Project Overview

This project delivers a full-cycle machine learning workflow to forecast daily sales for Rossmann stores with a time‑aware XGBoost model. The pipeline cleans and engineers features, then evaluates on a final six‑week holdout to mirror deployment.

Highlights

  • Objective: Predict daily log1p(Sales) per store.
  • Data coverage: 2013‑01‑01 → 2015‑07‑31; 844,338 open‑store rows; 1,115 stores.
  • Final test (log scale): RMSE 0.169, MAE 0.130, R² 0.833.
  • Key drivers: Promotions, competition distance, weekday and store effects.

Repository Structure

.
├── 01_EDA.ipynb / 01_EDA.pdf          # Cleaning, EDA, feature engineering
├── 02_Modeling.ipynb / 02_Modeling.pdf# Time-based split, tuning, evaluation
├── data/
│   ├── train.csv                      # Kaggle Rossmann 'train' (place here)
│   ├── store.csv                      # Kaggle Rossmann 'store' (place here)
│   └── data_processed.csv             # Exported from EDA for modeling
├── exports/                           # Created by the modeling notebook
│   ├── xgb_final_model.joblib
│   ├── X_test_processed.csv
│   ├── y_test_processed.csv
│   └── test_predictions.csv
└── Rossmann_Final_Report_Pro.md       # Project report

Quickstart

  1. Create an environment

    python -m venv .venv && source .venv/bin/activate  # or use conda
pip install -U pandas numpy scikit-learn xgboost optuna shap matplotlib

  2. Add data

    • Place train.csv and store.csv into data/.

  3. Run EDA

    • Open 01_EDA.ipynb and run all cells.
    • It writes data/data_processed.csv used by modeling.

  4. Run Modeling

    • Open 02_Modeling.ipynb and run all cells.
    • It creates a chronological split (last six weeks = test), tunes XGBoost, evaluates once on the holdout, and saves artifacts under exports/.

Results at a glance

  • Test performance (log scale): RMSE 0.169, MAE 0.130, R² 0.833.
  • Top signals: Promo variables, CompetitionDistance, weekday and store effects.

Reproduce Locally

  • EDA: cleans data, builds engineered features (for example, ActiveStoreCount, promo interactions), applies log transforms, and exports data_processed.csv.
  • Modeling: verifies leakage protection, encodes categoricals, tunes XGBoost, and exports the trained model and test artifacts.

Predict with the Saved Model

import joblib, pandas as pd, numpy as np
model = joblib.load("exports/xgb_final_model.joblib")
X = pd.read_csv("exports/X_test_processed.csv")  
pred_log = model.predict(X)
pred = np.expm1(pred_log)  # back-transform if training used log1p(Sales)

Justin Castillo
jcastillo.hotels@gmail.com
GitHub
LinkedIn