Supervised Contrastive Learning for Classification of Market Stock Series

In the complex world of financial markets, the quest for innovative methods to analyze and interpret market behavior is ongoing. This thesis explores the potential of Supervised Contrastive Learning (SCL) as a novel approach to classifying stock market states, aiming to offer a more nuanced understanding of market dynamics. By constructing a dataset from NASDAQ 100 index prices, the study employs a deep neural network model to examine how SCL performs in comparison to traditional machine learning and deep learning techniques. This thesis attempts to contribute to the broader discourse on financial technology innovation, underlining the importance of continued exploration and experimentation in the development of financial analysis tools.\ The focus is on the process and methodology of applying SCL to financial time series analysis, emphasizing the exploratory nature of this research in seeking new pathways for financial analysis. The research provides valuable insights into the applicability of SCL in financial markets, suggesting directions for future work in enhancing the accuracy and efficiency of market state classification. The task involves categorizing financial time series into three main trends: buy, hold, and sell, over different future time frames—specifically 3, 5, and 7 days following the targeted period. In addition to the empirical evaluation of SCL against traditional and deep learning models, this thesis embarks on a qualitative analysis of the latent representations generated by the SCL model, compared with other models, such as TS2Vec. This analysis seeks to uncover whether these representations can illuminate significant patterns within the financial time series data, offering insights into the underlying mechanisms of market behavior. The ultimate aim is to showcase how the Supervised Contrastive Learning (SCL) approach can be effectively applied to forecast financial time series.