Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Mathematics

Program Name/Specialization

Mathematics for Science and Finance

Faculty/School

Faculty of Science

First Advisor

Prof. Roman Makarov

Advisor Role

Supervisor

Abstract

Financial markets are driven by both quantitative data and qualitative narratives continu- ously disseminated through financial news. Understanding how these narratives shape price dynamics and investor behavior remains a critical challenge in financial modelling. This dissertation develops a comprehensive methodological framework that systematically inte- grates news sentiment into return and volatility forecasting models. The central objective is to leverage this qualitative data to better characterize asset return distributions, predict long-memory volatility, and infer latent market regimes. The empirical foundation rests on a massive dataset of over 347,000 firm-specific news headlines and corresponding financial data for 37 assets across equity, cryptocurrency, and commodity markets from March 2022 to January 2025. To quantify market sentiment, two advanced natural language process- ing techniques are employed: the domain-specific FinBERT transformer model [6] extracts continuous sentiment polarity scores, while a Random Forest classifier utilizing TF-IDF Bag-of-Words representations that computes directional news impact probabilities across four distinct intraday and interday trading horizons. Rather than treating sentiment as a simple standalone variable, this research embeds it deeply into the statistical structure of financial models. The study begins by demon- strating through Markov mixture models that asset return distributions fundamentally shift across sentiment states, identifying the Student’s t mixture distribution as the op- timal fit over Gaussian and Laplace alternatives. Building on this distributional insight, a Sentiment-Adjusted Markov-Switching FIGARCH (SA-FIGARCH) model is introduced. Empirical evaluations reveal that these sentiment-enhanced specifications significantly out- perform baseline models in capturing volatility clustering and long-memory effects, partic- ularly when overnight news accumulation is absorbed at the next day market open. The framework culminates in the development of sentiment-augmented Hidden Markov Models including Discrete (DHMM), Gaussian (GHMM), and Autoregressive (ARHMM) architec- tures coupled with a novel analogue-based forecasting algorithm which utilizes likelihoods. Results demonstrate that augmenting these state-space models with text-derived sentiment signals drastically refines the detection of unobservable market regimes. Most notably, the integrated framework improves out-of-sample directional forecasting accuracy in 94.4% of the analyzed assets, yielding an average performance gain of 20.30% compared to tradi- tional models. Finaly, this dissertation provides a robust, interpretable toolkit bridging behavioral finance and machine learning, proving that qualitative market narratives leave measurable, systematic, and highly exploitable traces in asset price dynamics.

Convocation Year

2026

Convocation Season

Fall

Download

Available for download on Tuesday, December 22, 2026

Share

COinS