Upsell Prediction
“Which subscribers will upgrade their plan?”
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 subscribers will upgrade their plan?
Carriers send 50M+ upgrade offers monthly with 2-3% conversion rates. Each wasted offer costs $0.50-$2.00 in delivery and discounting, totaling $25M-$100M in wasted marketing spend annually. Worse, poorly timed offers train subscribers to wait for discounts. The upgrade signal is in the intersection of usage patterns, network experience, social influence from contacts on higher plans, and historical response behavior.
Quick answer
Effective upsell prediction connects usage patterns, social influence from the call graph, offer response history, and plan comparison behavior in a relational model. Subscribers at 85%+ data utilization whose top contacts are on higher-tier plans, and who recently browsed the carrier app's plan comparison page, convert at 12x the base rate. SAP SALT benchmark shows relational graph ML achieves 91% accuracy vs 75% for XGBoost on flat tables.
Approaches compared
4 ways to solve this problem
1. Usage threshold triggers
Send upgrade offers to subscribers who exceed 80% of their data cap or incur overage charges.
Best for
Catches the most obvious upsell candidates. High intent, easy to implement with billing data alone.
Watch out for
Reaches only 10-15% of potential upgraders. Many subscribers manage their usage carefully and never hit caps, but would upgrade if they knew their friends all had unlimited plans.
2. Campaign response modeling (logistic regression)
Train a model on historical offer acceptance rates using subscriber demographics, tenure, and plan attributes.
Best for
Good for optimizing offer timing and channel. Works when you have enough historical campaign response data to train on.
Watch out for
Does not capture the social influence signal. Subscribers upgrade because their friends did, not because of demographics. Also misses offer fatigue: subscribers shown too many declined offers respond less to future ones.
3. RFM segmentation for upsell
Segment subscribers by recency of engagement, frequency of usage, and monetary value (ARPU), then target high-RFM segments.
Best for
Quick directional segmentation that works without ML infrastructure. Good for initial targeting before building models.
Watch out for
Static segments miss individual timing. A subscriber in the 'medium' segment might be ready to upgrade today based on recent usage spikes and social influence that the segment average hides.
4. KumoRFM (relational graph ML)
Connect subscribers, plans, usage data, offer history, and call-graph contacts into a relational model. The GNN learns social upgrade influence, offer fatigue patterns, and genuine intent signals.
Best for
Highest conversion rate. Captures social influence from the call graph (contacts on higher plans), offer fatigue from response history, and intent signals from app browsing behavior.
Watch out for
Requires offer history and call detail records for full effectiveness. If offer tracking is not in place, start collecting it before deploying.
Key metric: SAP SALT benchmark: relational graph ML achieves 91% accuracy vs 75% for XGBoost on flat tables in plan upgrade propensity tasks.
Why relational data changes the answer
Plan upgrades are socially influenced decisions. The signals live across usage data (consistently approaching plan limits), the call graph (whether a subscriber's frequent contacts are on higher-tier plans), offer history (how many offers they have declined and through which channels), and digital behavior (plan comparison page visits on the carrier app). A flat feature table with 'data_used_pct = 94%' misses the social context entirely: this subscriber's five closest contacts all upgraded to unlimited last quarter.
Relational models connect the usage data to the call graph and learn that subscribers whose contacts upgraded recently convert at 12x the base rate when approached within 30 days. They also learn offer fatigue: subscribers who received 3+ declined offers in 90 days respond 60% less to the next one, making timing as important as targeting. SAP SALT benchmark shows relational graph ML achieves 91% accuracy vs 75% for XGBoost on flat tables for propensity tasks. For a carrier sending 50M offers monthly, that accuracy gap eliminates $25M-$100M in wasted marketing spend.
Targeting upsells from usage data alone is like a car salesman pitching upgrades based on mileage. High mileage means the car is used a lot, but it does not tell you the driver's coworkers all just leased new BMWs, that they browsed the dealer website three times last week, and that the last two mailers went straight to the recycling bin. The call graph, digital behavior, and offer history complete the picture.
How KumoRFM solves this
Graph-learned network intelligence across your entire subscriber base
Kumo connects subscribers, plans, usage patterns, offer history, and response data into a graph where upgrade propensity propagates through communication networks. It learns that subscribers at 85%+ data utilization whose top contacts recently upgraded and who browsed the carrier app's plan comparison page convert at 12x the base rate. The model also learns offer fatigue: subscribers shown 3+ declined offers in 90 days respond 60% less to the next one.
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
SUBSCRIBERS
| subscriber_id | plan | tenure_months | monthly_arpu |
|---|---|---|---|
| SUB201 | Basic 5GB | 14 | $35 |
| SUB202 | Unlimited | 28 | $65 |
| SUB203 | Basic 5GB | 6 | $35 |
PLANS
| plan_id | name | monthly_cost | data_gb | tier |
|---|---|---|---|---|
| PLN01 | Basic 5GB | $35 | 5 | Entry |
| PLN02 | Unlimited | $65 | Unlimited | Mid |
| PLN03 | Unlimited Plus | $75 | Unlimited | Premium |
USAGE
| usage_id | subscriber_id | month | data_gb_used | overage_charges |
|---|---|---|---|---|
| U201 | SUB201 | 2025-02 | 4.7 | $0 |
| U202 | SUB201 | 2025-01 | 4.9 | $5.00 |
| U203 | SUB203 | 2025-02 | 2.1 | $0 |
OFFERS
| offer_id | subscriber_id | offer_type | sent_date | channel |
|---|---|---|---|---|
| OFF01 | SUB201 | Upgrade to Unlimited | 2025-02-01 | SMS |
| OFF02 | SUB201 | Upgrade to Unlimited | 2025-01-15 | |
| OFF03 | SUB203 | Add hotspot | 2025-02-10 | App push |
RESPONSES
| response_id | offer_id | action | timestamp |
|---|---|---|---|
| RSP01 | OFF01 | Opened | 2025-02-01 |
| RSP02 | OFF02 | Ignored | |
| RSP03 | OFF03 | Clicked | 2025-02-10 |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT BOOL(SUBSCRIBERS.PLAN_UPGRADE, 0, 30, days) FOR EACH SUBSCRIBERS.SUBSCRIBER_ID WHERE SUBSCRIBERS.PLAN != 'Unlimited Plus'
Prediction output
Every entity gets a score, updated continuously
| SUBSCRIBER_ID | CURRENT_PLAN | BEST_OFFER | UPGRADE_PROB_30D |
|---|---|---|---|
| SUB201 | Basic 5GB | Unlimited | 0.72 |
| SUB202 | Unlimited | Unlimited Plus | 0.15 |
| SUB203 | Basic 5GB | Unlimited | 0.08 |
Understand why
Every prediction includes feature attributions — no black boxes
Subscriber SUB201 -- Basic 5GB, 14-month tenure
Predicted: 72% upgrade probability within 30 days
Top contributing features
Data utilization (3-month avg)
94% of plan
30% attribution
Top contacts on higher plans
4 of 5 on Unlimited
22% attribution
Overage charges (last 90d)
$15.00 total
19% attribution
App plan-comparison page visits
3 in last 14d
17% attribution
Offer response history
1 opened, 1 ignored
12% 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 upsell prediction
How do you predict which telecom subscribers will upgrade?
Connect usage data, call-graph contacts, offer response history, and digital engagement into a relational model. The strongest signals are social (contacts on higher plans), behavioral (plan comparison page visits), and utilization-based (data usage above 85%). Graph ML identifies subscribers with genuine upgrade intent vs those who will ignore offers.
What is offer fatigue and how do you avoid it?
Offer fatigue occurs when subscribers receive too many upgrade offers and learn to ignore them, or worse, wait for bigger discounts. Relational models track offer response history and learn that subscribers who declined 3+ offers in 90 days respond 60% less. The model suppresses offers to fatigued subscribers and targets only those with fresh intent signals.
How does social influence drive plan upgrades?
Subscribers whose top-5 calling contacts are on higher-tier plans upgrade at 12x the base rate. This is the single strongest upsell signal in telecom, and it is invisible to models that do not incorporate the call graph. When a subscriber's social circle upgrades, they follow within 30-60 days.
What is the ROI of targeted upsell prediction in telecom?
A 30M-subscriber carrier that improves upsell conversion from 3% to 8% through targeted offers generates $140M in incremental annual ARPU. The model also eliminates $25M-$100M in wasted offer delivery and discounting costs by not sending offers to subscribers who will never convert.
What data do you need for telecom upsell prediction?
At minimum: subscriber profiles, plan details, and usage data. High-impact additions: call detail records (for the social influence graph), offer history with response data (for fatigue modeling), and digital engagement data (app and website plan-comparison page visits). The call graph and offer history provide the biggest lift over basic usage models.
Bottom line: A 30M-subscriber carrier that improves upsell targeting from 3% to 8% conversion generates $140M in incremental annual ARPU. Kumo identifies subscribers whose usage patterns, social network influence, and offer response history signal genuine upgrade intent, eliminating wasted offers that train subscribers to wait for discounts.
Related use cases
Explore more telecom 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.
