Temporal Sentiment Shift Detection in Financial Markets: Domain Expertise vs Architectural Innovation

Guduguntla, Pallavi (2025) Temporal Sentiment Shift Detection in Financial Markets: Domain Expertise vs Architectural Innovation. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (905kB) | Preview
Preview	PDF (Configuration Manual) Download (588kB) | Preview

Abstract

This research investigates relative effectiveness of domain-specific pre-training against architectural innovations in detecting temporal shifts of financial sentiment across various regimes of markets. While models constructed using transformers have revolutionized analysis of financial sentiment, their ability of detecting shifts of sentiment during regime shifts of markets is not deep enough with related research. This study addresses this gap by systematically comparing FinBERT's financial domain expertise against ModernBERT's architectural advances including extended context windows, Rotary Positional Embeddings, and alternating attention mechanisms. Using a comprehensive dataset of 37,509 financial news articles from 2009-2023, augmented with market regime classifications and temporal annotations, three models were fine-tuned and evaluated: FinBERT, ModernBERT-base, and ModernBERT-large. The evaluation framework incorporated traditional classification metrics, specialized temporal shift detection measures, and cross-regime performance analysis. Results demonstrate that FinBERT's domain-specific pre-training yields superior performance with 91.49% overall accuracy compared to ModernBERT-base (87.93%) and ModernBERT-large (88.82%). Crucially, FinBERT achieved 90.69% temporal accuracy on sentiment shift detection, outperforming both ModernBERT variants. All models exhibited robust performance across all regimes of markets, though with varying efficiency. Statistical testing substantiated significant performance differences (p < 0.001). Our results therefore indicate that architectural innovations of domain-specific financial intelligence provide greater value than architectural innovations for sentiment analysis of time series, with design implications for next-generation sentient financial AI systems. Our contribution entails substantiated models, integrated assessment measurements, as well as empirical evidence for design of sentiment analysis systems for dynamic financial markets.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Byrne, Brian UNSPECIFIED
Subjects:	H Social Sciences > HG Finance H Social Sciences > HG Finance > Fintech T Technology > T Technology (General) > Information Technology > Fintech
Divisions:	School of Computing > Master of Science in FinTech
Depositing User:	Ciara O'Brien
Date Deposited:	20 Aug 2026 11:04
Last Modified:	20 Aug 2026 11:04
URI:	https://norma.ncirl.ie/id/eprint/9572

Actions (login required)

View Item