mah-trigui/sea-turtle-rescue-forecast-Zindi-Competition

predict the number of sea turtles Local Ocean Conservation will rescue each week in Kenya # Sea Turtle Rescue Forecast Challenge This competition is hosted on Zindi, a machine learning platform for data science challenges. Here is the link to the competition: Sea Turtle Rescue: Forecast Challenge 🌾 - Knowledge Ranked in the TOP 66% --- **Competition:** Zindi — Predict weekly sea turtle rescues per capture site for 2019 **Target:** `Capture_Number` (count per site per week) **Metric:** RMSE **Data:** 29 capture sites, weekly rescue counts from 1998–2018 ## The Problem Local Ocean Conservation rescues sea turtles caught in fishing nets across 29 sites in Kenya. The goal: forecast how many turtles each site will rescue each week in 2019, so staff and budget can be allocated in advance. This is a multi-site time series forecasting problem with strong seasonality, sparse counts (many site-weeks with zero rescues), and a structural data gap in 2016. ## Architecture ``` 00_config.R → Constants, libraries, parameters 01_data_loading.R → Load train.csv, build site×week grids for train/test 02_feature_engineering.R → Site aggregates, week aggregates, frequency, elasticity, clustering 03_time_series.R → Per-site STL+ETS forecasts with expanding-window backtesting 04_model_supervised.R → CatBoost + XGBoost using TS forecasts as features 05_ensemble.R → Weighted blend (70% TS + 30% CatBoost) → submission MAIN.R → Run all steps sequentially ``` ## Key Engineering Decisions ### 1. Hybrid TS + Supervised (Time Series as a Feature) A pure time series model (STL+ETS per site) captures seasonality well but cannot learn cross-site patterns. A pure supervised model can use site features but struggles with temporal structure in sparse count data. The solution: **use the time series forecast as a feature** in the supervised model. The supervised model learns when and how to deviate from the TS baseline. But there's a training problem: if you only generate a TS forecast for 2019, the supervised model has n …