Sales Forecasting Description

Overview

The Sales Forecasting Description project is part of the Self-Paced Elevvo Machine Learning Internship Level 3.

This project focuses on analyzing and predicting sales trends using Walmart’s historical retail data. Sales forecasting is essential in retail for inventory management, staffing, financial planning, and demand prediction. By leveraging machine learning techniques, the project demonstrates how data-driven models can improve decision-making and optimize resources.

The workflow includes data preprocessing, feature engineering, exploratory data analysis (EDA), and implementation of advanced models such as XGBoost and LightGBM, with performance comparisons across different feature selection strategies.

---

Table of Contents

• Features
• Technologies Used
• Usage
• Data Sources
• Project Structure
• Results & Visualizations

---

Features

• Data preprocessing and cleaning
• Feature engineering and selection (manual, k-best, recursive feature elimination (RFE), full set)
• Exploratory data analysis (EDA) with Jupyter Notebook & interactive HTML export
• Implementation and comparison of XGBoost and LightGBM models
• Evaluation using R², RMSE, and R² drop (to check generalization and overfitting)
• Visualization of sales trends and model performance

---

Technologies Used

• Python
• Jupyter Notebook
• ydata-profiling
• Pandas, NumPy
• Scikit-learn
• Matplotlib, Seaborn
• XGBoost, LightGBM

---

Usage

1. Open Jupyter Notebook:

   jupyter notebook

2. Run the notebooks in order:

   • Data Analysis & Visualization.ipynb → Perform EDA and visualize sales trends
   • Data Preprocessing & Feature Engineering.ipynb → Prepare and clean the dataset
   • Feature Selection & Model Development.ipynb → Apply feature selection, train models, and compare results

3. Alternatively, explore SalesForecastingDescriptionEDA.html for a static, interactive version of the EDA results.

---

Data Sources

Dataset: Walmart Sales Forecasting Dataset (Kaggle)

Raw data files include:

• features.csv → Economic & seasonal features (CPI, unemployment, fuel price, etc.)
• stores.csv → Store metadata
• train.csv → Training dataset with historical weekly sales
• test.csv → Test dataset for evaluation and prediction

---

Project Structure

Sales-Forecasting-Description/
│
├── Data/
│   ├── Raw Data/
│   │   ├── features.csv
│   │   ├── stores.csv
│   │   ├── train.csv
│   │   └── test.csv
│
├── Predictions/
│   ├── Weekly_Sales_Prediction.xlsx   # Model output predictions
│
├── Notebooks/
│   ├── Data Analysis & Visualization.ipynb
│   ├── Data Preprocessing & Feature Engineering.ipynb
│   ├── Feature Selection & Model Development.ipynb
│
├── SalesForecastingDescriptionEDA.html
└── README.md

---

Results & Visualizations

LightGBM – Actual vs Predicted

XGBoost – Actual vs Predicted

Model Comparison – Validation R²

Model Comparison – Validation RMSE

Model Comparison Summary

Feature Set	XGB Val R²	LGBM Val R²	XGB Val RMSE	LGBM Val RMSE	XGB R² Drop	LGBM R² Drop
Manual	0.9301	0.9597	5804.13	4406.48	-0.0085	0.0077
K-best	0.9217	0.9362	6139.64	5544.41	-0.0262	0.0015
RFE	0.9170	0.9567	6321.51	4568.02	-0.0144	0.0065
All	0.9309	0.9603	5769.54	4371.43	-0.0107	0.0088

Key Insights:

• LightGBM consistently outperforms XGBoost across all feature sets.
• Manual and All feature sets yield the best results.
• Small R² drop values → Both models generalize well with only minor overfitting.