Sherzadd/turtle_rescue_forcast

Forecasting weekly sea turtle rescues in Kenya using historical data (1998–2018) from Local Ocean Conservation. # 🐢 Sea Turtle Rescue: Forecast Challenge (Zindi) Forecast the weekly number of turtles caught per capture site along the Kenyan coast to help **Local Ocean Conservation (LOC)** plan their staff schedules and budgets more efficiently. --- ## 📘 Overview Since 1998, **Local Ocean Conservation (LOC)** has worked with local fishers to rescue endangered sea turtles caught as bycatch. Each time a turtle is captured, LOC records the event, measures the animal, and releases it safely. This challenge aims to **forecast the number of turtles caught per week for each capture site in 2019**, using historical data from 1998–2018. Accurate forecasts will help LOC: - Plan staff schedules - Manage budgets - Improve the rescue program’s operational efficiency --- ## 📂 Dataset Description | File | Description | Usage | |------|--------------|--------| | **`train.csv`** | Historical turtle rescue data (1998–2018). Each row represents one turtle capture event, including date, capture site, and other variables. | ✅ Main training dataset | | **`Sample_sub.csv`** | Template submission file with all required IDs for 2019 (format: `CaptureSite_ _ `). | 📄 Use as reference for your `submission.csv` | | **`variable_definitions.csv`** | Column descriptions for `train.csv`. | 🔍 Reference only | | **`CaptureSite_category.csv`** | Metadata mapping each site to its category and type (geographic/hidden groupings). | 🧭 Optional — can be merged for extra features | --- ## 🎯 Objective Predict: > The number of turtles caught **per week per capture site** during **2019**. ### Target variable Weekly count of turtles per `(CaptureSite, Week, Year)`. ### Evaluation metric Zindi typically used **Mean Absolute Error (MAE)** — lower is better. --- ## 🧠 Approach Summary 1. **Aggregate** `train.csv` to weekly counts per capture site (1998–2018). 2. **Feature Engineering**: - Lag features (1, 2, 3, 4, 52 weeks) - Rolling means (4, 8, 12, 52 weeks) - Calendar features (week of year sine/cosine) - …