Tools
Tools
Romala, Naga Lakshmi Divya (2025) Multi-Horizon stock Price Volatility Forecasting using attention-Based Transformers. Masters thesis, Dublin, National College of Ireland.
Preview |
PDF (Master of Science)
Download (990kB) | Preview |
Preview |
PDF (Configuration Manual)
Download (651kB) | Preview |
Abstract
Volatility forecasting in financial markets remains critical for risk management and trading strategies, yet existing methods inadequately integrate heterogeneous data sources across multiple prediction horizons. This research addresses the fundamental question of how attention-based transformer architectures can effectively integrate financial news sentiment with market indicators to improve multi-horizon stock price volatility forecasting.
The study develops a novel dual-encoder transformer architecture that processes financial news sentiment through an ensemble of five specialized models (FinBERT, FinBERT-tone, Financial RoBERTa, VADER, and TextBlob) alongside historical market data to predict volatility at 1-day, 5-day, and 15-day horizons simultaneously. The methodology employs separate encoders for sentiment and market features, connected through multi-head attention mechanisms that learn dynamic cross-modal relationships. The implementation utilizes a dataset of 105,540 news articles paired with corresponding price data spanning 2010-2021, covering 50 major stocks across diverse sectors.
The experimental results show that the predictions are very accurate, with R² scores of 0.9919, 0.7185, and 0.7740 for the three different prediction horizons. All of these scores are higher than the target threshold of 0.7. The model works well in bull, bear, and sideways markets, with directional accuracy of up to 84.88% for 1-day forecasts. Attention weights analysis shows that the model changes the relative importance of sentiment and market characteristics based on the current situation.
This research presents a functional multi-horizon volatility forecasting system that accommodates diverse trading strategies and expands transformer applications to include financial forecasting. The successful amalgamation of ensemble sentiment analysis and market data through attention mechanisms establishes a novel standard for volatility forecasting systems.
| Item Type: | Thesis (Masters) |
|---|---|
| Supervisors: | Name Email Del Rosal, Victor UNSPECIFIED |
| Subjects: | H Social Sciences > HD Industries. Land use. Labor > HD61 Risk Management H Social Sciences > HG Finance > Fintech T Technology > T Technology (General) > Information Technology > Fintech H Social Sciences > HG Finance > Investment |
| Divisions: | School of Computing > Master of Science in FinTech |
| Depositing User: | Ciara O'Brien |
| Date Deposited: | 20 Aug 2026 11:51 |
| Last Modified: | 20 Aug 2026 11:51 |
| URI: | https://norma.ncirl.ie/id/eprint/9582 |
Actions (login required)
 |
View Item |