Docs/Causal Model

Causal Model

The causal model is Shingo's core intelligence layer. It maps the relationships between technical changes, user behavior, and revenue outcomes, identifying actual causes rather than statistical correlations.

Why causal inference matters

Correlation tells you two things happened around the same time. Causation tells you one thing made the other happen. The difference determines whether your team spends a sprint on the right fix or chases symptoms.

Shingo uses counterfactual analysis: “What would have happened if this deploy hadn't shipped?” By comparing user cohorts who were and weren't affected by a change, Shingo isolates the causal effect with a confidence score.

How the model is built

01

Data ingestion

Shingo pulls events from all connected sources: billing transactions from Stripe, user actions from PostHog/Amplitude, errors from Sentry, and deploys from GitHub. These are unified into a single event timeline per user.

02

Session reconstruction

Events are grouped into user sessions. Each session captures the full behavioral journey: page views, feature usage, errors encountered, and revenue actions (purchases, upgrades, cancellations).

03

Pattern compression

Thousands of sessions are compressed into behavioral patterns. Shingo identifies repeating sequences of actions that lead to specific revenue outcomes, filtering out noise and one-off events.

04

Causal graph construction

Patterns are connected into a directed causal graph. Each edge represents a causal relationship: 'deploy X caused behavior Y which caused revenue outcome Z.' Edges are weighted by confidence and dollar impact.

05

Counterfactual validation

Each causal link is validated through counterfactual analysis. Shingo compares affected vs. unaffected cohorts to confirm the causal relationship, producing a confidence score (typically 85-97%).

Confidence scores

Every finding includes a causal confidence score. This tells you how certain Shingo is that the identified cause actually produced the observed revenue effect.

90-100%

High confidence

Strong counterfactual evidence. Clear causal chain from change to revenue impact.

75-89%

Moderate confidence

Likely causal, but some confounding factors exist. Worth investigating.

< 75%

Low confidence

Suggestive but not definitive. Shingo flags these but recommends manual review.

What the model detects

Broken user flows

Code changes or deploys that silently break checkout, onboarding, or upgrade paths.

Permission gates

Access control changes that block revenue-generating actions for specific user segments.

Performance regressions

Latency or timeout changes that cause users to abandon high-value actions.

Feature flag side effects

Flag rollouts that unintentionally impact revenue in non-obvious ways.

Third-party breakage

Library updates or API changes from vendors that break your revenue flows.

Cohort-specific issues

Problems that only affect high-LTV users, enterprise accounts, or specific segments.