Cohort analysis на собеседовании Data Scientist

Зачем разбирать на собесе

Cohort analysis — стандарт product analytics. На собесе DS: «как считать retention», «зачем cohorts».

Что такое cohort

Group of users sharing characteristic — обычно signup date.

Acquisition cohort. Все, кто signed up в week / month X.

Behavioral cohort. Все, кто did action X (made purchase, completed onboarding).

Зачем cohort:

• Tracking metric over time без survivor bias.
• Сравнение cohorts — improving / degrading?
• Понимание product changes impact.

Retention cohort

Most common.

Cohort: users signed up in May 2026
Day 0:  100% (all 1000 users)
Day 1:  60%
Day 7:  40%
Day 30: 25%
Day 90: 18%

SQL:

WITH cohorts AS (
  SELECT user_id, DATE_TRUNC('week', signup_date) AS cohort_week
  FROM users
),
retention AS (
  SELECT c.cohort_week,
         (e.event_date - c.cohort_week) / 7 AS week_n,
         COUNT(DISTINCT e.user_id) AS active
  FROM cohorts c
  JOIN events e ON e.user_id = c.user_id AND e.event_date >= c.cohort_week
  GROUP BY 1, 2
)
SELECT cohort_week, week_n,
       100.0 * active / FIRST_VALUE(active) OVER (PARTITION BY cohort_week ORDER BY week_n)
FROM retention;

Behavioral cohort

Users defined by action, не sign-up.

Examples:

• Users who made first purchase в May 2026.
• Users who completed onboarding tutorial.
• Users who upgraded to premium.

Сравнение их retention с control (didn't do action).

Triangle visualization

Standard cohort table:

Cohort     | Day 0 | Day 1 | Day 7 | Day 30 | Day 90
2026-Apr   | 100%  | 65%   | 42%   | 28%    | 20%
2026-May   | 100%  | 68%   | 45%   | 30%    | --
2026-Jun   | 100%  | 70%   | 47%   | --     | --
2026-Jul   | 100%  | 72%   | --    | --     | --

Reading: Down — newer cohorts. Right — time since acquisition. Empty — еще не дожили.

Тренды:

• Newer cohorts retain better (good — improvements working).
• Newer cohorts drop faster (bad — quality customer worse).
• Same retention — stagnation.

Связанные темы

• Survival analysis на собесе DS
• Causal inference для DS
• Funnel conversion для PM
• Активation rate для PM
• Подготовка к собесу Data Scientist

FAQ

Как handle censoring?

Newest cohorts не дожили to day 30 — empty cells. Отдельно от 0% retention.

Это официальная информация?

Нет. Статья основана на стандартных подходах product analytics.

Тренируйте Data Science — откройте тренажёр с 1500+ вопросами для собесов.