Company-level churn moves with dozens of market and service indicators. Ranking them by correlation puts consequences and symptoms at the top. A time-series causal discovery algorithm — PCMCI with partial-correlation tests, as used in the project — separates the indicators that lead churn from those that follow it, and estimates how much each one moves it.
Each of the nine indicators is a candidate cause of churn at one, two or three weeks of lead. Series are detrended and standardized first, so a shared drift cannot pose as a relationship.
For every variable, a PC-style search keeps only lagged candidates that stay associated once the strongest other candidates are conditioned on. Each surviving link is then re-tested conditioning on both variables' parents — the MCI test — which removes links that only reflect autocorrelation or a common driver.
With the parents known, a regression of churn on them gives each lead's effect — the linear case of the outcome models (G-formula) used alongside propensity methods in the project. Business teams get a short list: which lever, how many weeks ahead, how much.