Cohort-Based Revenue Forecasting: Retention Curves, Not Blended Averages

Most SaaS revenue forecasts fail the same way: one churn number, applied to every customer, every month. That single blended rate — the one your billing dashboard shows — quietly assumes a customer who signed up yesterday behaves like one who has paid for a year. They don't. New customers churn three to ten times faster, and a forecast built on the blended rate overstates month-24 MRR by 30–45%.

This guide shows how to forecast revenue cohort by cohort instead: build the retention grid, read the curve, benchmark it against 2026 numbers, estimate it when you have almost no history, and turn it into an MRR forecast an investor can't poke holes in. One fictional startup — Lumo, a $15/month consumer photo tool — runs through every calculation.

Why a blended churn rate overstates your forecast

The early-life cliff and the steady-state plateau

Every subscription product has a retention curve, and almost every retention curve has the same two features: a cliff in months 1–3, then a plateau.

Lumo adds roughly 1,000 new paying customers a month. Of each cohort, 70% are still paying after one month, 55% after three, 46% after six, 36% after twelve. Beyond the first year, survivors churn at about 3% a month — right in the healthy steady-state band. Month-1 churn is 30%. Month-13 churn is 3%. Same product, same price, a 10x difference in behavior depending on customer age.

Now look at what Lumo's dashboard reports. Blended churn is cancellations divided by the customer base, and the base is dominated by survivors — people who already made it past the cliff. With this curve and steady inflow, Lumo's base settles around 18,300 customers and loses about 1,000 a month: blended churn of 5.5%. A shade above the healthy 3–5% B2C band from our SaaS churn rate guide — worth watching, nothing dramatic. The number is mathematically true and predictively useless, because no actual customer churns at 5.5%. Newcomers churn much faster; veterans churn much slower.

The divergence, in numbers

Feed identical inputs into two forecast engines: 1,000 new paying customers a month at $15 ARPU, starting from zero. The blended engine applies 5.5% monthly churn to whatever base exists. The cohort engine applies Lumo's real curve to each monthly cohort separately.

Forecast month	Blended engine (5.5%/mo)	Cohort engine	Blended overstates by
6	$78,500 MRR	$57,900 MRR	+36%
12	$134,400 MRR	$95,000 MRR	+41%
18	$174,200 MRR	$125,100 MRR	+39%
24	$202,600 MRR	$150,100 MRR	+35%

At month 24 the blended model shows $2.43M of exit ARR. The cohort model — the one that matches how customers actually behave — shows $1.80M. That is $630,000 of phantom ARR, and it exists because the blended engine retains 94.5% of every new cohort in its first month when reality retains 70%. A growth forecast's future revenue comes mostly from cohorts that don't exist yet, so mispricing the first months of a cohort's life poisons the whole projection.

The direction of the error isn't random. Measured on a mature, survivor-heavy base, the blended rate is always too kind to new customers. (Measured on a very young, fast-growing base, it can err the other way.) Either way it is an artifact of base composition, not a behavioral constant — which disqualifies it as a forecasting input.

What a cohort is — and how to build the grid

A cohort is every customer whose first paid month is the same calendar month. Not signup month, not trial-start month — first paid month. Trials that never convert belong to your funnel math, not your retention math.

The cohort retention grid puts signup months in rows and months-since-signup in columns. Each cell is the share of the cohort still paying at that age. Here is Lumo's actual grid for January–May:

Cohort	Size	M0	M1	M2	M3	M4
January	850	100%	71%	61%	56%	52%
February	920	100%	69%	59%	54%	51%
March	1,010	100%	72%	62%	56%	—
April	1,080	100%	68%	58%	—	—
May	1,140	100%	70%	—	—	—

The formula behind each cell:

Retention(a) = Customers from the cohort still active at age a ÷ Cohort size at month 0

The grid is a triangle because recent cohorts haven't lived long enough to fill their rows. Reading it is a skill worth 20 minutes of practice:

• Across a row you see one cohort's life: January lost 29 points in month 1, 10 more by month 2, then the decay slowed. That is the curve.
• Down a column you see cohort quality over time at a fixed age. Lumo's M1 column reads 71, 69, 72, 68, 70 — stable within ±2 points. A downward trend here means acquisition quality is slipping (usually a channel-mix shift), and it shows up months before it dents blended churn.
• Down a diagonal you see calendar effects. A price change, an outage, or a dunning fix hits every cohort in the same calendar month — a diagonal stripe across the grid.

The empty lower-right corner is the future. Forecasting is the act of filling in that triangle with a fitted curve.

The three retention-curve shapes

Cohort curves come in three shapes, and each implies a different business.

Fast decay: the consumer curve

Lumo's shape. 25–40% of the cohort gone in month 1, decay slowing sharply, plateau by month 4–6 at 2.5–5% monthly churn. Most B2C and prosumer self-serve products look like this. The implication: your economics are decided at the cliff. Survivors are fine — it's how many customers reach the plateau that sets lifetime value, which is why month-1 retention is the single highest-leverage number in a consumer subscription business.

Flattening: the good B2B curve

SMB B2B done well: 80–90% month-1 retention, gentle decay, flattening toward 2–3% monthly churn within the first year (1–2% for mid-market). The tail carries the value: the area under the retention curve is your expected customer lifetime, and with a flattening curve most of that area sits beyond month 6. This is the geometry behind LTV in SaaS — LTV is mostly a question of how flat your tail is.

The smile: negative churn

Logo retention can only fall. Revenue retention can rise. When surviving customers expand — more seats, higher tiers, more usage — a cohort's revenue retention can dip to 85–90% by month 6 and climb back above 100% by month 18–24. That is negative churn at the cohort level, and it changes the forecast qualitatively: each cohort becomes an appreciating asset, and the base grows even with zero new sales. If your cohorts smile, a straight-line LTV formula understates you badly.

Cohort retention benchmarks for 2026

Logo retention of paying customers on monthly plans, measured against cohort size at month 0:

Retention at age	B2C self-serve	SMB B2B	Mid-market
Month 1	60–75%	80–90%	92–97%
Month 3	45–60%	70–82%	87–93%
Month 6	35–50%	62–75%	82–90%
Month 12	25–40%	50–65%	75–85%

Two notes on reading this table. First, B2C month-12 retention of 25–40% looks brutal next to the "3–5% monthly churn is healthy" benchmark — both are correct. The cliff removes 30–40% of a consumer cohort early; the survivors then churn at 3–5% monthly, and compounding the two produces exactly these month-12 numbers (Lumo: 70% × eleven months of plateau churn ≈ 36%). Second, the bands are wide because price point and onboarding vary; sitting at the top of your segment's band at month 12 puts you in the upper quartile.

Estimating a curve with under six months of data

You have four cohorts, none older than five months, and an investor wants a 36-month forecast. You cannot observe month-12 retention that hasn't happened. You can borrow it.

The method: benchmark curves within a segment differ in level far more than in shape. Express the shape as ratios to month 1. For B2C self-serve, typical ratios are M3/M1 ≈ 0.78, M6/M1 ≈ 0.65, M12/M1 ≈ 0.51 (Lumo's curve: 0.79, 0.66, 0.51). Then:

1. Measure your own months 1–3 from the grid — the cliff is observable almost immediately.
2. Anchor the borrowed shape to your level. If your M1 is 64%, project M12 ≈ 64% × 0.51 ≈ 33%.
3. Beyond month 12, hold conditional churn flat at your segment's steady state — for B2C, pick 4% until data argues otherwise.

Use the pessimistic half of the benchmark range while you're young. Early cohorts are full of friendlies — your network, early adopters, people who found you organically — and scaled paid channels almost always retain worse.

Re-fit the curve on a schedule, not on a feeling: monthly, as each new diagonal of the grid fills in; immediately, if two consecutive cohorts land more than 5 points off the projection at the same age; and from scratch after any change to pricing, onboarding, or channel mix — post-change cohorts belong to a new curve family and should not be averaged with the old one.

Expansion revenue inside cohorts

Everything above tracked logos. Serious forecasts track a second curve per cohort:

Cohort revenue retention(a) = MRR from the cohort at age a ÷ the cohort's MRR at age 0

Lumo sells a $5/month add-on, and survivors adopt it over time: average ARPU among retained customers climbs from $15 at signup to $18 by month 12. Logo retention at month 12 is 36%, but revenue retention is 36% × 18/15 = 43.2%. Layering ARPU-by-age into the forecast lifts Lumo's month-24 MRR from $150,100 to roughly $168,400 — a 12% difference a logo-only model simply cannot see.

Per-cohort NRR is also the durability metric investors check. Take each vintage's revenue retention at a fixed age — say month 12 — and line the vintages up. Lumo's reading 43% steady across vintages is fine; a B2B product whose M12 revenue retention climbs from 95% to 105% across vintages is showing compounding in its rawest form. A declining vintage trend is the earliest honest warning that the business is degrading, and no blended metric will surface it.

Turning cohorts into an MRR forecast

The whole machine is one formula applied repeatedly:

MRR(t) = Σ over cohorts c: New paid customers(c) × Retention(t − c) × ARPU at age (t − c)

Three inputs, each independently testable:

1. New cohort inflow — how many customers start paying each month. This comes out of your acquisition funnel: sessions × visit-to-signup × signup-to-paid. Benchmarks for each stage are in our conversion funnel guide.
2. The retention curve — fitted from the grid, one curve per segment or plan.
3. ARPU by age — flat if you have no expansion, rising if you do.

Each cohort then decays along the curve while new cohorts stack on top. Lumo's first six months, at flat 1,000 customers/month inflow and flat $15 ARPU:

Cohort	Jan	Feb	Mar	Apr	May	Jun
January	$15,000	$10,500	$9,000	$8,250	$7,800	$7,350
February	—	$15,000	$10,500	$9,000	$8,250	$7,800
March	—	—	$15,000	$10,500	$9,000	$8,250
April	—	—	—	$15,000	$10,500	$9,000
May	—	—	—	—	$15,000	$10,500
June	—	—	—	—	—	$15,000
Total MRR	$15,000	$25,500	$34,500	$42,750	$50,550	$57,900

Read it as geology: each row is a layer that thins as it ages, each column sums to the month's MRR. Growth is a race between the thickness of the newest layer and the combined decay of every layer underneath. The model also makes your ceiling visible: with flat inflow, Lumo's MRR plateaus near $275,000 — 1,000 customers × 18.3 expected paid months × $15. Nothing but more inflow, better retention, or higher ARPU moves that ceiling, and the cohort model tells you which lever pays best before you spend on any of them.

Four mistakes that corrupt cohort forecasts

1. Averaging cohort percentages

A simple average of retention rates ignores cohort size. A 400-customer cohort at 62% month-3 retention averaged with a 1,600-customer cohort at 54% gives 58%; the weighted truth is 55.6%. Two and a half points at month 3 compounds into a real revenue miss by month 24. Worse is averaging across eras: if recent cohorts retain worse than last year's, the long-run average flatters you exactly when you need the warning. Fit the curve to recent vintages; use old cohorts only to estimate tail shape.

2. Mixing billing cycles in one curve

An annual prepaid customer cannot cancel in months 1–11. Blend annual and monthly plans into one curve and you get a flattering flat line followed by a cliff at month 12–13 that your forecast never saw coming. Keep separate curves: monthly plans get the monthly retention curve; annual plans get renewal rates (first renewal typically 40–60% for B2C, 70–85% for SMB B2B). Model refunds on annual plans in month 0–1, where they actually occur.

3. Trusting mid-contract retention on annual plans

Month-6 "retention" of an annual cohort is contractual, not behavioral — it includes customers who mentally churned in week 3 and are merely locked in. The honest mid-contract signal is usage, not billing status. Be especially suspicious of cohorts acquired on deep annual discounts: they post beautiful year-1 retention and then collapse at first renewal, which is survivorship bias wearing a party hat.

4. Ignoring seasonality of cohort quality

Cohorts are not interchangeable across the calendar. January resolution-buyers and Black Friday discount-hunters routinely retain 5–12 points worse at month 3 than list-price cohorts from March. If a quarter of your annual inflow arrives in Q4 promos, forecasting Q4 cohorts with your average curve overstates next year's revenue. Tag cohorts by acquisition month, channel, and promo status, and give promo cohorts their own haircut.

What a 5-point drop in month-1 retention costs

Cohort models earn their keep in sensitivity analysis, because you can move one behavioral assumption and watch it propagate.

Drop Lumo's month-1 retention from 70% to 65% — five points, the kind of slip a broken onboarding flow causes — with conditional retention after month 1 unchanged, so every later point on the curve scales by 65/70:

• Month-24 MRR: $150,100 → $140,500
• Year-2 exit ARR: $1.80M → $1.69M, a loss of $116,000 (−6.4%)
• Cumulative revenue over 24 months: down $133,000
• Expected paid lifetime: 18.3 → 17.1 months; revenue per acquired customer $275 → $256, which with unchanged CAC takes ~7% straight off LTV:CAC

The upside is symmetric: five points of month-1 improvement adds roughly the same $116k of exit ARR, making onboarding the cheapest ARR Lumo can buy. Run these curve shifts as named scenarios — base, cliff-worsens, cliff-improves — alongside your CAC and pricing cases; the mechanics are covered in our guide to scenario planning and sensitivity analysis.

How this shows up in your financial model

Cohort logic is the revenue backbone of a fundable model — the difference between "we grow 15% month over month" and a machine an analyst can audit. Within the full structure of a SaaS financial model, three checks separate the two:

1. Bottom-up revenue. MRR must be built as funnel → new cohorts → retention curve → ARPU, not as last month × growth rate. Top-down growth assumptions are the first thing a diligence analyst deletes.
2. Model-to-grid reconciliation. The retention the model assumes at month 12 should match the retention your grid actually shows at month 12. A 5-point gap between assumed and observed retention is a credibility problem before it is a math problem.
3. NRR by vintage. Flat-to-rising cohort revenue retention across vintages is the strongest durability evidence a pre-Series-B company can show.

A cohort engine also fixes cash timing. Because revenue arrives cohort by cohort, gross margin, payback, and operating cash flow inherit the correct timing — blended models routinely misplace cash needs by a quarter or more, which is an expensive thing to discover mid-raise.

Next steps

1. Export your billing history and build the grid: rows = first-paid month, columns = months since, cells = share still paying.
2. Fit your curve — M1, M3, M6, M12 plus a steady-state tail — and compare it to the benchmark table above. Under six months of data, borrow the shape and anchor it to your observed months 1–3.
3. Rebuild your revenue forecast as inflow × retention curve × ARPU-by-age, and retire the blended churn rate from forecasting duty.
4. Run the ±5-point month-1 scenario so you know your year-2 ARR spread before an investor asks.

CashQuil runs this entire pipeline natively — traffic channels through funnel conversion into per-cohort retention curves, MRR, and net revenue, with scenario and sensitivity analysis built in and full XLSX export, on a 3-day free trial.

Start free →

---