Cross-Sell Optimization
“Which customers should receive a mortgage offer?”
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 customers should receive a mortgage offer?
Banks spend $50-100 per direct-mail mortgage offer, yet conversion rates on untargeted campaigns hover at 0.5-1.0%. That means $5M-$10M spent on 100K mailers to acquire 500-1,000 loans. Worse, irrelevant offers erode customer trust. The signals for mortgage readiness are scattered across transaction data (rent payments, home-improvement spend, real-estate site visits), life-event indicators (income growth, new savings patterns), and product holdings (checking maturity, investment growth). No single table contains the answer.
Quick answer
The most effective cross-sell models connect transaction data (rent payments, home-improvement spend), savings trajectories, digital behavior, and product holdings into a unified view. Relational ML models that read these connected tables produce 3-5x higher conversion rates than rule-based segmentation. On the SAP SALT benchmark, graph-based approaches achieve 91% accuracy vs 75% for XGBoost on flat tables.
Approaches compared
4 ways to solve this problem
1. Rule-based segmentation
Target customers based on static rules: age 28-40, income above $75K, no existing mortgage, tenure above 2 years.
Best for
Fast to implement. No model needed. Works as a starting filter for campaign teams already running offers.
Watch out for
Conversion rates hover at 0.5-1.0%. You are mailing 100K offers to get 500 loans, wasting $5M+ in campaign costs and annoying 99K customers with irrelevant mail.
2. XGBoost on flat feature table
Engineer features like savings balance growth, income bracket, product count, and tenure, then train a gradient-boosted classifier.
Best for
Solid lift over rules. Can 2x conversion rates with good feature engineering. Widely understood by marketing analytics teams.
Watch out for
Misses the behavioral sequences that signal intent: Zillow transactions followed by Home Depot spend followed by savings acceleration is a mortgage-readiness signal that no single aggregated column captures.
3. Collaborative filtering
Recommend products based on what similar customers adopted: 'customers like you also got a mortgage.'
Best for
Works well for digital product recommendations with high-volume interaction data.
Watch out for
Ignores individual behavioral signals. Two customers in the same segment can have completely different product readiness based on their recent transaction patterns.
4. KumoRFM (relational graph ML)
Connect customers, transactions, balances, product holdings, and digital events. Write a PQL query to predict mortgage adoption. The GNN learns cross-table intent signals automatically.
Best for
Captures the full behavioral story: real-estate service transactions + savings account acceleration + home-improvement spend spike = high mortgage propensity. Lifts conversion 3-5x while reducing offer volume by 60%.
Watch out for
Requires multiple normalized tables with foreign keys. Best results need at least 6 months of transaction history.
Key metric: SAP SALT benchmark: relational graph ML achieves 91% accuracy vs 75% for XGBoost on flat tables in product propensity tasks.
Why relational data changes the answer
Mortgage readiness is not stored in any single table. The signals are distributed: transaction data shows rent payments and real-estate site visits, account balances show savings acceleration, product holdings show which products the customer already has, and digital events show rate-comparison page views. A flat feature table can capture some of these as aggregated columns, but it loses the temporal sequencing that distinguishes browsing from buying.
Relational models read the connected graph and learn patterns like 'savings growing at $2,700/month for 3 months, 3 Zillow transactions in the last 60 days, $4,300 at Home Depot in the last quarter, and direct deposits up 12% year-over-year.' On the SAP SALT benchmark, relational approaches achieve 91% accuracy vs 75% for XGBoost on flat tables. That accuracy gap means 3-5x higher conversion rates and 60% fewer wasted offers.
Sending mortgage offers based on demographics alone is like a car dealer mailing everyone in a zip code. What you actually want to know is who has been visiting dealership websites, whose lease is ending in 3 months, and whose family just grew. That is all in the transaction and event data. Relational ML reads the shopping behavior, not just the profile.
How KumoRFM solves this
Relational intelligence built for banking and financial data
Kumo builds a relational graph connecting customers to their transactions, account balances, product holdings, digital behavior, and demographic data. The model learns that Customer C-10042 has rising direct-deposit amounts, increasing savings-account balances, recurring Zillow transactions, and home-improvement spend at Home Depot. These cross-table signals produce a mortgage-propensity score far more accurate than rule-based targeting, lifting conversion rates 3-5x while reducing offer volume by 60%.
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 | age | segment | products_held | tenure_years |
|---|---|---|---|---|
| C-10042 | 34 | Mass Affluent | 3 | 4.2 |
| C-10078 | 52 | Premier | 6 | 12.1 |
| C-10115 | 28 | Mass Market | 1 | 1.8 |
TRANSACTIONS
| txn_id | customer_id | merchant_category | amount | timestamp |
|---|---|---|---|---|
| T-001 | C-10042 | Real Estate Services | $350 | 2025-08-12 |
| T-002 | C-10042 | Home Improvement | $2,140 | 2025-08-20 |
| T-003 | C-10078 | Grocery | $187 | 2025-09-01 |
ACCOUNT_BALANCES
| customer_id | account_type | balance | 3mo_trend | snapshot_date |
|---|---|---|---|---|
| C-10042 | Savings | $67,400 | +$8,200 | 2025-09-01 |
| C-10042 | Checking | $12,300 | +$1,100 | 2025-09-01 |
| C-10078 | Investment | $340,000 | +$15,000 | 2025-09-01 |
PRODUCT_HOLDINGS
| customer_id | product | status | open_date |
|---|---|---|---|
| C-10042 | Checking | Active | 2021-06-15 |
| C-10042 | Savings | Active | 2021-06-15 |
| C-10042 | Credit Card | Active | 2022-01-10 |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT BOOL(PRODUCT_HOLDINGS.PRODUCT = 'Mortgage', 0, 6, months) FOR EACH CUSTOMERS.CUSTOMER_ID WHERE PRODUCT_HOLDINGS.PRODUCT != 'Mortgage'
Prediction output
Every entity gets a score, updated continuously
| CUSTOMER_ID | NAME | MORTGAGE_PROPENSITY | RANK | RECOMMENDED_ACTION |
|---|---|---|---|---|
| C-10042 | James Whitfield | 0.74 | 1 | Priority Outreach |
| C-10115 | David Park | 0.31 | 2 | Nurture Campaign |
| C-10078 | Maria Gonzalez | 0.08 | 3 | Suppress |
Understand why
Every prediction includes feature attributions — no black boxes
Customer C-10042 (James Whitfield)
Predicted: 74% mortgage propensity in next 6 months
Top contributing features
Real estate service transactions
3 in 60d
28% attribution
Savings account growth trajectory
+$8.2K/3mo
25% attribution
Home improvement spend spike
$4,300 in 90d
21% attribution
Income growth (direct deposit trend)
+12%
15% attribution
Age and tenure fit (34, 4.2 yrs)
High
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 cross-sell optimization
How do you build a cross-sell model for banking?
Start with product holdings (what each customer already has) and transaction data (behavioral signals for what they need next). The best models connect these with savings trajectories, digital engagement, and life-event indicators. Relational ML automates the feature discovery across these tables, finding patterns like savings acceleration + real-estate transactions that signal mortgage readiness.
What is a good conversion rate for banking cross-sell?
Untargeted campaigns typically convert at 0.5-1.0%. Well-built propensity models using flat-table ML lift this to 1.5-2.5%. Relational ML models that capture cross-table behavioral signals achieve 3-5% conversion rates, a 3-5x improvement over rules-based targeting.
What data do you need for a product propensity model?
At minimum: customer profiles, product holdings, and transaction history. For strongest results, add account balance trajectories, digital engagement events (page views, rate comparisons), and income signals (direct deposit trends). The power is in connecting these tables, not in any single data source.
How do you avoid annoying customers with irrelevant offers?
High-precision propensity scoring is the answer. By targeting only the top-scored 20-30% of customers instead of a broad segment, you reduce offer volume by 60% while increasing conversion 3-5x. The customers you contact are the ones actually showing intent signals, so the offer feels relevant rather than intrusive.
Can ML predict which banking product a customer needs next?
Yes. Relational ML models predict product adoption by reading cross-table behavioral signals. The model does not just know that a customer is 34 and earns $95K. It knows they started Zillow transactions, accelerated savings, and increased home-improvement spend. That behavioral context predicts mortgage readiness far better than demographics.
Bottom line: Increase mortgage cross-sell conversion from 0.8% to 3.5% while mailing 60% fewer offers, saving $3M in campaign costs and generating $40M+ in new loan originations per quarter.
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.
