Electronic Theses and Dissertations

Date of Award

5-1-2026

Document Type

Thesis

Degree Name

M.S. in Engineering Science

First Advisor

Hong Xiao

Second Advisor

David Harrison

Third Advisor

Kristin Davidson

School

University of Mississippi

Relational Format

dissertation/thesis

Abstract

Long-term time-series forecasting is a fundamental problem in real-world applications such as financial modeling and weather forecasting. While Transformer-based architectures have shown strong performance in sequence modeling like natural language processing and computer vision, their direct application to time-series forecasting faces challenges including quadratic computational complexity and capturing inherent temporal dependencies. Recent works such as Informer and Autoformer address these limitations through sparse attention mechanisms and series decomposition, but introduce trade-offs between predictive accuracy and computational efficiency.

In this thesis, we propose Kairosformer, a Transformer-based architecture for long-horizon time-series forecasting that integrates probabilistic attention with series-wise temporal aggregation. By combining efficient attention mechanisms with temporal decomposition, Kairosformer improves scalability while maintaining predictive performance.

We conduct experiments on multiple benchmark datasets under various prediction horizons. The results demonstrate that Kairosformer consistently improves trade-offs between predictive accuracy and computational efficiency compared to Transformer-based models. A ranking-based evaluation further shows that Kairosformer A2 provides the best overall performance, outperforming baseline models when both accuracy and efficiency are considered. In particular, Kairosformer shows improved runtime efficiency and stable performance in long-horizon forecasting, emphasizing its potential applicability to time-series prediction tasks.

These findings suggest that combining temporal structure with efficient attention mechanisms is a promising direction for scalable and accurate time-series forecasting.

Share

COinS
 
 

To view the content in your browser, please download Adobe Reader or, alternately,
you may Download the file to your hard drive.

NOTE: The latest versions of Adobe Reader do not support viewing PDF files within Firefox on Mac OS and if you are using a modern (Intel) Mac, there is no official plugin for viewing PDF files within the browser window.