Predictive Revenue Analytics for SaaS

Predictive revenue analytics for SaaS starts with a simple observation: billing data already contains the signals that predict revenue changes 30–90 days before they show up in a P&L. The problem isn't that these signals are hidden — they're sitting in every Stripe account, in every subscription event log, in every invoice record. The problem is that most SaaS companies and their investors still forecast revenue from spreadsheets, and spreadsheets don't read billing data. They extrapolate from the last three months. That's not prediction. That's a straight line drawn through a curve.

Why SaaS revenue forecasting from spreadsheets fails

Linear extrapolation misses inflection points

Revenue forecasting from a three- or six-month trailing average assumes the next quarter will look like the last one. That assumption holds in exactly the conditions where a forecast isn't interesting — stable growth, stable churn, stable expansion. The moment any of those shift, the linear model breaks. And the shift is the thing you actually need to predict.

An inflection point has a signature in billing data weeks before it appears in top-line MRR. Failed payments accelerate, trial conversions slow, expansion deals stall, or contraction events cluster in the same cohort. Each of these signals is small in isolation. Together, they trace a direction change that linear extrapolation will miss until the quarterly close.

The math is straightforward. If a company adds $50K in new MRR per month and churns $15K, the linear forecast says $35K net growth next month. But if failed payment velocity doubled last month, if two annual contracts signaled contraction at renewal, and if trial-to-paid conversion dropped 200 basis points — the actual net growth next month is likely $18K–$22K. The spreadsheet won't catch this until the revenue actually declines. Billing data caught it 45 days ago.

Self-reported projections embed optimism bias

CFO projections at portfolio companies are systematically optimistic. This isn't a criticism — it's a structural feature of how SaaS companies budget. The annual plan sets a growth target. The quarterly forecast bends toward it. When the board asks for a 90-day outlook, the CFO projects from the plan, adjusting for known losses but rarely incorporating the leading signals that haven't yet materialized as revenue events.

The gap between CFO projections and billing-data actuals widens predictably with time horizon. At 30 days, the divergence is 3–5% — manageable and within forecast tolerance. At 60 days, it grows to 7–12%. At 90 days, CFO projections diverge from billing-data-derived forecasts by 12–22%. That gap is nearly all one-directional: projections run high. The billing data doesn't have an incentive structure. It records what happened, and when you read the patterns correctly, what is about to happen.

Six billing-data signals that predict revenue changes

Billing data isn't one signal. It's a taxonomy of subscription events, each with a different lead time, a different predictive weight, and a different revenue implication. The six signals below cover the full spectrum of MRR movement — involuntary loss, voluntary loss, contraction, expansion, conversion, and structural risk. Each one is computable from Stripe event data without any product telemetry, CRM integration, or self-reported pipeline.

Failed payment velocity (30-day involuntary churn predictor)

Failed payment velocity is the rate of change in payment failures over a trailing 14-day window, expressed as a percentage of active subscriptions. Not the failure count — the acceleration. A company with 1,000 customers and 20 failed payments per month is at steady state. The same company with 20 failures in week one and 35 in week two has a velocity problem. That acceleration predicts a spike in involuntary churn within 30 days.

The signal compounds with recovery rate data. If the same acceleration coincides with a drop in retry success (the percentage of failed payments recovered within the dunning window), the involuntary churn forecast steepens. Failed payment velocity alone explains 40–55% of variance in next-month involuntary churn — more than any other single billing signal.

Cohort revenue decay rate (60-day retention predictor)

Cohort decay is the rate at which a signup cohort's MRR declines over time, measured month-over-month within each cohort. A healthy SaaS business shows cohort decay that flattens: steep early losses (month 1–3), then stabilization. The predictive signal is when recent cohorts decay faster than historical ones.

If the Q1 cohort retained 82% of MRR at month three, and the Q2 cohort retained only 74% at month three, the trajectory predicts lower net retention 60 days out. The signal is stronger than aggregate churn rate because it isolates the behavioral shift to specific customer vintages rather than averaging across the entire base. When two consecutive cohorts show accelerating decay, the model treats it as a systematic shift, not noise.

Expansion timing patterns (90-day growth predictor)

Expansion revenue in SaaS follows timing patterns that are more regular than most teams realize. Usage-based upgrades cluster around billing cycle boundaries. Seat-based expansions correlate with customer headcount changes that are themselves seasonal. Annual plan upgrades concentrate in Q4 and Q1.

The predictive signal is the deviation from established timing patterns. If a company historically sees 60% of its expansion revenue in months 4–7 of a customer's lifecycle, and current-quarter expansion in that window is running 25% below trend, the 90-day growth forecast adjusts downward. The inverse is equally useful: expansion ahead of historical pace predicts a growth acceleration that top-line MRR hasn't captured yet.

Trial-to-paid conversion trend

Trial conversion rate as a snapshot is a lagging indicator — it tells you what happened 14 or 30 days ago, depending on trial length. The conversion trendis a leading indicator. A 200-basis-point decline in trailing 30-day trial conversion predicts lower new MRR 60–90 days out, because today's trials are the pipeline for next quarter's new subscriptions.

The signal is more granular than most reporting surfaces. Conversion by plan tier, by acquisition channel (when tagged in Stripe metadata), and by trial length each carry independent predictive weight. A company whose enterprise trial conversion is stable but whose self-serve conversion is declining has a different revenue trajectory than one where both are falling. The billing data distinguishes these without any CRM dependency.

Contraction clustering

Contraction events — downgrades, seat removals, plan decreases — are individually noisy. A single downgrade might be a customer right-sizing after an initial over-purchase. But when contractions cluster by cohort, by plan tier, or by time window, the signal sharpens. Three downgrades from the same enterprise tier in the same month is a pricing signal, not customer noise.

Contraction clustering predicts net revenue retention 60–90 days out. The model measures the Gini coefficient of contraction events across customer segments — evenly distributed contractions are normal attrition; concentrated contractions in a single segment predict accelerating revenue loss in that segment. The distinction matters for forecasting because segment-concentrated contraction compounds: the first wave triggers more in the same cohort as customers compare notes or respond to the same external pressure.

Payment method distribution

The mix of payment methods across a customer base is a structural risk signal. Credit card subscriptions carry 2–5% baseline involuntary churn from card failures. ACH/bank transfer subscriptions carry 0.5–1%. Invoice/wire subscriptions carry near-zero involuntary churn but higher voluntary churn risk (contract non-renewal).

The predictive signal is the shift in payment method distribution over time. A company migrating customers from cards to ACH reduces its involuntary churn floor. A company whose new signups are disproportionately card-based relative to its existing base will see involuntary churn increase as those cohorts mature. The distribution shift predicts the churn trajectory 90 days out with no reference to individual customer behavior — purely from the structural characteristics of the payment method mix.

From signals to forecasts — building a predictive revenue analytics model from billing data

Individual signals are informative. The forecast is the weighted combination. Each of the six signals has a different lead time, a different confidence interval, and a different MRR component it predicts. Failed payment velocity predicts involuntary churn MRR at 30 days with high confidence. Expansion timing patterns predict growth MRR at 90 days with moderate confidence. The model assigns weights based on historical accuracy within each company's own billing data, then combines them into a composite MRR forecast that decomposes into new, expansion, contraction, and churn.

The weighted average of all six signals catches revenue direction changes 2–3 months ahead of spreadsheet extrapolation. The advantage isn't precision at the dollar level — no model predicts exact MRR to the cent. The advantage is directional accuracy: knowing that revenue growth is decelerating before the deceleration shows up in the numbers. For portfolio operators and investors, directional accuracy at 60–90 days is worth more than dollar precision at 30 days, because it's the difference between a proactive intervention and a retrospective explanation.

The model improves with portfolio-level data. A single company has one billing history, one churn distribution, one expansion pattern. The signals are real but sparse. A portfolio with ten companies has ten billing histories, ten churn distributions, ten expansion patterns. The model learns that failed payment velocity above 1.5x baseline predicts involuntary churn increases not because one company showed it, but because eight of ten showed it. Cross-company signal density is the same structural advantage that makes portfolio churn prediction more accurate than single-company models.

Calibration is ongoing. The model compares its 30-, 60-, and 90-day predictions against actual MRR outcomes each month. Signals whose predictions diverge from actuals get down-weighted. Signals that consistently predict direction changes before they appear get up-weighted. This recalibration loop means the model adapts to each company's specific billing dynamics over time, rather than applying a static formula.

The practical output is a decomposed MRR forecast: predicted new MRR, predicted expansion MRR, predicted contraction MRR, and predicted churn MRR at 30, 60, and 90 days. Each component shows which signals are driving it and how each signal has changed relative to its trailing baseline. A PE firm reviewing a portfolio company doesn't get a single number — they get a transparent model they can interrogate. If the 90-day forecast shows declining net MRR, they can trace it to contraction clustering in the mid-market tier plus declining trial conversion in the self-serve channel. That specificity makes the difference between a forecast and a diagnosis.