This project focuses on forecasting weekly sales for a subset of Walmart stores using various machine learning and time series models.
The analysis covers models such as Ridge Regression, Random Forest, XGBoost, SARIMA, and LSTM, and explores how different factors like holidays, fuel prices, and temperature influence sales. The project also includes an extensive exploratory data analysis to understand the impact of holidays, fuel prices, temperature, and other factors on sales.
The goal of this project is to provide accurate sales forecasts for Walmart, enabling better inventory management, staffing, and financial planning. This is crucial for optimizing operations and enhancing customer satisfaction in the competitive retail landscape.
The dataset includes weekly sales data from 45 Walmart stores, along with associated store attributes and external factors such as holidays, temperature, CPI, and fuel prices. Special attention was given to data cleaning, particularly in handling missing values and anomalies.
Implementing various models and tuning them for optimal performance.
- SARIMA: Adjusted for seasonality and stationarity.
- Ridge Regression: Implemented with sklearn but showed high WMAE error.
- Random Forest and XGBoost Regressors: Included hyperparameter tuning.
- LSTM and MLP Models: Explored for deep learning approaches.
- The SARIMA model proved to be the most effective, particularly due to its handling of seasonality and trends in the sales data.
- Machine learning models like XGBoost and Random Forest showed promise but required extensive tuning.
- Deep learning approaches, while insightful, were limited by computational demands.
- Challenges such as hyperparameter tuning and feature selection were pivotal in model performance.
- The potential for integrating more dynamic features like economic indicators or social media trends could be explored to improve forecasting accuracy.
This project encountered challenges in hyperparameter tuning, feature selection, and computational constraints. Future work could explore integrating more external data sources and experimenting with advanced neural network architectures.