Cohort Retention & Attrition

Week 9 · Forecasting Wrap · how attrition erodes a cohort's revenue
Built-in AI tutor. Not sure why a lower attrition rate compounds into such a different long-run number? Ask the helper on this page. It will not tell you the retention percentage before you commit a prediction.

A cohort is the customers you acquired in one period. The moment they sign, the clock starts — some churn fast, some churn slow, and how fast depends on where they came from. A referral customer already trusts you; a paid-ad customer had to be sold. Same formula everywhere: revenue(t) = customers × ARPU × (1 − attrition)t, t in years. Different origin, different attrition rate, and the gap compounds for years after the acquisition is old news.

A · Predict before you look

Build a cohort

Pick a channel and a cohort size. Before the curve draws, commit a number: what percent of this cohort's total revenue do you think is still here after 24 months? Then reveal it.

$1,331.81
ARPU (per retained customer)
10%/yr
Annual attrition

Before you look: what percent of this cohort's revenue is still around after 24 months (2 years)? A wrong guess is fine — a guess is the point.

%

–
Retention at 24 mo
–
Revenue at selected horizon
–
Total revenue, 0–H
matches the answer key
Read the gap, not just the number. All three lines start at 100% — every cohort's revenue is intact on day one. By month 24, referral (10%/yr attrition) is still at 81%. Paid-ads (28%/yr) is down to 51.8%. Same starting revenue, same formula, different origin — the attrition rate alone is what drives that gap.
B · Same headcount, different mix

Shift the mix — origin matters at equal volume

Hold January 2025's actual combined acquisition — 1,998 customers — constant, and change only how they split across channels. A dollar of paid-ad revenue and a dollar of referral revenue look identical on day one, but they diverge over time because the two channels churn at different rates.

Referral16.6%
Paid ads52.6%
Organic (the remainder)30.8%
–
This mix, total revenue 0–48 mo
–
If 100% referral, 0–48 mo
–
If 100% paid ads, 0–48 mo
C · Cohorts stack up over time

This channel's 2025 cohorts, rippling forward

Every month brought a new cohort through the channel you picked in section A. Each one starts decaying the moment it lands. The revenue in any given month is the sum of every cohort still alive — this quarter's acquisitions are still echoing two or three years out.

D · The forecasting tie-in

A one-point miss on attrition compounds

Attrition is not an observed fact — it is a forecast input, the same way a growth rate or a discount rate is. Nudge the assumption by one percentage point in either direction for the cohort from section A and read what happens to the dollar forecast as the horizon stretches out.

Horizon−1 pt attritionBaseline attrition+1 pt attritionSpread (hi−lo) ÷ base
Where this goes next. This is exactly why the week's hedging block (later in the deck) sizes a pad from measured error rather than a guess: a backtest MAE/RMSE tells you how wrong your method has been before; attrition is exactly the kind of consequential, unobserved assumption that deserves the same discipline — flag it, and say how much the forecast moves if it is off by a point.