Banking Customer Churn Prediction
“Which banking customers will close their accounts in the next 90 days?”
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 banking customers will close their accounts in the next 90 days?
Retail banks lose 10-15% of depositors annually, with each lost household representing $3,000-$8,000 in lifetime revenue. By the time a customer calls to close an account, it is already too late. Most churn models rely on static snapshots of balance and tenure, missing the behavioral signals buried in transaction graphs: reduced direct-deposit frequency, declining debit-card spend, new ACH transfers to competitor accounts, and fading branch or app engagement. A top-20 US bank estimated $150M in annual deposit runoff from customers it could have saved with 30 days of lead time.
Quick answer
The best way to predict banking churn is to combine transaction-level behavioral signals (declining spend frequency, new transfers to competitors, reduced app logins) with account and product data using a graph-based ML model. Flat-table approaches miss the cross-table patterns that actually precede account closure. On the RelBench benchmark, relational models score 76.71 vs 62.44 for single-table baselines.
Approaches compared
4 ways to solve this problem
1. Rules-based triggers
Flag customers when balance drops below a threshold or login frequency falls to zero.
Best for
Simple, fast to deploy; good for obvious cases like zero-balance accounts.
Watch out for
Catches only the most extreme cases. By the time balance hits zero, the customer is already gone.
2. XGBoost on a flat feature table
Build a wide table with aggregated features (avg balance, login count, tenure) and train a gradient-boosted classifier.
Best for
Solid baseline that works when you have a good feature engineering team and moderate data complexity.
Watch out for
Requires months of manual feature work. Misses sequential patterns like 'declining weekly grocery spend followed by new ACH to a competitor.'
3. Graph analytics (PageRank, community detection)
Build a transaction graph and compute centrality or clustering features, then feed them into a downstream model.
Best for
Adds network-level signals that flat tables cannot capture, like how connected a customer is to other churners.
Watch out for
Graph features are static snapshots. You lose the temporal dimension: when connections changed matters as much as what they are.
4. KumoRFM (relational graph ML)
Point Kumo at your raw tables (customers, accounts, transactions, digital events) and write a two-line PQL query. The GNN learns temporal, cross-table patterns automatically.
Best for
Highest accuracy with minimal feature engineering. Captures multi-hop, time-aware signals like 'declining debit spend at habitual merchants + new competitor ACH + reduced app logins.'
Watch out for
Requires relational data in normalized tables. Not the right tool if all you have is a single pre-aggregated CSV.
Key metric: RelBench benchmark: relational models score 76.71 vs 62.44 for single-table baselines on customer churn prediction tasks.
Why relational data changes the answer
Churn is not a single-table event. The signals live across transactions (declining spend frequency at habitual merchants), digital events (fewer app logins, rate-comparison page views), account balances (drawdown velocity), and product holdings (closing secondary accounts first). A flat feature table forces you to pre-aggregate all of this into static columns, destroying the temporal sequences that actually predict departure.
Relational models read these tables as a connected graph and learn patterns like 'customer who used to spend at Whole Foods every Tuesday stopped 3 weeks ago, started ACH transfers to Marcus, and viewed competitor rates twice.' On the RelBench benchmark, this multi-table approach scores 76.71 vs 62.44 for single-table baselines. That gap is not a rounding error. It is the difference between catching a churner 60 days out and catching them the day they call to close.
Predicting churn from a balance snapshot is like diagnosing a patient from a single blood-pressure reading. You might catch the extreme cases, but you will miss the patient whose blood pressure has been slowly climbing for six months, who just stopped taking their medication, and who canceled their next appointment. The medical chart (the relational graph) tells a story the single number never can.
How KumoRFM solves this
Relational intelligence built for banking and financial data
Kumo connects accounts, transactions, customer profiles, product holdings, branch interactions, and digital-channel events into a single relational graph. Instead of hand-engineering 200+ features, you write a two-line PQL query. Kumo's graph neural network learns patterns like declining transaction frequency at merchants where a customer used to spend weekly, new recurring transfers to Ally or Marcus, and reduced mobile-app logins. These cross-table signals surface churn risk 60-90 days before closure, giving retention teams time to intervene with targeted offers.
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
CUSTOMERS
| customer_id | name | segment | tenure_years | region |
|---|---|---|---|---|
| C-10042 | James Whitfield | Mass Affluent | 8.3 | Northeast |
| C-10078 | Maria Gonzalez | Premier | 12.1 | Southeast |
| C-10115 | David Park | Mass Market | 2.7 | West |
ACCOUNTS
| account_id | customer_id | type | balance | opened_date |
|---|---|---|---|---|
| A-50001 | C-10042 | Checking | $45,230 | 2017-04-12 |
| A-50002 | C-10042 | Savings | $112,500 | 2017-04-12 |
| A-50003 | C-10078 | Checking | $28,900 | 2013-01-08 |
TRANSACTIONS
| txn_id | account_id | type | amount | merchant | timestamp |
|---|---|---|---|---|---|
| T-900001 | A-50001 | debit | $127.40 | Whole Foods | 2025-09-01 |
| T-900002 | A-50001 | ACH_out | $5,000 | Marcus Savings | 2025-09-03 |
| T-900003 | A-50003 | direct_dep | $4,200 | Employer | 2025-09-15 |
DIGITAL_EVENTS
| event_id | customer_id | channel | action | timestamp |
|---|---|---|---|---|
| E-001 | C-10042 | mobile_app | login | 2025-09-01 |
| E-002 | C-10042 | mobile_app | view_rates | 2025-09-02 |
| E-003 | C-10078 | web | bill_pay | 2025-09-14 |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT BOOL(ACCOUNTS.STATUS = 'closed', 0, 90, days) FOR EACH CUSTOMERS.CUSTOMER_ID WHERE ACCOUNTS.TYPE = 'Checking'
Prediction output
Every entity gets a score, updated continuously
| CUSTOMER_ID | NAME | SEGMENT | CHURN_PROB | RISK_TIER |
|---|---|---|---|---|
| C-10042 | James Whitfield | Mass Affluent | 0.87 | Critical |
| C-10078 | Maria Gonzalez | Premier | 0.12 | Low |
| C-10115 | David Park | Mass Market | 0.64 | High |
Understand why
Every prediction includes feature attributions — no black boxes
Customer C-10042 (James Whitfield)
Predicted: 87% probability of account closure within 90 days
Top contributing features
New ACH transfers to competitor bank
3 in 30d
34% attribution
Declining debit-card transaction frequency
-62%
22% attribution
Mobile app login decline
-78%
18% attribution
Balance drawdown velocity
-$12K/mo
15% attribution
Rate comparison page views
4 visits
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 banking customer churn prediction
What is the best ML model for bank churn prediction?
Graph neural networks that operate on relational data (transactions, accounts, digital events) outperform flat-table models like XGBoost by a wide margin. On the RelBench benchmark, relational models score 76.71 vs 62.44 for single-table approaches. The key advantage is capturing temporal, cross-table behavioral sequences rather than static aggregated features.
How early can you predict customer churn in banking?
With relational ML, reliable churn signals appear 60-90 days before account closure. The earliest indicators are subtle: declining transaction frequency at habitual merchants, new ACH transfers to competitor institutions, and reduced digital engagement. Traditional models that rely on balance thresholds typically detect churn only 10-15 days out.
What data do you need for a banking churn model?
At minimum: customer profiles, account details, and transaction history. For best results, add digital engagement events (app logins, page views), product holdings, and branch interaction data. The power comes from connecting these tables, not from any single source. More tables with clear foreign keys means more signal for the model.
How do you reduce false positives in churn prediction?
Most false positives come from models that react to single signals (e.g., a balance dip) without context. Relational models reduce false positives by requiring convergence across multiple behavioral dimensions: a balance dip alone is not alarming, but a balance dip plus new competitor transfers plus declining app usage is a strong composite signal.
What is the ROI of a banking churn prediction model?
A top-20 US bank with 10M retail customers losing 10-15% annually can expect $30-50M in recovered deposit runoff by retaining even 20-30% of at-risk customers identified 60+ days early. The model cost is a fraction of one retained Premier household's lifetime value ($50K-$100K).
Bottom line: Identify at-risk depositors 60-90 days before closure and retain 20-30% with targeted interventions, saving $30-50M in annual deposit runoff for a top-20 bank.
Related use cases
Explore more financial services 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.
