📈 Dynamic Supply Chain Demand Forecasting using Machine Learning

Python Xgboost scikit-learn Streamlit demand-forecasting supply-chain Machine Learning

A machine learning-based retail sales forecasting system that predicts daily store sales using historical sales data, store characteristics, promotions, holidays, competition information, and time-based features.

🚀 Live Demo

Demand-forcasting-demo.mp4

🌐 Streamlit App: https://dynamic-demand-forecasting.streamlit.app/

---

📌 Project Overview

Accurate demand forecasting helps retailers make better decisions about inventory, staffing, and promotional planning.

This project develops a Retail Sales Forecasting system using Machine Learning to predict daily sales for retail stores.

The project covers the complete Machine Learning workflow:

Data Collection → Data Preprocessing → EDA → Feature Engineering → Model Training → Model Evaluation → Deployment

---

🎯 Objective

The main objectives of this project are to:

• Predict daily retail store sales.
• Identify important factors affecting sales.
• Compare different regression algorithms.
• Select the best-performing model based on evaluation metrics.
• Deploy the final model as an interactive Streamlit application.
• Support better demand and inventory planning.

---

📊 Dataset

• Dataset: Rossmann Store Sales
• Source: Kaggle
• Final Dataset Size: 844,338 rows
• Features: 18
• Target Variable: Sales

The dataset contains information related to:

• Store characteristics
• Promotions
• Holidays
• Competition
• Product assortment
• Historical sales
• Date and seasonal patterns

---

🛠️ Technologies Used

• Python
• Pandas
• NumPy
• Scikit-learn
• XGBoost
• Matplotlib
• Seaborn
• Joblib
• Streamlit

---

🔄 Machine Learning Workflow

1. Data Preprocessing

• Merged sales.csv and store.csv using the Store column.
• Handled missing values.
• Removed the Customers feature to avoid data leakage during forecasting.
• Filtered records where stores were open and sales were greater than zero.
• Converted the Date column into a datetime format.

2. Feature Engineering

Created additional time-based features:

• Year
• Month
• Day
• WeekOfYear
• IsWeekend

3. Categorical Encoding

Categorical variables were processed using:

ColumnTransformer + OneHotEncoder

This converts categorical features such as StoreType, Assortment, and PromoInterval into numerical features suitable for Machine Learning models.

---

📊 Exploratory Data Analysis

Several analyses were performed to understand sales patterns:

• Sales distribution
• Average monthly sales
• Average sales by day of week
• Sales during promotional and non-promotional periods
• Sales distribution across store types
• Numerical feature correlation analysis
• XGBoost feature importance

Key Observations

• Sales show noticeable variation across different months.
• Promotional periods are associated with higher average sales.
• Sales patterns vary across different store types.
• Store type and promotion-related features were among the important predictors in the final XGBoost model.

---

🤖 Models Compared

Four regression algorithms were trained and evaluated:

1. Linear Regression
2. Decision Tree Regressor
3. Random Forest Regressor
4. XGBoost Regressor

The models were evaluated using:

• MAE — Mean Absolute Error
• MSE — Mean Squared Error
• RMSE — Root Mean Squared Error
• R² Score

---

🏆 Final Model Performance

Metric	XGBoost
MAE	757.67
RMSE	1066.65
R² Score	0.8821

The XGBoost Regressor achieved the best overall performance among the models evaluated and was selected as the final model.

R² Score: 88.21%

The model explains approximately 88.21% of the variation in sales on the test data.

---

🔍 Important Features

Feature importance analysis was performed using the trained XGBoost model.

Some of the important features included:

• StoreType_b
• Promo
• Promo2SinceYear
• CompetitionDistance
• PromoInterval
• Promo2
• CompetitionOpenSinceYear
• CompetitionOpenSinceMonth
• Assortment
• Store

This helps provide insight into the factors influencing the model's sales predictions.

---

🌐 Streamlit Application

The trained XGBoost model was deployed using Streamlit.

The application allows users to provide store-related inputs and receive an estimated daily sales prediction.

Application Workflow

User Input → Preprocessing → Trained XGBoost Model → Sales Prediction

The trained model is saved using Joblib and loaded by the Streamlit application during prediction.

---

📁 Project Structure

Dynamic-Supply-Chain-Demand-Forecasting/
│
├── app.py
├── README.md
├── LICENSE
├── requirements.txt
├── .gitignore
│
├── data/
│   ├── sales.csv
│   └── store.csv
│
├── models/
│   └── xgboost_sales_model.pkl
│
├── notebooks/
    └── Retail_Sales_Forecasting.ipynb

---

💡 Business Applications

The forecasting system can support:

• 📦 Inventory planning
• 🏪 Store-level demand planning
• 👥 Staff allocation
• 📢 Promotional planning
• 📊 Sales analysis
• 🚚 Supply chain decision-making

Accurate demand predictions can help businesses reduce the risk of overstocking and understocking and make more informed operational decisions.

---

👨‍💻 Author

Muhammed Fayis

Data Science Machine Learning