Sales Forecasting Using Time Series Analysis with XGBoost

Project Overview 🎯

This project aims to forecast sales for thousands of product families sold at Favorita stores using historical data. The analysis is performed using XGBoost, a powerful machine learning algorithm. Through feature engineering and model tuning, the project provides both short-term and long-term sales predictions, with a focus on generating a 30-day forecast.

Project Workflow 📈

1. Data Collection and Preprocessing:

   • Used historical sales data provided by Favorita stores.
   • Processed the data to handle missing values and outliers.
   • Performed feature engineering to create lag features and rolling statistics.
   • Applied one-hot encoding for categorical variables and scaling for numerical variables.

2. Exploratory Data Analysis (EDA):

   • Visualized sales trends over time.
   • Analyzed seasonal effects, promotions, and other external factors influencing sales.

3. Feature Engineering:

   • Created lag features and rolling windows to capture historical patterns.
   • Engineered calendar-related features such as day of the week, month, and promotions.

4. Model Building:

   • Implemented XGBoost for time series forecasting.
   • Fine-tuned the model using RMSE and RMSLE as evaluation metrics.
   • Generated a 30-day sales forecast.

5. Model Evaluation:

   • Used RMSE and RMSLE to assess the model’s accuracy.
   • Compared the model’s predictions against actual sales data for validation.

Key Features 🛠️

• XGBoost Model: Applied an advanced machine learning approach for time series forecasting.
• Feature Engineering: Integrated lag and rolling statistics to enhance forecast accuracy.
• Data Visualization: Plots and charts to illustrate key trends and model performance.
• 30-Day Sales Forecast: Focused on predicting future sales, with results analyzed for potential improvements.

Challenges & Improvements 🔧

• Current Challenge: The 30-day forecast results show a linear pattern, lacking the nuanced shape of actual sales data.
• Future Work: Exploring advanced techniques such as model stacking, ensemble methods, and deeper time-series models (e.g., ARIMA, LSTM) to improve forecast accuracy.

Results 📊

• Achieved a RMSLE of 0.409.
• Significant improvement in prediction accuracy through feature engineering.

Check it out on Kaggle 🔗

You can also explore the project on Kaggle: https://www.kaggle.com/code/walidkw/store-sales-ts-forecasting-xgboost