On this page
Most "AI churn prediction" requires months of integration, product usage telemetry, and a data science team to train a model on one company's data. That model doesn't transfer. It can't predict churn for a different company, a different vertical, or a different price point. Portfolio-level prediction from billing data is structurally different: it works across companies, requires no integration beyond Stripe, and the cross-company training data makes predictions more accurate than any single company could achieve alone — because the model has seen the same churn pattern play out across hundreds of subscription businesses, not just one.
30–90 days
Advance warning from billing signals
3–8x
More training data in a portfolio vs single company
68–82%
Prediction accuracy from billing data alone
Why portfolio churn prediction is different
Single-company churn prediction is a machine learning problem with a fundamental constraint: limited training data. A SaaS company with 500 customers and 3% monthly churn generates approximately 180 churn events per year. Training a model to predict which customers will churn next month on 180 positive examples is barely viable — the model learns the specific patterns of that one company, at that one price point, in that one market.
Portfolio prediction inverts this constraint. A 10-company portfolio with 5,000 aggregate customers generates 1,800 churn events per year. The model trains on churn patterns across multiple price points, multiple verticals, and multiple customer segments. It learns that a payment failure followed by a downgrade inquiry followed by a support ticket about cancellation is a universal sequence, not a company-specific one. The cross-company training data is the advantage no single company can replicate.
The integration requirement is also fundamentally different. Single-company churn prediction typically requires product usage data — login frequency, feature adoption, session duration — because billing data alone is considered insufficient. That's true for one company because the billing signal is sparse. At the portfolio level, the density of billing signals across companies provides enough statistical power to predict from billing data alone. The model doesn't need to know how often a user logged in last week. It needs to know the payment trajectory, the plan trajectory, and the interaction trajectory with the billing system — all of which are visible in Stripe.
Customer Churn Rate
Percentage of customers who cancel their subscription within a given period.
Billing signals that predict churn
Billing data contains more predictive signal than most teams realize. The stereotype is that billing data is a lagging indicator — it tells you someone churned, not that they're about to. That's true of the churn event itself. The billing signals that precedechurn by 30–90 days are leading indicators, and they're available for every customer without any additional integration.
| Signal | Lead time | Predictive strength | What it indicates |
|---|---|---|---|
| Failed payment (not recovered) | 7–30 days | Very high | Involuntary churn imminent |
| Downgrade | 30–90 days | High | Reduced commitment, value questioning |
| Billing cycle switch (annual → monthly) | 60–120 days | High | Testing exit; wants flexibility to cancel |
| Coupon/discount applied | 30–60 days | Moderate | Price sensitivity; retention offer accepted |
| Invoice dispute or refund | 14–45 days | High | Active dissatisfaction with value received |
| Payment method removal/change | 7–21 days | Moderate | Account maintenance or pre-cancellation cleanup |
| Subscription pause | 30–90 days | Very high | Explicit disengagement; 40–60% never resume |
The power of these signals increases when combined. A customer who downgrades, then fails a payment, then contacts support about cancellation has a churn probability above 85% — the sequence is nearly deterministic. A customer with a single failed payment and no other signals has a 15–25% churn probability, depending on whether the payment is recovered. The model learns these conditional probabilities from the portfolio's aggregate experience, not from a textbook.
Involuntary churn signals: failed payments as the canary
Failed payments are the most predictive single signal because they follow a clear temporal pattern. A first failure is retried automatically. If the retry succeeds, the customer is retained. If the first retry fails, the probability of eventual recovery drops to 40–55%. After a second retry failure, it drops to 15–25%. After the third, it's below 10%.
The prediction model uses this decay curve to assign churn probabilities in real time. A customer whose first retry just failed is "at risk" (45–60% probability of eventual churn). A customer whose second retry failed is "high risk" (75–85%). This granularity enables stage-specific intervention: a first-failure customer gets an automated card update prompt; a second-failure customer gets a personal outreach from the account manager.
The cross-company model advantage
A single-company churn model is trained on one distribution of customers, one pricing structure, and one market. When that market shifts, the model's predictions degrade because it has no reference frame for the new conditions. A portfolio model has seen multiple markets, multiple pricing shifts, and multiple economic cycles across its constituent companies. It's structurally more robust.
The cold-start problem illustrates this. When a company joins a portfolio, a single-company model starts from zero — it needs 6–12 months of churn history before it can make reliable predictions. A portfolio model can predict churn for the new company from day one, because the billing patterns it's learned from other companies transfer. A customer at the new company who fails a payment and downgrades within the same month triggers the same risk assessment the model has validated across the rest of the portfolio.
Segment-level learning is the subtler advantage. A portfolio with companies at different ACV tiers builds an implicit segmentation model. The model learns that churn patterns at $29/mo (high volume, card-failure-driven, price-sensitive) are different from churn patterns at $5,000/mo (low volume, contract-driven, relationship-dependent). When a new company joins at $200/mo, the model already has a prior for that segment's churn behavior — calibrated from real data, not from industry benchmarks.
Seasonal and macro patterns that single companies miss
SaaS churn has seasonal patterns that are invisible at the single-company level but clear at the portfolio level. January churn spikes (new year budget reviews), Q1 and Q3 payment failure increases (card expirations from holiday reissues), and August retention dips (decision-maker vacations) are all real patterns — but a single company with 50 monthly churn events can't distinguish seasonal signal from noise.
A portfolio model sees the seasonal pattern across 500 monthly churn events. It learns that a January churn spike of 15% over baseline is expected, not alarming. It adjusts predictions accordingly — avoiding false alarms in January and maintaining sensitivity in March. This calibration is impossible for a single company unless it has 5–10 years of history, and most SaaS companies are younger than that.
Net Revenue Retention
Revenue retained from existing customers including expansion, contraction, and churn.
What predictions look like in practice
Churn prediction isn't a magic number — it's a risk tier assigned to each customer based on their billing behavior trajectory. Useful prediction divides the customer base into actionable segments: customers who need no intervention, customers who need automated intervention, and customers who need human attention.
A practical risk-tier system works with three levels. Low risk (75%+ of customers): on-time payments, stable plan, no recent changes. These customers are retained by the product, not by intervention. Medium risk (15–20% of customers): one or more billing signals in the past 60 days — a downgrade, a failed payment that was recovered, a discount request. These customers get automated attention: a check-in email, a usage summary, a renewal reminder. High risk (5–10% of customers): multiple correlated signals — failed payment plus downgrade, subscription pause, or explicit cancellation inquiry. These customers need human outreach.
The portfolio view adds a layer that individual companies don't have: the cross-company risk summary. If 3 of 10 portfolio companies show elevated high-risk customer percentages in the same month, the signal is portfolio-level, not company-specific. It might indicate a macroeconomic headwind, a payment processor issue, or a competitive shift. The portfolio operator who sees this pattern can coordinate the response across companies rather than letting each react in isolation.
From prediction to intervention
Prediction without intervention is an interesting dashboard that doesn't save revenue. The value of churn prediction is entirely in the actions it enables — and the right action depends on the type of risk.
Involuntary churn risk (payment failures) calls for mechanical intervention: optimized retry timing, card update prompts, pre-expiration notifications. These are automatable and the ROI is directly measurable — recovered payments are dollars saved. The portfolio operator's role is to ensure every company has this automation in place and to benchmark recovery rates across the portfolio.
Voluntary churn risk (downgrades, disengagement, cancellation signals) calls for relationship intervention. The billing data identifies who is at risk and when the risk appeared. The whyrequires the company's customer success team to investigate. The portfolio model can tell you that a customer who downgraded last month has a 45% probability of churning within 90 days. It cannot tell you whether that customer is frustrated with features, facing budget pressure, or evaluating a competitor.
The intervention economics matter. At median SaaS LTV, saving a $200/mo customer from churn preserves $2,400–$4,800 in lifetime value (12–24 month average lifespan). If the model identifies 20 high-risk customers per company per month, and intervention saves 30% of them, each company preserves $14,400–$28,800 in annual LTV. Across a 10-company portfolio, that's $144K–$288K per year in retained value — from a model that runs on data the portfolio is already collecting.
The feedback loop: intervention outcomes improve the model
Every intervention outcome — customer saved, customer churned despite outreach, customer who wasn't flagged but churned anyway — feeds back into the model. The false-negative rate (customers who churned without being flagged) reveals gaps in the signal set. The false-positive rate (customers flagged as high risk who didn't churn) calibrates the threshold.
This feedback loop is another portfolio advantage. A single-company model gets 15 data points per month (180 annual churn events divided by 12). A portfolio model gets 150. The model converges faster, recalibrates more accurately, and adapts to market shifts more quickly because it has 10x the feedback volume. After 6 months, the portfolio model has seen 900 churn outcomes. A single-company model has seen 90.
North Metric computes churn risk scores from billing data across every connected portfolio company. The model uses the cross-company signal library — payment trajectories, plan changes, billing pattern anomalies — to assign risk tiers updated daily. The portfolio view surfaces which companies have elevated risk concentrations and where intervention is most likely to preserve revenue, giving operators the information they need to act before the churn appears in next quarter's numbers.