June 29, 2025
The project, implemented using Python, forecasts store sales using historical unit sales data for thousands of items across multiple stores. Preprocessing involved handling missing values, encoding categories, and extracting date features, ensuring proper time-series formatting. Feature engineering included lag features, rolling stats, store-item patterns, and Ecuador national and local holidays, along with temporal patterns like trends and seasonality. ML models: The first used Linear Regression (scikit-learn) for baseline predictions, followed by Deep Learning (LSTM networks with TensorFlow) to improve accuracy by capturing complex temporal patterns.