A claim about Quenchly

Usage decays in the weeks before a voluntary cancel.

Weekly event counts for accounts that later cancel, aligned to the cancel date, fall through the final six weeks. The last four weeks run at 80% of the trailing baseline or less.

holds on the current window

as of moments ago

Usage before voluntary cancel

WITH cancels AS (
  SELECT s.account_id, e.at AS canceled_at
  FROM subscription_events e
  JOIN subscriptions s ON s.id = e.subscription_id
  WHERE e.kind = 'cancel' AND e.reason = 'requested'
),
events AS (
  SELECT c.account_id,
         CAST((julianday(c.canceled_at) - julianday(a.at)) / 7 AS INTEGER) AS weeks_before
  FROM app_events a
  JOIN users   u ON u.id = a.user_id
  JOIN cancels c ON c.account_id = u.account_id
  WHERE a.at <  c.canceled_at
    AND datetime(a.at) >= datetime(c.canceled_at, '-84 days')
)
SELECT weeks_before,
       COUNT(*)                     AS events,
       COUNT(DISTINCT account_id)   AS accounts
FROM events
GROUP BY weeks_before
ORDER BY weeks_before DESC;

View data table

weeks_before	events	accounts
11	4968	327
10	5156	330
9	5017	326
8	6084	400
7	6728	398
6	6472	400
5	5995	398
4	5761	474
3	5197	477
2	4098	466
1	3076	446
0	2061	410

holds on the current window

Customer-success tooling sells on one chart: engagement falling before the churn event, early enough to act on. In random demo data the chart is flat until the account disappears. In Quenchly every account carries a usage latent that fades over its final six weeks before a requested cancel, so the event volume in app_events drops first and the cancel row arrives afterward. Payment-failure cancels do not fade the same way, because the customer did not decide to leave.

What you should see

Twelve rows, weeks 11 down to 0 before the cancel. Event counts hold roughly level through the first six, then fall week over week to the end. Divide the sum of weeks 3 to 0 by the average of weeks 11 to 4 and the ratio is 0.8 or lower. It averages about 0.69 from one seed to the next. Repeat the query with reason = 'payment_failure' and the slope mostly disappears.

Run it yourself

SQLite one file, ready to query CSV tar.gz, one file per table