Logo - Branislav Mateas - BMateas.com

RET / Retention signals

Repeated return behavior mattered more than launch volume

Renewal followed sustained recurrence—not a first-week spike or one magic activation threshold.

Signal trail

  1. 1

    Early launches

    Weak

  2. 2

    Total activity

    Useful

  3. 3

    Active-day recurrence

    Most useful

The brief

The question behind the work

A confidential subscription business wanted to understand which observable usage patterns were associated with renewal. The work was intended to help lifecycle and product teams identify risk without turning correlation into a causal claim.

Which aspects of customer activity best identify renewal risk—and which behaviors are reliable enough to guide lifecycle experiments?

Evidence available

  • Completed subscription cycles with a defined renewal outcome
  • Early usage, full-cycle activity, active days, and usage concentration
  • Acquisition route, renewal configuration, billing, and market descriptors
  • Confidence intervals, regression, discrimination, and interaction checks

How I worked

From raw evidence to a defensible answer

  1. 01

    Freeze a mature cohort

    Include only subscriptions that completed the decision window, then define renewal once and remove duplicate outcome fields.

  2. 02

    Engineer competing signals

    Contrast early launches, total activity, unique active days, launches per active day, and concentrated versus distributed use.

  3. 03

    Compare signal quality

    Use interval estimates, regression, and discrimination checks to see which measures add information beyond acquisition context.

  4. 04

    Test the shape of the relationship

    Look for a genuine threshold, a gradual curve, and differences between customers with automatic and manual renewal paths.

  5. 05

    Design the intervention

    Translate recurrence and inactivity into targeting rules, then require randomized holdouts before claiming an effect on renewal.

What the evidence said

The findings that survived validation

F1

Renewal rose with sustained engagement, but the pattern was gradual rather than governed by a universal activation threshold.

F2

Early launch volume added little useful information compared with behavior observed across the customer lifecycle.

F3

Unique active days were more informative than raw volume; repeated launches on the same day contributed little.

F4

Renewal configuration had a separate association, while recurrence remained useful for identifying risk in manual-renewal journeys.

Decisions

Signal → interpretation → move

D1

A strong onboarding burst

Interpretation

First-week launch count was a weak long-term retention signal in this analysis.

Business move

Stop using early intensity as the primary activation rule.

D2

High total activity

Interpretation

Volume carried information, but could still be concentrated into a short burst.

Business move

Separate how much someone used the product from how often they returned.

D3

More distinct active days

Interpretation

Recurrence was the most useful measured behavioral signal, without proving causation.

Business move

Target increasing inactivity gaps and test habit-supporting messages with holdouts.

Delivery & limits

What the client could use next

Delivered

  • Mature-cohort and outcome definition
  • Behavioral feature specification
  • Renewal-by-engagement diagnostic
  • Predictor-strength comparison
  • Lifecycle targeting hypotheses
  • Measurement and validation agenda

Where the evidence stops

  • The work was observational and did not prove that increasing activity causes renewal.
  • No holdout or temporal validation was documented, so the analysis is not described as a deployed predictive model.
  • Feature timing needs to be confirmed before any operational score is used to rule out leakage.

Your decision

Have a question your current dashboard cannot answer?

Let’s find the right grain, evidence, and test before another metric becomes the strategy.

Start a conversation