Comparative Analysis of Time Series Forecasting Libraries Across Multiple Sectors

Abstract

This thesis presents a comparative analysis of various time-series forecasting models applied to datasets from retail sales, exchange rates, and bike-sharing systems. The study evaluates traditional statistical methods, machine learning techniques, and deep learning approaches, focusing on their accuracy, efficiency, and flexibility. The research highlights the strengths and limitations of each model, emphasizing the impact of incorporating exogenous variables.

Traditional models like SARIMA demonstrated robustness in capturing long-term trends and seasonal variations, particularly in retail sales data, but struggled with high volatility. Advanced machine learning models, such as those implemented in GluonTS, showed significant improvements with the inclusion of exogenous variables, reducing RMSE and MAE values notably. However, there were some compatibility issues that impacted the usability of some models.

The findings indicate that no single model universally outperforms others across all datasets. Instead, the effectiveness of forecasting models is highly dependent on the dataset’s specific characteristics and the inclusion of relevant exogenous variables. This research underscores the importance of model selection and customization based on the unique requirements of each forecasting task, contributing valuable insights into the field of time series analysis.