How ehgeraldo/rossmann_store_sales works
Project Type and Purpose
• This is a data-science / machine-learning forecasting pipeline.
• Purpose: Ingest historical store data, run transformations / EDA, train & tune models, and expose a prediction API for six-week-ahead daily sales forecasts.
Main High-Level Components
a. Data Storage
• Raw CSV files: “data/train.csv”, “data/test.csv”, “data/store.csv”
b. ETL & Preprocessing Layer
• Notebooks c0 → c7 (Planning, Fill NAs, Descriptive Stats, Feature Engineering, Data Preparation)
c. Exploration & Feature Selection
• Notebooks c4 → c8 (Univariate → Multivariate EDA, Boruta Selection)
d. Model Training & Validation
• Notebooks c9 → c15 (Linear & Non-Linear Models, Cross-Validation, Hyperparameter Tuning, Error Interpretation, Rossmann Classifier)
e. Model Artifact Store
• Serialized model files (e.g., pickle / joblib) on disk or model registry
f. Deployment / Serving Layer
• Notebook c16 (API_Handler) → a REST API service (Flask or FastAPI)
• Client (web dashboard or external consumer) calls the API
Relationships & Data Flow
• Data Storage → ETL/Preprocessing: raw CSVs flow into transformation routines
• ETL → EDA & Selection: cleaned data feeds exploratory plots and feature pickers
• EDA → Model Training: selected features and splits feed into training routines
• Training → Model Store: best model and tuned parameters are saved
• Model Store → API Service: the production model is loaded for inference
• Client → API Service: external calls for daily-sales predictions
Architectural Patterns & Technologies
• Pipeline Pattern: a linear, staged data-processing flow
• Layered Architecture: Data Layer → Processing Layer → Model Layer → Service Layer
• Key Libraries: pandas / numpy (ingest & transform), scikit-learn (modeling & CV), BorutaPy (feature selection), Flask/FastAPI (serving)
• Notebooks should be refactored into Python modules for production