As machine learning continues to transform various academic disciplines and industries, its applications in time series forecasting and financial data analysis have attracted substantial attention. This dissertation addresses these two critical areas, with a focus on the practical advancements of state-of-the-art machine learning techniques.

In the domain of time series forecasting, we develop and implement cutting-edge machine learning models, including novel architectures based on Transformers and fully linear structures. These models are applied to general time series data and real-world forecasting problems, demonstrating superior performance in accuracy and efficiency compared to traditional methods and other baselines. Additionally, we explore the potential of foundational models for time series analysis, contributing to the growing field of long-term forecasting with large deep learning models.

In the financial domain, the thesis investigates the application of machine learning techniques, particularly regime-switching models, which offer advantages over regime-agnostic approaches. We integrate feature selection methods that enhance regime-aware factor allocation in continuous jump models. Furthermore, we examine the use of variational autoencoders and dynamic programming algorithms in conjunction with regime identification, expanding the analytical toolkit for financial modeling.

Co-Advisers: Vincent Poor and John Mulvey

Zoom Link: https://princeton.zoom.us/j/96865839940