Most subscription revenue forecasts are wishful thinking dressed up in a spreadsheet. They start with last month's growth rate, project it forward 24 months, and arrive at a number the founder wants to hear. Three months later the actual revenue is half of the forecast, the spreadsheet gets quietly updated, and everyone pretends it always said that. The honest alternative is a cohort-based model: separate the subscribers you already have (whose future revenue is determined by their retention curve) from the subscribers you'll acquire (whose volume depends on marketing efficiency), then add them up. This guide walks through how to build that model, where to source the inputs, how to fit a retention curve from your actual data, how to run scenarios that bracket the uncertainty, and the specific conditions under which the forecast stops being reliable — usually exactly when you need it most.
Why most subscription forecasts are wrong
The standard subscription forecast is built on three assumptions that don't hold: (1) past growth rate continues, (2) churn is constant, (3) ARPU is stable. None of those are usually true, but together they produce a smooth exponential curve that looks confident. Reality is bumpier — growth rate decays as the easy customers get acquired, churn front-loads in the first 90 days, ARPU drifts as subscribers add and remove items. A forecast that ignores all three produces a number that's directionally optimistic by 20-40% over a 12-month window.
Cohort-based forecasting fixes this by treating each signup month as a separate population with its own retention curve. The current month's revenue is the sum of revenue from each surviving cohort. Future revenue is projected by extrapolating each cohort's curve forward and adding revenue from future cohorts (which you estimate based on acquisition trajectory). It's more work but it's honest — and crucially, it makes the assumptions visible so you can argue about them.
Multiplying last month's MRR by 1.10 every month for 24 months produces a forecast where you do 10× your current revenue. If that's your model, you're not forecasting — you're storytelling. The reality is that growth rate decays predictably; even great subscription businesses go from 20%/mo growth to 5%/mo growth as they scale. A flat-multiplier forecast misses that decay every time.
The cohort model in 10 lines
Forget the spreadsheet for a minute. Conceptually a cohort revenue forecast is straightforward.
- Group historical subscribers by signup month. The January 2025 cohort, the February 2025 cohort, and so on.
- For each cohort, compute the survival rate at month 1, 2, 3, etc. — what % of the original cohort is still subscribed at each age.
- Compute average revenue per surviving subscriber at each cohort age (usually close to ARPU but drifts slightly).
- Each cohort's contribution to a future month's revenue is (cohort_size × survival_at_that_age × ARPU_at_that_age).
- Sum across all cohorts to get total revenue for that future month.
- For future cohorts (not yet acquired), estimate their size from your acquisition trajectory and apply the same survival curve.
- Total forecast = revenue from existing cohorts (extrapolated forward) + revenue from future cohorts (projected from acquisition).
The hard parts are: (1) how do you extrapolate the survival curve past the months you have actual data for, and (2) how do you estimate future acquisition. Both are genuinely uncertain — that's why scenario planning matters more than precision in the central forecast.
Fitting a retention curve from your data
You have actual survival data for the months you've been operating, but you need to project the curve forward. The shape of subscription retention is well-studied: it's almost always best fit by a power-law or shifted-Beta-Geometric (sBG) distribution. Both have the property of front-loading churn (more loss in early months) and asymptoting to a long-term retention floor.
A practical approach: fit a simple model with two parameters. Parameter 1 is the early-churn rate (high in months 1-3, drops fast). Parameter 2 is the long-term floor (what fraction of subscribers retain past month 12 or so). Most consumer subscriptions converge to a floor between 20% and 45% — anything above 50% is exceptional, anything below 20% is a leaky product. Once you fit those two parameters to your actual data, you can extrapolate the curve forward.
- Power-law fit — y = a × x^(-b). Simple, works on most data. Two parameters, fit by log-log regression.
- sBG (shifted Beta-Geometric) — the academically rigorous choice; available as a Python or R library. Two parameters, fit by max-likelihood.
- Eyeball + interpolate — if you only have 4 months of data, formal fitting is overconfident. Eyeball the curve, draw a smooth extrapolation, and treat the long-term floor as an explicit assumption.
- Use industry benchmarks for the floor — if you have less than 12 months of data, anchor the long-term retention floor to category benchmarks (coffee subscriptions: 30-40% floor; supplements: 25-35%; discovery boxes: 15-25%).
If your oldest cohort has 35% retention at month 12 and you have no later data, don't assume month-24 retention is 35% — assume it's 30-32% (continued slow decay) or in the optimistic case 30% with a flat tail. Anyone projecting month-24 retention higher than their best observed month-12 figure is fabricating data.
Forecasting future cohort acquisition
The cohort math handles existing subscribers cleanly. New subscribers are harder because their volume depends on marketing spend, channel saturation, conversion rate, and seasonality — none of which extrapolate linearly. Two practical approaches.
Approach 1: tie acquisition to a spend or input metric. If you can spend $5,000/mo on ads at a $50 CAC, you'll acquire ~100 subscribers. Project spend forward, project CAC efficiency forward, multiply for monthly acquisition. This works when you have one or two dominant channels and you understand their saturation point — CAC tends to rise as you scale spend, so don't assume linear returns.
Approach 2: use last-3-month rolling average plus a growth multiplier. If you acquired 80, 90, 100 subscribers in the last three months, project 110 next month, 120 the one after, etc. — but explicitly assume the growth rate decays. A 25% monthly growth rate from a small base is plausible; the same growth rate sustained for 12 months is not. Apply a decay factor that brings the growth rate down to category-norm (typically 3-8% monthly for sustained subscription growth).
If your forecast shows revenue tripling in 12 months and acquisition tripling along with it, ask: where exactly does that acquisition come from? Which channel? At what CAC? If you can't answer those three questions, the forecast is fiction. Honest forecasts include the channel math, not just the headline number.
Scenarios beat point forecasts
A single-line forecast pretends you know the future. You don't. The honest output of a subscription forecast is three numbers — a base case, an upside, and a downside — bracketing the range of plausible outcomes. The width of the range tells you something important: if base and downside are within 10% of each other, the model is robust; if they're 50% apart, you're operating in high uncertainty and should plan for that.
- Base case — your best-guess retention curve, your best-guess acquisition trajectory, no surprises.
- Upside — retention 10-15% better than observed (a save flow lands well), acquisition 20-30% above plan (a marketing push works).
- Downside — retention 10-15% worse (a churn spike), acquisition 20-30% below plan (channel saturates, ad costs rise), or both.
- Stress test — flat acquisition for 6 months. What happens to MRR? This tells you how dependent your business is on continued growth vs how much existing subscriptions carry you.
The most useful single number from this exercise isn't the base case — it's the downside MRR 6 months out. That tells you your minimum sustainable revenue if everything goes wrong on both sides. If that number doesn't cover your fixed costs, you're more fragile than you think.
When forecast accuracy breaks down
Cohort forecasting works when the underlying business is stable in shape — same product, same acquisition channels, same pricing. It breaks badly under regime change. Some specific situations where you should distrust the forecast entirely:
- You just launched a new product line — retention curves for the new line aren't established. Run the forecast on existing products only and treat the new line as upside.
- You changed pricing — old cohorts' retention may not predict new cohorts' retention. Wait 90 days before incorporating new-pricing cohorts into the model.
- You changed the discount level — same problem; price-sensitivity composition of new cohorts shifts.
- You shifted acquisition channels — TikTok-acquired subscribers retain differently than Facebook-acquired ones. Channel-mix shift breaks the model until you have separate retention curves per channel.
- Macro shock — inflation, recession, supply chain disruption. Retention worsens across the board; your historical curve is now too optimistic by an unknown amount.
- Operational change — moved 3PL, changed packaging, raised shipping prices. Each can move retention 3-8% in either direction with no immediate signal.
Don't treat the forecast as a one-time exercise. Rebuild it every quarter with fresh data, and compare the previous quarter's prediction against actual. If you were off by more than 15%, figure out which assumption broke. Calibration improves with practice — most merchants are too optimistic in their first 2-3 forecasts and gradually converge to honest.
What a usable forecast actually looks like
Pulling it together. A subscription revenue forecast you can actually use in board meetings, investor conversations, or your own planning has these properties.
- Built from cohort math, not from a single growth multiplier
- Retention curves come from your actual data (or, for early-stage, anchored to category benchmarks with explicit assumption notes)
- Future acquisition tied to a spend model or a decaying growth rate, not flat exponential
- Three scenarios (base, upside, downside) — not a single line
- Long-term retention floor is stated explicitly as an assumption
- Stress test included: what happens if acquisition flatlines for 6 months
- Quarterly rebuild cadence, with last quarter's prediction shown alongside this quarter's actuals
- Sensitivity table: a 10% change in churn → what impact on 12-month MRR
- Acknowledges break conditions (pricing change, new channel, macro shock)
- Conservative on the upside; the optimistic case isn't the base case
If your current forecast has fewer than half of these, it's a story, not a model. Adopt the rest one at a time — start with cohort math (the biggest lift), then add scenarios, then add the stress test. Most subscription founders get to a usable forecast in 4-6 weeks of effort, and the upgrade pays for itself the first time it warns you about a downside scenario you'd otherwise have walked into.
Making the forecast operationally useful
A forecast that sits in a spreadsheet and gets opened once a quarter is half-useful. To make it actually drive decisions, connect it to the operating numbers you watch weekly. The bridge is simple: when this week's actual MRR diverges from the forecast trajectory by more than 5%, that's a signal — either the forecast was wrong (rebuild) or something changed in the business (investigate).
The same applies to the cohort-level reading. If the most recent cohort is tracking 10% below the survival curve at month 2, you have an early warning that something about new subscribers is different — could be acquisition-channel mix, could be product, could be a fulfillment hiccup. Catching that at month 2 is much cheaper than catching it at month 6.
Combine the forecast with the rest of your analytics dashboard. The cohort retention table feeds the forecast. The dashboard surfaces deviations. The forecast tells you whether those deviations matter to the 12-month picture or are weekly noise.
Subscription forecasting FAQ
How far out can I forecast subscription revenue?
Cohort-based models stay reasonably accurate to 12 months. Past that, retention-curve extrapolation gets shakier and future-acquisition assumptions dominate the result. For 18-24 months, treat the forecast as a directional planning tool, not a precise number.
Can I forecast with less than 6 months of data?
Yes, but you'll lean on industry benchmarks for the retention floor and acquisition trajectory. State those benchmarks explicitly as assumptions. A 6-month-old store's forecast should have wider scenarios than a 3-year-old store's.
What retention shape should I expect?
Front-loaded churn (highest in months 1-3), then a softer decline, then a long-term floor. Most consumer subscriptions reach the floor between months 9-15. Anything claiming flat retention from day one is either measurement error or a freemium model masquerading as a subscription.
How do I model price increases in the forecast?
Apply the new price to new cohorts immediately and to existing cohorts on their renewal anchor date (assuming grandfathering rules permit). Expect a temporary churn bump on existing cohorts when the increase takes effect — 5-15% additional churn the month of the price change is typical.
Should I forecast in customers or in MRR?
Both. Customers tells you operational load (fulfillment, support). MRR tells you financial. They diverge when ARPU shifts — a forecast that only shows one will mislead you. Most spreadsheets have a cohort table for customer counts and a parallel one for MRR.
How do I handle multi-cadence subscriptions?
If you sell monthly and quarterly together, build separate cohort tables for each cadence. Their retention curves are different — quarterly subscribers churn less per period but each loss is worth 3 months of MRR. Sum the two forecasts.
What's the right way to model expansion revenue?
If your business has expansion (upsells, cross-sells, add-ons), model it as a per-cohort ARPU drift over time. Most cohorts show 5-15% ARPU growth in the first year as add-ons accumulate. Use your actual cohort ARPU trajectories rather than a constant ARPU.
How do I forecast in a hyper-seasonal business?
Build a monthly seasonality multiplier from at least 18 months of history (one full year + the next year's comparison). Apply it to acquisition projections, not to the retention curve directly — retention is usually less seasonal than signups.
What if my churn is irregular — some months 5%, others 12%?
First, check if the volatility is from a single cohort or random noise across cohorts. Cohort-month volatility is normal; multi-cohort-wide volatility means something is shifting (acquisition mix, operations). Smooth with a 3-month rolling average for the forecast input, but investigate the underlying cause.
How do I model the impact of a new save flow on the forecast?
Conservatively. New save flows typically reduce churn by 1-3 percentage points monthly, sometimes 5+ for previously-poor flows. Add the assumed reduction only to cohorts that experience the new flow (not retroactively to old cohorts), and treat the upside scenario as the optimistic version.
Should I share the forecast with investors?
Share the base case with the assumptions explicit, and ideally show the three scenarios. Investors who've seen subscription businesses before will respect honest scenario bracketing more than a single-line forecast that obviously can't be that precise. A confident forecast with hidden assumptions is a red flag.
What's the most common forecasting mistake?
Anchoring retention to the best cohort instead of the average cohort. Your March cohort retained 65% at month 6 — great. But the average cohort retained 52%. The forecast should use the average, not the best. Cherry-picking the best cohort produces forecasts that consistently overshoot reality by 15-20%.