Early Churn Prediction for Mobile Apps

Impala Studios faces a critical challenge: a large share of users install its Weather Radar app but stop using it within the first month, cutting into subscription income and slowing growth. This thesis develops and validates an automated, AWS-native churn-prediction system that flags at-risk users as early as seven days after installation. Leveraging Amplitude event logs, we engineer 20+ week-one behavioral features and train a classifier model, achieving 0.94 AUC and 0.95 precision on held-out data. We embed the model in two SageMaker Pipelines (one for retraining, one for nightly inference) and surface scores via S3/Athena for BI tools and alerts. Offline and live-cohort validations confirm that the system meets all functional and non-functional requirements. Our results pinpoint onboarding friction and lack of early perceived value as the primary churn drivers, guiding targeted product interventions and establishing a reusable framework for all Impala apps.