Methodology

The explanation engine analyzes stored metric history, then generates a narrative from the available analysis. The method that runs depends on the data and enabled capabilities. These are descriptions of implemented methods, not measured customer accuracy or causal guarantees.

Tier 1: Pattern Correlation Pearson correlation compares metric series aligned by date. It requires at least three shared observations, skips zero-variance series and returns up to ten relationships with absolute correlation above 0.5. These findings describe co-movement, not causation.

Tier 2: Temporal Analysis PELT changepoint detection uses a squared-error cost and BIC-style penalty to identify shifts in a metric series. Temporal analysis compares the order of detected shifts and their magnitude, with a minimum history requirement. The lead-time calculation assumes uniformly spaced observations. Temporal precedence alone does not establish a cause.

Tier 3: Quasi-Experimental Estimation When Tier 3 is enabled and sufficient history is available, the in-process difference-in-differences estimator splits the analysis window at its midpoint. It compares metric groups, ranks candidate controls by pre-period similarity, and applies sample-size, parallel-trend and statistical-significance checks. This is not automatic identification of a verified natural experiment.

An optional counterfactual service can run interrupted time-series estimation when it is configured and sufficient pre-period history exists. The service also implements synthetic control for requests with treated and control units, but the explanation pipeline does not currently supply those units. Low-confidence bootstrap results are not used as Tier 3 explanation evidence.

Tier 3 includes a causal or correlational evidence label derived from estimator checks. Difference-in-differences uses a significance threshold of p < 0.05 with assumption checks; optional counterfactual estimates compare the observed result with a 90% credible interval. These checks do not rule out unobserved confounding or validate the business cause. An inconclusive result may remain correlational or fall back to a lower tier.

Graph Discovery: Not Live in Explanations Graph discovery is not connected to customer-facing explanations. We do not claim graph-derived causal answers.

Data Requirements and Limits If the requested metric has insufficient stored history, the explanation engine returns an insufficient-data response rather than a generated explanation. Source data gaps, irregular sampling and estimator assumptions limit what analysis can establish. Review the underlying data and consider alternative explanations before acting. No customer accuracy percentages, outcome benchmarks or universal causal-answer guarantees are claimed here.