Julie-Montague/Zindi-Maize-Price-Prediction-Challenge

Using historical prices of dry maize in Kenya, this project develops a machine learning solution to predict average weekly prices of maize in the counties of Kiambu, Kirinyaga, Mombasa, Nairobi and Uasin-Gishu. # AgriBora-Maize-Price-Prediction-Challenge Competition link : zindi.africa ## Project Overview This project develops machine learning models to predict maize prices using time-series data. Built as part of a Zindi data science competition, the project focuses on: - Time-series feature engineering - Model comparison and evaluation - End-to-end data processing pipeline ## Key Contributions - Implemented full pipeline: preprocessing → feature engineering → modeling → evaluation - Engineered time-series features (lags, rolling statistics, trends) - Compared multiple models (LightGBM, XGBoost) - Evaluated performance using RMSE and MAE ## Tech Stack - Python (pandas, numpy, sklearn) - ML Models : CatBoostRegressor, RidgeRegressor, MLPRegressor, HistGradientBoostingRegressor - Time-series feature engineering - Matplotlib / Seaborn ## Results - Built predictive models for maize price forecasting - Demonstrated impact of feature engineering on model performance - Evaluated models using standard regression metrics (RMSE, MAE) ## Reproducibility All steps from data preprocessing to model training are included in this repository and can be reproduced. --- ## OVERVIEW Using historical prices of dry maize in Kenya, this project develops a machine learning solution to predict average weekly prices of maize in the counties of Kiambu, Kirinyaga, Mombasa, Nairobi and Uasin-Gishu. By forecasting average weekly dry maize prices, the models aim to provide short-horizon market intelligence that can help farmers decide when to sell after storing produce in certified warehouses. Reliable two-week-ahead forecasts strengthen agriBORA’s storage–credit–market workflow by enabling better timing decisions for delayed selling and improving expected returns. ## OBJECTIVES 1. **Build a clean weekly panel** (county × week-start date) from the challenge data, robust to duplicate rows and irregular reporting. 2. **Improve …