Customer Lifetime Value
“What is each customer's 3-year lifetime value?”
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
What is each customer's 3-year lifetime value?
Banks allocate relationship-manager time, fee waivers, and retention budgets uniformly across segments, wasting resources on low-value accounts while under-serving high-potential customers. A regional bank with 2M retail customers found that its top 8% of households generated 42% of total revenue, yet these customers received the same service level as accounts generating $200/year. Without accurate forward-looking LTV, banks cannot differentiate treatment, leading to $20-40M in misallocated service costs and lost high-value relationships annually.
Quick answer
The most accurate CLV models for banking connect customer profiles, product holdings, transaction histories, branch interactions, and digital engagement into a relational graph. This approach produces LTV estimates 30-40% more accurate than segment-based averages because it captures forward-looking signals like investment account growth, product consolidation patterns, and engagement trajectory rather than static snapshots.
Approaches compared
4 ways to solve this problem
1. Segment-based averages
Assign each customer a segment (Mass Market, Mass Affluent, Premier) and use the segment's historical average revenue as the LTV estimate.
Best for
Simple and fast. No model required. Works as a starting point for resource allocation decisions.
Watch out for
Treats all Premier customers the same. A Premier customer growing their investment account by $15K/quarter has very different forward value than one whose balances are flat. The average hides the variance.
2. RFM scoring (Recency, Frequency, Monetary)
Score customers on recency of last transaction, frequency of interactions, and monetary value of their accounts. Combine into a composite LTV score.
Best for
Better than segment averages. Captures basic behavioral differences within segments. Well-understood by marketing teams.
Watch out for
Still a static snapshot. Does not capture trajectories: a customer whose frequency is declining rapidly looks the same as one whose frequency just dipped once and recovered.
3. XGBoost on engineered features
Build features from product count, balance trends, transaction frequency, tenure, and channel usage. Train a regression model to predict 3-year revenue.
Best for
Meaningful accuracy improvement over RFM. Handles nonlinear interactions between features.
Watch out for
Manual feature engineering collapses rich temporal data into static aggregates. 'Average quarterly balance growth' loses the information about whether growth is accelerating, decelerating, or spiking from a single event.
4. KumoRFM (relational graph ML)
Connect customer profiles, product holdings, revenue history, engagement events, and transaction data. Write a PQL query to predict 3-year revenue sum. The GNN learns forward-looking value signals from the full relational graph.
Best for
Captures the full trajectory: investment account growing at $15K/quarter + increasing branch advisor visits + mortgage consolidation signals = high and rising LTV. These cross-table trajectories are invisible to flat models.
Watch out for
Requires revenue history data at the customer-quarter level. If revenue attribution is messy (common in shared household accounts), data cleaning is the bottleneck.
Key metric: Relational ML produces CLV estimates 30-40% more accurate than segment-based averages, enabling precise resource allocation across customer tiers.
Why relational data changes the answer
Customer value is not a number in a single table. It is the sum of behavior across products (checking, savings, investment, mortgage), channels (branch, app, web), and time (growing, stable, declining). A flat feature table collapses all of this into static aggregates like 'total balance' and 'product count,' destroying the trajectories that differentiate a $47K customer from a $2K customer.
Relational models read the full customer-product-engagement graph and learn sequences like 'investment account growing at $15K/quarter, branch advisor meetings twice per month, mortgage rate inquiry last week, and 12-year tenure with no product closures.' On RelBench, relational approaches score 76.71 vs 62.44 for single-table baselines on prediction tasks. For LTV, that accuracy gap means the difference between allocating a relationship manager to a $47K customer vs wasting that manager's time on a $2K account.
Estimating customer value from segment averages is like valuing a house by its neighborhood average. Two houses on the same street can differ by $500K based on renovation history, lot size, and recent upgrades. The appraisal (relational model) inspects the house itself. The Zillow estimate (segment average) just looks at the zip code.
How KumoRFM solves this
Relational intelligence built for banking and financial data
Kumo connects customer profiles, product holdings, transaction histories, branch interactions, digital engagement, and life-event signals into a relational graph. The model predicts that Customer C-10078 will generate $47,200 over the next 3 years because her investment account is growing, she is adding direct-deposit payroll, and her branch-visit frequency suggests she will consolidate a competitor mortgage. These cross-table signals produce LTV estimates 30-40% more accurate than segment-based averages.
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 | segment | tenure_years | products_held | relationship_mgr |
|---|---|---|---|---|
| C-10078 | Premier | 12.1 | 6 | RM-042 |
| C-10042 | Mass Affluent | 4.2 | 3 | RM-018 |
| C-10115 | Mass Market | 1.8 | 1 | None |
REVENUE_HISTORY
| customer_id | quarter | fee_revenue | interest_revenue | total |
|---|---|---|---|---|
| C-10078 | 2025-Q2 | $1,240 | $2,800 | $4,040 |
| C-10042 | 2025-Q2 | $320 | $890 | $1,210 |
| C-10115 | 2025-Q2 | $45 | $120 | $165 |
PRODUCT_HOLDINGS
| customer_id | product | balance | monthly_activity |
|---|---|---|---|
| C-10078 | Investment Account | $340,000 | 12 trades |
| C-10078 | Mortgage | $280,000 | 1 payment |
| C-10042 | Checking | $12,300 | 45 txns |
ENGAGEMENT_EVENTS
| customer_id | channel | event_type | frequency_30d |
|---|---|---|---|
| C-10078 | Branch | advisor_meeting | 2 |
| C-10078 | Mobile | investment_review | 8 |
| C-10042 | Mobile | balance_check | 22 |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT SUM(REVENUE_HISTORY.TOTAL, 0, 36, months) FOR EACH CUSTOMERS.CUSTOMER_ID
Prediction output
Every entity gets a score, updated continuously
| CUSTOMER_ID | SEGMENT | PREDICTED_3YR_LTV | CURRENT_ANNUAL | LTV_TIER |
|---|---|---|---|---|
| C-10078 | Premier | $47,200 | $16,160 | Platinum |
| C-10042 | Mass Affluent | $18,400 | $4,840 | Gold |
| C-10115 | Mass Market | $2,100 | $660 | Standard |
Understand why
Every prediction includes feature attributions — no black boxes
Customer C-10078 (Maria Gonzalez)
Predicted: $47,200 predicted 3-year lifetime value
Top contributing features
Investment account growth trajectory
+$15K/qtr
28% attribution
Product depth (6 products held)
Top 5%
23% attribution
Branch advisor meeting frequency
2x/month
19% attribution
Mortgage consolidation signals
Rate inquiry
17% attribution
Tenure and relationship stability
12.1 years
13% 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 customer lifetime value
How do you calculate customer lifetime value in banking?
The most accurate approach uses relational ML to predict forward-looking revenue by connecting customer profiles, product holdings, transaction history, and engagement data. This captures growth trajectories (investment account growing, product consolidation signals) rather than just current balances. The result is a 3-year revenue prediction per customer, 30-40% more accurate than segment-based averages.
What is the difference between CLV and current revenue?
Current annual revenue is a backward-looking measure. CLV predicts forward value, which can be very different. A Mass Affluent customer currently generating $4,800/year but growing savings rapidly, visiting branches for advisor meetings, and showing mortgage-readiness signals may be worth $18,400 over 3 years. A Premier customer generating $16K/year but showing disengagement signals may be worth less than projected.
How do banks use customer lifetime value predictions?
Three primary uses: (1) allocate relationship-manager time to the highest-value and highest-growth customers, (2) size retention offers proportionally to the value at risk when a high-LTV customer shows churn signals, and (3) prioritize acquisition channels that attract high-LTV customers rather than volume.
What data do you need for a banking CLV model?
Customer profiles, product holdings with balances, revenue history (fee + interest income per quarter), transaction data, and engagement events (branch visits, app logins, advisor meetings). The more tables you connect with clear foreign keys, the more forward-looking signals the model can learn.
How accurate are ML-based CLV predictions?
Relational ML models produce CLV estimates 30-40% more accurate than segment-based averages, measured by mean absolute error on held-out data. On RelBench benchmarks, relational models score 76.71 vs 62.44 for single-table approaches. The accuracy matters because it drives resource allocation: every dollar of RM time spent on a low-value customer is a dollar not spent on a high-value one.
Bottom line: Allocate relationship-manager time and retention budgets to the top 8% of customers who drive 42% of revenue, recovering $20-40M in annual misallocated service costs.
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.
