🛒 Rossmann Store Sales Forecasting

📌 Project Overview

Rossmann operates over 3,000 drug stores in 7 European countries. In this Kaggle competition, the task is to forecast daily sales for 6 weeks across 1,115 Rossmann stores in Germany. Accurate forecasts help store managers improve planning and resource allocation.

This project implements a complete machine learning pipeline to tackle this business problem, including:

• Data Cleaning
• Exploratory Data Analysis
• Feature Engineering
• Model Training
• Ensemble Learning
• Final Submission to Kaggle

---

📁 Dataset Description

train.csv

• Historical sales data from 2013–2015.
• Includes Store ID, Sales, Customers, Open/Closed status, and Promotion info.

test.csv

• Same structure as train.csv without Sales.
• Used for making final predictions.

store.csv

• Store metadata: StoreType, Assortment level, Competition distance, Promo2 status, etc.

sample_submission.csv

• Format template to structure predictions for Kaggle evaluation.

---

📊 Workflow Summary

🔹 1. Load & Merge Data

• Merged train.csv and test.csv with store.csv on the Store column.

🔹 2. Handle Missing Values

• Filled missing numerical values with median (e.g. CompetitionDistance).
• Filled PromoInterval with the most frequent value.

🔹 3. Feature Engineering

• Extracted Day, Month, Year from the Date column.
• Filtered training data to include only rows where stores were open (Open == 1).

🔹 4. EDA (Exploratory Data Analysis)

• Visualized sales distributions.
• Examined relationships between Sales and Promo, DayOfWeek, Month, Year.
• Generated a correlation heatmap to identify feature relationships.

🔹 5. Data Preprocessing

• Applied MinMaxScaler to normalize numeric features.
• Applied OneHotEncoder to convert categorical columns into numeric format.

🔹 6. Modeling

Trained a variety of models:

• Linear Models: Linear Regression, Ridge, Lasso, ElasticNet
• Tree-Based Models: Decision Tree, Random Forest

🔹 7. Ensemble Learning

• Combined Ridge, Decision Tree, and Random Forest predictions using weighted averaging.
• Weights: 0.6 * RandomForest + 0.3 * DecisionTree + 0.1 * Ridge

🔹 8. Evaluation Metrics

• RMSE: Root Mean Squared Error
• RMSPE: Root Mean Square Percentage Error (used by Kaggle)

Model	RMSE	RMSPE
RandomForest	1154.43	18.74%
DecisionTree	1469.77	21.88%
Ridge	2718.66	48.44%
Ensemble	1306.09	18.76%

🔹 9. Feature Importance

Visualized top 10 features contributing to each model's prediction using bar charts.

🔹 10. Final Submission

• Predictions were made on the test set.
• Final results were stored in submission.csv as per Kaggle format:

csv
Id,Sales
1,5263.0
2,6064.0
...

🗂️ Project Structure

Rossmann-Sales-Forecasting/
├── data/
│   ├── train.csv
│   ├── test.csv
│   ├── store.csv
│   └── sample_submission.csv
│
├── notebooks/
│   └── Rossmann_Sales_Forecasting.ipynb
│   └── Updated_Rossmann_Sales_Forecasting.ipynb
|
├── outputs/
│   └── submission.csv
│
├── README.md
└── .gitignore