TikTok Retention Dashboard
ABOUT THIS PROJECT
End-to-end analytics project analysing user retention, churn behaviour, and feature impact across a synthetic TikTok EMEA user base of 50,000 users spanning January to December 2024.
User segments are modelled with weighted sampling (Casual 40%, Lurker 25%, Creator 20%, Power User 15%), churn timing via geometric distribution, and session counts via Poisson. Features a cohort retention heatmap, Kaplan-Meier-style survival curves per segment, logistic regression churn predictor with ROC curve, and a Streamlit dashboard with KPI cards, feature impact analysis, and sidebar filters.
WHAT I LEARNED
Geometric distribution is the right model for churn timing — each day is a Bernoulli trial, and the math produces survival curves that actually match how real platform users behave
Poisson distribution for session counts felt obvious in hindsight, but getting the segment-conditional parameters right took more iteration than expected
Cohort heatmaps tell a story that line charts can't — seeing retention decay visually across 12 monthly cohorts makes the seasonal patterns immediately obvious
Logistic regression as a churn predictor is underrated for explainability — the feature coefficients directly answer which product features actually matter for retention
Building synthetic data that's statistically believable is harder than it sounds — if the distributions don't reflect real platform behaviour, the analysis conclusions mean nothing
TECH STACK
LIVE PREVIEW
Having trouble viewing? Open in new tab