Renewal Prediction
“Which subscriptions will renew at their next renewal date?”
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
By submitting, you accept the Terms and Privacy Policy.
Loved by data scientists, ML engineers & CXOs at
A real-world example
Which subscriptions will renew at their next renewal date?
Renewal forecasting based on CSM gut feel or simple usage thresholds misses 40% of at-risk accounts. By the time a customer signals non-renewal, it is often too late to intervene. For a SaaS company with $500M ARR, improving renewal prediction accuracy by 15% protects $25M in at-risk revenue.
Quick answer
Renewal prediction forecasts which subscriptions will renew at their next renewal date by learning from usage patterns, support history, seat adoption, and cross-account relationships. The best models identify at-risk accounts 6-8 weeks before the renewal decision, giving customer success teams time to intervene with targeted save plays.
Approaches compared
4 ways to solve this problem
1. CSM Gut Feel / Manual Forecasting
Customer success managers assess renewal likelihood based on their relationship knowledge, recent conversations, and quarterly business reviews. Scores are entered manually into the CRM.
Best for
Small account portfolios (under 50 accounts per CSM) where deep relationship knowledge is the primary signal.
Watch out for
Does not scale. CSMs miss 40% of at-risk accounts because they overweight recent interactions and underweight product usage data. Forecasting accuracy is typically 55-65%.
2. Usage Threshold Models
Set thresholds like 'accounts with under 30% feature adoption are at risk.' Simple to build, easy to explain, and requires no ML infrastructure.
Best for
Products with a single dominant usage metric (API calls, data volume) where the threshold is well-calibrated.
Watch out for
Static thresholds miss compound signals. An account with 80% feature adoption but 3 escalated P1 tickets and a champion who stopped logging in is at higher risk than one with 25% adoption and a growing user base.
3. Traditional ML (Gradient Boosted Trees)
Train a binary classifier on features like usage trends, NPS scores, support ticket counts, and days to renewal. Requires a well-maintained feature pipeline.
Best for
Teams with established ML infrastructure and clean, pre-aggregated account-level features.
Watch out for
Feature engineering is the bottleneck, and the model treats each account in isolation. It cannot see that 3 accounts in the same industry vertical all have declining usage, or that a specific partner integration failing is driving churn across multiple accounts.
4. KumoRFM (Graph Neural Networks on Relational Data)
Connects subscriptions, usage events, support tickets, renewals, and industry/partner relationships into a relational graph. Learns compound renewal signals from the full account ecosystem.
Best for
B2B SaaS companies with multi-table account data, industry relationships, and partner ecosystems.
Watch out for
Requires usage event data with timestamps. If your product does not log user-level events, the graph has fewer signals to work with.
Key metric: Graph-based renewal prediction models achieve 91% accuracy on SAP's SALT benchmark vs 75% for single-table models, giving CS teams 6-8 weeks of lead time on at-risk accounts.
Why relational data changes the answer
Subscription SUB402 has a Pro plan at $3,200 MRR. A flat model sees it uses 2 of 12 features, has 3 of 10 seats active, and has a pending support escalation. That is concerning but not conclusive since many healthy accounts under-adopt features. The relational graph adds critical context: accounts in the same industry with similar adoption profiles renewed at only 45% last quarter. The champion user's login frequency dropped 60% in 30 days. And the support ticket escalation trend is accelerating, not stabilizing.
Renewal decisions are not made in isolation. They are influenced by the vendor's performance across the customer's entire industry cohort, the health of key user relationships, and the trajectory of support interactions. Graph neural networks capture these multi-table, multi-hop signals naturally. A flat model requires an ML engineer to manually compute 'industry renewal rate' and 'champion engagement score' and 'ticket escalation velocity' as separate features. The GNN learns these compound patterns directly from the relational structure, and it discovers patterns that no engineer would think to encode, like 'accounts whose CSM has 3+ other declining accounts are themselves 2x more likely to churn.'
Predicting renewal with flat data is like predicting whether a tenant will renew their lease by only looking at their rent payment history. A relational model also sees that the building's maintenance requests are rising, two neighbors already moved out, and the landlord's response time has doubled. The tenant's own behavior matters, but the building-level signals are what push the prediction from 'maybe' to 'probably not.'
How KumoRFM solves this
Relational intelligence for customer retention
Kumo learns renewal signals from the full relational graph — usage event sequences, support ticket escalation patterns, multi-seat adoption curves, and how renewal behavior propagates through industry and partner networks. Unlike threshold-based models, Kumo captures compound signals that emerge weeks before the renewal decision.
From data to predictions
See the full pipeline in action
Connect your tables, write a PQL query, and get predictions with built-in explainability — all in minutes, not months.
Your data
The relational tables Kumo learns from
SUBSCRIPTIONS
| subscription_id | customer_id | plan | mrr | renewal_date |
|---|---|---|---|---|
| SUB401 | C501 | Enterprise | $12,500 | 2025-04-15 |
| SUB402 | C502 | Pro | $3,200 | 2025-04-01 |
| SUB403 | C503 | Enterprise | $18,000 | 2025-05-10 |
RENEWALS
| renewal_id | subscription_id | amount | timestamp |
|---|---|---|---|
| RN201 | SUB401 | $150,000 | 2024-04-15 |
| RN202 | SUB402 | $38,400 | 2024-04-01 |
| RN203 | SUB403 | $216,000 | 2024-05-10 |
USAGE_EVENTS
| event_id | subscription_id | feature | count | timestamp |
|---|---|---|---|---|
| UE501 | SUB401 | API calls | 12,400 | 2025-03-01 |
| UE502 | SUB402 | Dashboard views | 340 | 2025-03-01 |
| UE503 | SUB403 | Data exports | 890 | 2025-03-02 |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT COUNT(RENEWALS.*, 0, 30, days) > 0 FOR EACH SUBSCRIPTIONS.SUBSCRIPTION_ID
Prediction output
Every entity gets a score, updated continuously
| SUBSCRIPTION_ID | TIMESTAMP | TARGET_PRED | True_PROB |
|---|---|---|---|
| SUB401 | 2025-03-05 | True | 0.91 |
| SUB402 | 2025-03-05 | False | 0.28 |
| SUB403 | 2025-03-05 | True | 0.85 |
Understand why
Every prediction includes feature attributions — no black boxes
Subscription SUB402 — Pro plan ($3,200 MRR)
Predicted: False (28% renewal probability)
Top contributing features
Feature adoption breadth (30d)
2 of 12 features
31% attribution
Active seats vs licensed seats
3 of 10
24% attribution
Support ticket escalation trend
+3 P1 tickets
20% attribution
Similar-industry renewal rate
45%
14% attribution
Champion user login frequency
-60%
11% attribution
Feature attributions are computed automatically for every prediction. No separate tooling required. Learn more about Kumo explainability
PQL Documentation
Learn the Predictive Query Language — SQL-like syntax for defining any prediction task in 2–3 lines.
Python SDK
Integrate Kumo predictions into your pipelines. Train, evaluate, and deploy models programmatically.
Explainability Docs
Understand feature attributions, model evaluation metrics, and how to build trust with stakeholders.
Frequently asked questions
Common questions about renewal prediction
How early can you predict subscription non-renewal?
The most actionable prediction window is 6-8 weeks before the renewal date. This gives customer success teams time to execute a save play: escalate to an executive sponsor, offer a migration plan, or adjust pricing. Predictions made closer than 2 weeks to renewal rarely leave enough time for meaningful intervention.
What is the most important signal for renewal prediction?
No single signal dominates. The compound of usage breadth (features adopted), usage depth (intensity of core features), and relationship health (support tickets, champion engagement) is more predictive than any individual metric. Graph models capture these interactions automatically.
How does seat adoption affect renewal probability?
Seat utilization is one of the strongest renewal signals for multi-seat products. Accounts using fewer than 30% of licensed seats renew at roughly half the rate of fully adopted accounts. But the trajectory matters more than the level: an account growing from 3 to 7 active seats is healthier than one steady at 5.
Can renewal prediction models account for economic conditions?
Yes, when industry and firmographic data is included in the graph. The model learns that renewal rates in a specific industry vertical are declining, and accounts matching that profile inherit a higher risk score even if their individual usage looks healthy.
Bottom line: A SaaS company with $500M ARR that improves renewal prediction accuracy by 15% protects $25M in at-risk revenue — giving CS teams the lead time to intervene 6-8 weeks before the renewal decision.
Related use cases
Explore more retention use cases
Topics covered
One Platform. One Model. Infinite Predictions.
KumoRFM
Relational Foundation Model
Turn structured relational data into predictions in seconds. KumoRFM delivers zero-shot predictions that rival months of traditional data science. No training, feature engineering, or infrastructure required. Just connect your data and start predicting.
For critical use cases, fine-tune KumoRFM on your data using the Kumo platform and Research Agent for 30%+ higher accuracy than traditional models.
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
