Loyalty Program Optimization
“Which loyalty tier will each customer reach 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 loyalty tier will each customer reach in the next 90 days?
Loyalty programs are expensive — a typical retailer spends 2-3% of revenue on rewards. Without predicting tier movement, you over-reward customers who would have stayed anyway and under-reward those on the cusp of upgrading. For a retailer doing $2B in revenue, optimizing tier targeting by just 15% saves $9M in rewards spend while lifting tier-upgrade rates by 20%.
Quick answer
Loyalty program optimization uses multi-class prediction to forecast which tier each customer will reach in the next 90 days. This lets marketing teams invest rewards where they change behavior rather than rewarding customers who would have stayed loyal anyway, turning loyalty programs from a cost center into a growth engine.
Approaches compared
4 ways to solve this problem
1. Points Threshold Rules
Set fixed point thresholds for each tier and project when customers will cross them based on current accumulation rate. The standard approach for most loyalty programs.
Best for
Simple programs with linear point accumulation where spending velocity is stable.
Watch out for
Cannot predict behavioral shifts. A customer who suddenly starts shopping a new category will cross the threshold faster, but rules only extrapolate the current rate.
2. RFM-Based Segmentation
Segment customers by recency, frequency, and monetary value, then map segments to likely tier outcomes. Quick to implement with no ML required.
Best for
Retailers with straightforward tier structures and limited cross-category complexity.
Watch out for
Misses referral network effects, cross-category purchase patterns, and reward redemption behavior that strongly influence tier progression.
3. Multi-Class ML (Random Forest / XGBoost)
Train a multi-class classifier on features like transaction velocity, points balance, and tenure. Predict the probability of each tier outcome.
Best for
Teams with ML infrastructure and a well-maintained feature pipeline. Good accuracy on stable customer segments.
Watch out for
Requires manual feature engineering. Cannot capture how tier movement propagates through household or referral graphs. A customer whose spouse just upgraded to Gold behaves differently, but the flat model does not see this.
4. KumoRFM (Graph Neural Networks on Relational Data)
Connects customers, transactions, rewards, and referral networks into a relational graph. Predicts tier movement as a multi-class classification, learning from cross-category purchase behavior and how tier upgrades propagate through connected customers.
Best for
Retailers with complex loyalty programs, referral bonuses, household linking, or cross-brand point earning.
Watch out for
The graph advantage is largest when customer connections (referrals, households, shared stores) are meaningful. Purely anonymous programs with no relational structure benefit less.
Key metric: SAP SALT benchmark shows 91% accuracy for multi-table relational models on customer behavior prediction vs 75% for single-table approaches, directly applicable to tier movement forecasting.
Why relational data changes the answer
Grace Kim is a Silver member spending $1,240/month across 4 categories, with an 85% reward redemption rate. A flat model might predict she will stay Silver. But the relational graph reveals that 2 of Grace's referred contacts just upgraded to Gold, her transaction frequency trend is up 42% over the past 90 days, and she recently started shopping in the premium electronics category where average basket size is 3x higher.
These cross-entity signals compound. The referral graph shows that customers whose referrals upgrade are 2.8x more likely to upgrade themselves within 60 days. The cross-category expansion signal adds another multiplier. No amount of feature engineering on a single customer-level table captures this compound effect. Graph neural networks propagate information across the customer-transaction-reward-referral graph, discovering these multi-hop patterns automatically. The result is tier predictions that account for the full context of each customer's position in the loyalty ecosystem, not just their individual spending trajectory.
Predicting tier movement with a flat model is like predicting whether a student will make the honor roll by only looking at their test scores. A relational model also sees that they joined a study group (referral network), started taking advanced classes (cross-category expansion), and their classmates in the same study group all improved. The individual metrics matter, but the relational context is what distinguishes an accurate prediction from a guess.
How KumoRFM solves this
Relational intelligence for customer retention
Kumo predicts the loyalty tier each customer will reach as a multi-class classification — learning from transaction velocity, reward redemption patterns, cross-category purchase behavior, and how tier movement propagates through referral and household graphs. This lets marketing invest rewards where they change behavior, not where they reward inertia.
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 | loyalty_tier | signup_date | city |
|---|---|---|---|---|
| C301 | Grace Kim | Silver | 2023-04-10 | Seattle |
| C302 | Hank Morales | Gold | 2022-09-15 | Austin |
| C303 | Ivy Nguyen | Bronze | 2024-01-20 | Denver |
TRANSACTIONS
| txn_id | customer_id | amount | timestamp |
|---|---|---|---|
| T8001 | C301 | $185.00 | 2025-02-25 |
| T8002 | C302 | $420.00 | 2025-03-01 |
| T8003 | C303 | $67.50 | 2025-02-28 |
REWARDS
| reward_id | customer_id | points_earned | tier_at_time | timestamp |
|---|---|---|---|---|
| R601 | C301 | 370 | Silver | 2025-02-25 |
| R602 | C302 | 840 | Gold | 2025-03-01 |
| R603 | C303 | 135 | Bronze | 2025-02-28 |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT CUSTOMERS.LOYALTY_TIER FOR EACH CUSTOMERS.CUSTOMER_ID
Prediction output
Every entity gets a score, updated continuously
| CUSTOMER_ID | TIMESTAMP | PRED_TIER | CONFIDENCE |
|---|---|---|---|
| C301 | 2025-03-05 | Gold | 0.74 |
| C302 | 2025-03-05 | Platinum | 0.61 |
| C303 | 2025-03-05 | Bronze | 0.88 |
Understand why
Every prediction includes feature attributions — no black boxes
Customer C301 — Grace Kim
Predicted: Gold tier (74% confidence)
Top contributing features
Transaction frequency trend (90d)
+42%
30% attribution
Points accumulation rate
1,240/month
25% attribution
Cross-category purchase diversity
4 categories
18% attribution
Referral network tier movement
2 contacts upgraded
16% attribution
Reward redemption rate
85%
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 loyalty program optimization
How do you predict loyalty tier changes with AI?
The most accurate approach uses multi-class classification on relational data. Connect your customers, transactions, rewards, and referral tables into a graph, then predict the probability of each tier outcome (Bronze, Silver, Gold, Platinum) over a 90-day horizon. This captures both individual spending trajectory and network effects from connected customers.
What percentage of loyalty program spend is wasted?
Industry research consistently shows 30-40% of loyalty rewards go to customers who would have maintained their behavior without the incentive. By predicting which customers are on the cusp of upgrading vs. those who are stable, you redirect rewards budget to where it actually changes behavior.
Can you prevent loyalty tier downgrade with predictive models?
Yes. Predicting tier movement works in both directions. Customers predicted to downgrade can receive targeted retention offers (bonus points events, exclusive access) before they lose tier status. This is often more cost-effective than trying to re-engage them after they have already been downgraded.
How do referral networks affect loyalty tier prediction?
Referral networks are one of the strongest signals for tier movement. When a customer's referrals upgrade, the referrer is 2-3x more likely to upgrade themselves within 60 days. This social proof effect is invisible to flat models but is naturally captured by graph neural networks.
Bottom line: A $2B retailer optimizing loyalty tier targeting by 15% saves $9M in rewards spend annually while lifting tier-upgrade rates by 20% — turning the loyalty program from a cost center into a growth engine.
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.
