Prediction is the easy part. In a large dataset of SaaS cancellation conversations, login frequency alone predicts churn with ~70% accuracy at 60 days out. Add feature usage depth and support ticket sentiment and you're at ~85%.
The hard part is the next step: what do you do with the prediction? A "this customer is at risk" alert is worthless without understanding why they're at risk. The same risk score means completely different things. One customer needs a feature that exists but they haven't found, another is comparing you to a competitor, another had a payment fail silently.
The SHAP explanations in your approach are a step in the right direction since they surface which features drove the prediction. But behavioral features still only tell you what happened, not why. "Usage dropped 40%" doesn't tell you if the customer is evaluating competitors, hit a bug they never reported, or just went on vacation.
Most churn prediction tools stop at the score and leave the "why" to CSMs who don't have time. The companies with the lowest churn in the dataset automate the "why": structured conversations triggered by behavioral signals, not just alerts.
Hey HN — built this because every churn prediction notebook on GitHub
uses the Kaggle Telco dataset and outputs a confusion matrix that no
founder can act on.
ChurnGuard connects to your Stripe account, engineers behavioral
features from real billing data, trains an XGBoost model on your
actual churned vs retained customers, and uses SHAP to explain
why each customer is at risk in plain English.
The LLM layer (Groq free tier) generates a specific 30-day retention
playbook per at-risk customer — not "schedule a call" but actual
messaging, offer amounts, and who should own each action.
Entire stack is free. Works in Google Colab. Happy to answer
questions about the architecture.
Prediction is the easy part. In a large dataset of SaaS cancellation conversations, login frequency alone predicts churn with ~70% accuracy at 60 days out. Add feature usage depth and support ticket sentiment and you're at ~85%.
The hard part is the next step: what do you do with the prediction? A "this customer is at risk" alert is worthless without understanding why they're at risk. The same risk score means completely different things. One customer needs a feature that exists but they haven't found, another is comparing you to a competitor, another had a payment fail silently.
The SHAP explanations in your approach are a step in the right direction since they surface which features drove the prediction. But behavioral features still only tell you what happened, not why. "Usage dropped 40%" doesn't tell you if the customer is evaluating competitors, hit a bug they never reported, or just went on vacation.
Most churn prediction tools stop at the score and leave the "why" to CSMs who don't have time. The companies with the lowest churn in the dataset automate the "why": structured conversations triggered by behavioral signals, not just alerts.
Took a quick look and I liked the it. I will be using it for my own saas applications.
Thanks for creating this project. Will give it a star.
Hey HN — built this because every churn prediction notebook on GitHub uses the Kaggle Telco dataset and outputs a confusion matrix that no founder can act on.
ChurnGuard connects to your Stripe account, engineers behavioral features from real billing data, trains an XGBoost model on your actual churned vs retained customers, and uses SHAP to explain why each customer is at risk in plain English.
The LLM layer (Groq free tier) generates a specific 30-day retention playbook per at-risk customer — not "schedule a call" but actual messaging, offer amounts, and who should own each action.
Entire stack is free. Works in Google Colab. Happy to answer questions about the architecture.