Skip to content
Back to the site

Method demonstration

Proof, on data where the answer is known.

Your own data has no answer key, so a churn model can score well on it and still be wrong. This page runs ours on a simulated game where we planted the truth, blind to it, and shows what came back.

What it recovers, every time:

the real driver of churn
accepted
a convincing decoy
rejected
the sampling trap that flatters an ordinary model
refused

What it recovered, blind to the truth

FIG. 1 Accepts the real driver, rejects the decoy
[ Illustrative synthetic data — demonstration only ]
Accepts the real driver, rejects the decoy Bar chart. A validation gate accepts the real driver of churn — a large, clear effect well above the accept threshold — and rejects a convincing decoy that sits near zero. Illustrative synthetic data, a demonstration on planted truth. 0.0 0.5 1.0 1.5 2.0 2.5 the real driver accepted a decoy rejected accept threshold effect size

The real driver clears the accept threshold; a convincing decoy with no real effect does not.

FIG. 2 The recovered exit shape
[ Illustrative synthetic data — demonstration only ]
The recovered exit shape Histogram. The exit shape recovered from synthetic telemetry (bars) closely tracks the planted shape (reference curve) — a right-skewed distribution. Illustrative synthetic data, a demonstration on planted truth. 0 1 2 3 how abruptly players leave → recovered planted

Read blind, the recovered shape of exit tracks the shape we planted.

FIG. 3 Two kinds of exit, counted separately
[ Illustrative synthetic data — demonstration only ]
Two kinds of exit, counted separately Line chart. Competing-risks cumulative incidence over eighteen weeks: two kinds of exit counted separately. One exit begins accumulating from the start; the other stays flat for five weeks, then climbs. Counted as one curve, that difference would disappear. Illustrative synthetic data. 0.0 0.1 0.2 0 5 10 15 one kind of exit the other cumulative incidence weeks

Competing-risks cumulative incidence (Aalen–Johansen): one exit starts accumulating from week one, the other stays flat for five weeks and then climbs. Counted as a single churn curve, that difference — and the different fix each one calls for — disappears.

And it holds under the tools a quant team expects by name — run against the planted truth, every one recovers the real cause and refuses the confound:

  • time-varying Cox regression
  • Aalen–Johansen competing risks
  • random survival forest
  • Harrell’s C · time-dependent AUC
  • King–Zeng case-control correction
Book a call See it move — self-exciting engagement The retraining trap The shutter Read the thinking

or write: hello@dividual.studio