p25/p50/p75/p95) and a list of evidence_refs that justify the claim.
  • User-drawn: Analysts can add edges by hand. Each edit validates that the graph remains acyclic before committing. Hand-drawn edges are tagged user_created and persisted with full provenance.

Why explicit beats implicit

A regression hides causal assumptions. A DAG makes them visible. When analysts disagree, they can point to a specific edge and argue about whether it should be there. When an intervention proposal is evaluated, the DAG tells you exactly which downstream variables to expect to change — and which should be held fixed in the counterfactual. This is the difference between reasoning and curve-fitting.

From the book: Beyond Synthesis, Chapter 9–12: “The Causal Lens.”