Loop'n — Crowd-prediction ML across 1000+ attractions
A crowd-prediction system for theme parks, powered by Prophet, covering 1000+ attractions worldwide and updating intra-day based on weather.
- Python
- Pandas
- Facebook Prophet
› prédiction · 24 h · attraction
Context
Loop’n is an app for theme park visitors: real-time information, visit planning, queue predictions. The user-facing promise is simple — “what time should I ride this attraction?” — but it demands a data pipeline that holds at global scale.
The constraints: cover hundreds of parks with very different attendance behaviours, integrate weather data that can flip the picture in a few hours, and stay actionable on the user side (a prediction that arrives too late is worthless).
My role
ML dev. I designed and operated the crowd-prediction system in Python.
What I built
- Prediction model based on Pandas for data wrangling and Facebook Prophet for time-series — a pragmatic choice: Prophet absorbs multiple seasonalities (day of week, school holidays, park events) without heavy tuning
- Coverage: 1000+ attractions worldwide, with models tuned per attraction
- Continuous intra-day updates: current weather, event cancellations and other signals refresh predictions without waiting for the nightly batch
- Integration with park operational data when available, fallback on proxies otherwise
The hard part: Prophet is easy to spin up but hard to industrialise once you multiply models by a thousand. The discipline of monitoring predictions (drift, anomalies) matters more than fine-tuning each model.
Results
The system powers the app in production. 1000+ attractions covered, predictions updated intra-day. Long-running work on prediction observability to spot drift per park.