Citation

Guang, Jonathan Chung Rong and Shah, Nathar (2026) A Hybrid LLM Framework for Financial News Insight Generation: Integrating Summarization, Domain Adaptation, and Reinforcement-Based Evaluation. Procedia Computer Science, 283. pp. 1581-1599. ISSN 18770509

Text main.pdf - Published Version Restricted to Repository staff only Download (805kB)

Abstract

The surge in financial news and disclosures has created a growing demand for automated systems capable of distilling complex, domain-specific information into actionable insights. This paper presents a hybrid insight generation framework that leverages both general-purpose and fine-tuned language models to analyze financial news and regulatory documents in real time. The system integrates GPT-4o for high-fluency summarization with a locally hosted, LoRA-fine-tuned Mistral-7B-Instruct model trained on over 2 million financial news articles and 10-K filings from 2014 to 2025. Through a modular architecture, the system enriches prompts with technical indicators, performs dual inference (explicit and implicit), and fuses model outputs via an editorial refinement stage. Evaluation is conducted on both short-form (FiQA-2) and long-form (FNS-2020) financial datasets using metrics such as ROUGE, BLEU, BERTScore, BLEURT and Q² F1 BERTScore. Experiments on FiQA-2 and FNS-2020 datasets show the hybrid model improves BERTScore by 3.2% and BLEURT by 4.1% over baselines. Results show that the hybrid model effectively combines linguistic fluency with domain reasoning, outperforming individual models on key insight generation tasks. The pipeline also includes a fallback classifier to filter non-financial queries, ensuring robustness and relevance in user interactions. The primary contribution of this study is the design of a hybrid inference framework that fuses general-purpose and domain-adapted LLMs, enabling real-time financial insight generation with improved factual grounding and fluency.

Item Type:	Article
Uncontrolled Keywords:	Hybrid Large Language Models (Hybrid LLMs); Financial Natural Language Processing (Financial NLP)
Subjects:	H Social Sciences > HG Finance > HG4001-4285 Finance management. Business finance.
Divisions:	Faculty of Computing and Informatics (FCI)
Depositing User:	Ms Rosnani Abd Wahab
Date Deposited:	02 Sep 2026 05:50
Last Modified:	02 Sep 2026 05:50
URII:	http://shdl.mmu.edu.my/id/eprint/16523

Downloads

Edit (login required)