Write a decision memo from a retention analysis

A retention decision memo should lead with the proposed action and the evidence that supports it. Explain the population, return definition, observation horizon and limitations before making a causal claim. A chart with a declining line is not, by itself, a recommendation.

The example below uses the original synthetic product analytics dataset. It demonstrates a reviewable decision process, not an observed result for NeuraPath or any real product.

Reproduce the facts before writing the recommendation

Define activation as a valid publication within seven elapsed days of signup, with an exclusive upper boundary. Define retention as a meaningful report event on the UTC calendar date seven dates after signup. Include only users whose required windows are fully observed by January 16 at 00:00 UTC.

The contracts differ intentionally. The retention tutorial explains the difference between calendar ages and elapsed durations.

from datetime import date, timedelta
from build_and_verify import database, active_ids, ACTIVATION_SQL

db = database()
activation = dict(db.execute(ACTIVATION_SQL).fetchall())
rows = []
cutoff = date(2026,1,16)
for user,signup in db.execute("SELECT user_id,signup_at FROM users WHERE user_kind='external' AND is_bot=0"):
    start = date.fromisoformat(signup[:10])+timedelta(days=7)
    end = start+timedelta(days=1)
    if end <= cutoff and user in activation:
        active = set(active_ids(db,str(start)+'T00:00:00Z',str(end)+'T00:00:00Z'))
        rows.append((user,activation[user],int(user in active)))
db.close()
assert rows == [('U1',1,1),('U2',0,0),('U3',1,1),
                ('U4',0,0),('U5',1,0),('U6',0,1)]
activated = [r for r in rows if r[1]]
not_activated = [r for r in rows if not r[1]]
assert sum(r[2] for r in rows) == 3 and len(rows) == 6
assert sum(r[2] for r in activated) == 2 and len(activated) == 3
assert sum(r[2] for r in not_activated) == 1 and len(not_activated) == 3
print({'mature_users':6,'day7_returning':3,
       'activated_returning':'2/3','not_activated_returning':'1/3'})

U9 is excluded because its day-seven return period has not occurred. U6 returns on its calendar day seven but did not activate within the strict elapsed-time window. Preserving that distinction prevents the memo from quietly redefining the metric to simplify the story.

Example decision memo

Decision requested: prioritize an investigation of publication friction and validate the activation proxy before proposing an onboarding rollout. The current evidence does not support claiming that increasing publication will cause better retention.

Observed evidence: among six mature external nonbot users, three have a qualifying day-seven return event. Two of three activated users return, compared with one of three nonactivated users. These are small descriptive counts from a synthetic teaching fixture.

Interpretation: activation and later return are associated in this sample, but motivation, account needs and other differences could explain the pattern. The groups were not randomized. The observed difference is not a causal effect estimate.

Diagnostic evidence: U4 has an invalid publication event; U2 publishes after the activation window; U6 publishes exactly at its exclusive boundary. These cases suggest separate checks for output validity, time to completion and boundary interpretation. They do not establish the underlying reasons for those behaviors.

Recommended next work: inspect the publication journey, confirm event validity against application outcomes, and collect evidence about obstacles to successful publication. If a specific onboarding change follows, design its assignment, primary outcome, guardrails and observation horizon before evaluating results.

Decision limits: six users provide little evidence about stability across cohorts or customer types. Source completeness is assumed by this controlled fixture. A real memo must verify production coverage and document missing or delayed events.

Follow-up artifact: a reproducible cohort table, a short issue list linked to verified evidence, and an experiment plan only after a concrete intervention and measurement contract are defined.

Make uncertainty actionable

“More research is needed” is too vague to guide work. Name the uncertainty that could change the decision. Here, event validity affects whether activation was measured correctly; user-level investigation affects which intervention is plausible; a suitable experiment affects whether a causal benefit can be established.

Avoid converting a descriptive association into a numerical uplift promise. Also avoid assuming that nonreturn on one particular day means permanent churn. A user may return later or use the product on a different cadence.

Keep the memo independently reviewable

Attach the query or executable calculation, data cutoff, metric definitions and counts. A reviewer should be able to reconstruct the headline without a private dashboard session. Keep customer-level records restricted appropriately in a real setting; a decision memo rarely needs raw personal information.

Exercise: change the return metric to any qualifying event during calendar days seven through thirteen. Identify which cohorts now lack a complete window before calculating the revised result. Explain whether the original recommendation changes and why.

NeuraPath's Data Analytics with Generative AI course connects technical analysis with business communication. A strong portfolio memo makes a decision easier to assess while keeping facts, interpretations and proposed tests distinct.

Continue learning

This article is part of the Customer and product analytics sequence. Use the neighbouring tasks when you need the prerequisite or the next application.

• Review the preceding task in Design a product analytics event naming standard.
• Return to the cluster foundation in Define active users before calculating DAU and MAU.