Telco Churn Study
500 samples. Same baseline model. Same parameters. The difference is synthetic data.
12% → 72% detection56% → 90% campaign match
- Telecom
- Churn
- ML + SLM
The dataset had more than 7,000 samples.
We used 500 on purpose. In production you often do not have the luxury of a large labelled set. The point was not more data. It was the right approach.
Fifty of those 500 had churned. That is 10 percent — the classic imbalanced set.
We took a different path.
The usual answer is to wait a year and collect more.
Same model, different data.
The platform learned the patterns, the distributions, and the relationships in those 500 samples.
It then generated 10,000 new rows that held those patterns. No real row was copied.
Before
12 / 100
detected
After
72 / 100
detected
A domain-specific churn agent.
Finding the customer is half the work. What do you say to them?
We fine-tuned a small language model on synthetic examples to match the offer to the reason.
Before
56%
campaign match
After
90%
campaign match
Half the time the offer was wrong. A network complaint drew a loyalty discount.
| Reason | RealMismatch | SyntheticMatch |
|---|---|---|
| Slow connection | Loyalty discount | Speed upgrade |
| Data limits | VIP support | Unlimited data |
| Competitor offer | Network guarantee | Price match |
| Product issues | Family bundle | Tech support plus |
| Family needs | Speed upgrade | Family bundle |
| Service quality | Loyalty discount | Premium SLA |
| Billing disputes | Network guarantee | Account credit |
| Support problems | Speed upgrade | VIP support |
| Price concerns | Tech support | Loyalty discount |
| Network issues | Family bundle | Network guarantee |
Time to a result.
1 week
This study
Traditional
- 6–12 months of data collection
- Legal and compliance review
- Privacy risk assessment
- Data security overhead
- Hope the patterns do not move
This study
- 500 labelled samples, no extra collection year
- Privacy review stays with the customer
- No real customer row was copied
- Same baseline model, same parameters
- Result in 1 week
Results summary.
| Scenario | Detection | Campaign match |
|---|---|---|
| Baseline (original data) | 12% | 56% |
| Synthetic ML only | 72% | 56% |
| Full synthetic (ML + SLM) | 72% | 90% |
Beyond churn.
Same method. Different jobs.
- 01
Upsell prediction
Who buys?
What to offer?
- 02
Fraud detection
Which transaction?
How to explain?
- 03
Customer segmentation
What groups?
How to engage?
- 04
Chatbot training
What context?
How to respond?
See it work.
Start free. Ask a question, or generate the data you need.