Repurposing language models for FX volatility forecasting: a data-efficient and context-aware approach
File(s)
Published version
Author(s)
Nguyen, Quoc Anh
Guo, Ce
Luk, Wayne
Type
Conference Paper
Abstract
Forecasting foreign exchange (FX) volatility is a critical task in finance. The high data demand of deep learning models creates a practical challenge, because data availability is limited by constantly shifting market dynamics. To address this problem, we introduce Vola-BERT, a novel predictive model that solves this data-efficiency problem by repurposing a pre-trained bidirectional language model. Our core insight is that the model’s value lies not in linguistic knowledge, but in its sophisticated capacity to model complex sequential patterns, which we adapt to the financial time-series domain. Vola-BERT is further enhanced by a new semantic conditioning framework that transparently integrates non-time-series data, such as high-impact economic events. This mechanism transforms the model from a black-box forecaster into an interpretable analytical tool capable of directly quantifying the impact of real-world market drivers. Our empirical study against 11 baseline models shows that Vola-BERT achieves state-of-the-art forecasting accuracy, consistently outperforming baselines under both data-rich and data-scarce conditions. This advantage is particularly pronounced in the data-scarce scenario, where it maintains high accuracy using only 10 percent of the training data. This work demonstrates that repurposing language models is a promising, data-efficient direction for context-aware numerical analysis.
Date Issued
2025-11-14
Date Acceptance
2025-10-10
Citation
ICAIF '25: Proceedings of the 6th ACM International Conference on AI in Finance, 2025, pp.465-473
ISBN
9798400722202
Publisher
ACM
Start Page
465
End Page
473
Journal / Book Title
ICAIF '25: Proceedings of the 6th ACM International Conference on AI in Finance
Copyright Statement
Copyright © 2025 Copyright held by the owner/author(s). This work is licensed under a Creative Commons Attribution International 4.0 License (http://creativecommons.org/licenses/by/4.0/)
License URL
Identifier
10.1145/3768292.3770386
Source
ACM International Conference on AI in Finance
Subjects
CCS Concepts • Applied computing → Forecasting
Decision analysis
• Information systems → Business intelligence
• Computing methodologies → Natural language generation
Temporal reasoning
Transfer learning Volatility Forecasting, Large Language Models, Data-Scarce Learning, Explainable AI, Forex, Time Series Forecasting
Start Date
2025-11-15
Finish Date
2025-11-18
Coverage Spatial
Singapore