Loyalty & retention

Streak & Retention Hazard Model

The cheapest day to save a user is almost never day one. This finds the day.

Model 7 of The Marketing Quant Python Kit. Runs entirely in your browser — nothing you type is sent anywhere.

Your streak

%

Share who come back after their first day. This is the hardest day.

%

Where daily retention settles once the habit is formed.

How fast continuation climbs toward the asymptote. Higher means the habit locks in sooner.

days
%

Percentage-point improvement in continuing, for someone who receives the reward.

$

With a fixed budget the ranking inverts: early days can no longer reach everyone, so the cheap late day stops being obviously best.

Where to place the reward

Cheapest day: day 29

Cost per additional survivor at the horizon.

$20.00

Baseline survivors

Reaching day 30 with no reward at all.

431.6

Users reached on that day

454

Spend

$909

Extra survivors bought

45.4

By day

DayPaidSpend$/survivor
110,000$20,000$254.87
33,675$7,349$127.95
71,677$3,353$70.00
14995$1,991$43.63
21686$1,371$30.17
29454$909$20.00

Early rewards reach everyone but most of them were going to continue anyway. Late rewards reach a small, self-selected group who are far more likely to be saved by it.

Every input here is an assumption about your product, not a benchmark. Fit day-1 continuation, asymptote and k to one real cohort before trusting the output — the shape of the curve drives the answer far more than the reward cost does.

Free resource

Get the Gamification Mechanics Playbook

Streaks, leaderboards and loyalty tiers as measurable systems. Where streaks break, why global leaderboards drive off 95% of players, and the retention lift a tier has to clear.

No spam. Unsubscribe anytime.

Browse all free guides →
Read next

The maths behind this

Other tools

Want this fitted to your actual numbers?

The defaults here are illustrative. Fitted to your own data — real hazard rates, real margins, real conversion — the same models tell you what to do next rather than what is theoretically possible.

Get it modelled properly