Personalized Chain-of-Thought Summarization of Financial News for Investor Decision Support

1 st Tianyi Zhang, 2 nd Mu Chen · 2025

Read Personalized Chain-of-Thought Summarization of Financial News for Investor Decision Support by 1 st Tianyi Zhang, 2 nd Mu Chen (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-05508.txt.

Introduction

Tianyi Zhang and a co-author build a chain-of-thought summarizer for financial news, tuned to each investor's own keywords. Advisors and investors face a flood of news where noise hides the market signals that drive timely calls.

Their frame turns raw stories into short, event-led summaries that stress only the facts a user's keywords mark as key. Those summaries then feed a language model that writes an investor-focused account. The middle layer aims to cut overload while it keeps the facts that a given reader trades on, a help for fast, clear briefs.

Catalog record

CREATOR1 st Tianyi Zhang, 2 nd Mu Chen
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE28 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-05508

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Personalized Chain-of-Thought Summarization of Financial News for Investor Decision Support 1 st Tianyi Zhang, 2 nd Mu Chen Abstract Financial advisors and investors struggle with information overload from financial news, where irrelevant content and noise obscure key market signals and hinder timely investment decisions. To address this, we propose a novel Chain-of-Thought (CoT) summarization framework that condenses financial news into concise, event-driven summaries. The framework integrates user-specified keywords to generate personalized outputs, ensuring that only the most relevant contexts are highlighted. These personalized summaries provide an intermediate layer that supports language models in producing investor-focused narratives, bridging the gap between raw news and actionable insights. I Introduction In today’s financial markets, the stream of information is overwhelming, often repetitive and noisy to investors’ actual needs [ 1 ] . However, extracting relevant insights from this information flow is critical for effective decision-making. Traditional summarization methods, while useful in reducing length, often fail to capture the context [ 2 ] and personalization [ 3 ] required for financial professionals to act confidently on the information presented. Recent advances in Large Language Models (LLMs) have created opportunities for more sophisticated text summarization, yet these models often face challenges such as hallucination [ 4 ] , lack of domain grounding [ 5 ] , and limited personalization [ 6 ] . Hallucination represents a critical issue where LLMs generate factually incorrect or fabricated information that appears plausible but contradicts the source material, particularly problematic in financial contexts where accuracy is paramount for investment decisions [ 7 ] . The lack of domain grounding manifests when models fail to maintain specialized terminology, context-specific relationships, and industry-relevant focus, leading to summaries that may be linguistically coherent but lack the precision required for professional financial analysis [ 8 ] . Limited personalization capabilities prevent these models from adapting their outputs to specific user needs, interests, or expertise levels, resulting in generic summaries that may not highlight the most relevant information for particular stakeholders such as financial advisors focusing on specific market sectors or investment strategies [ 9 ] . These shortcomings make it difficult for financial professionals to fully trust or act on automatically generated outputs, as inaccurate information can lead to significant financial losses and regulatory compliance issues. To address these challenges, Chain-of-Thought (CoT) reasoning has emerged as a promising approach [ 10 ] [ 11 ] . By explicitly modeling intermediate reasoning steps, CoT methods improve coherence, logical flow, and interpretability, enabling summaries that filter out irrelevant details while maintaining structured and contextually faithful content [ 12 ] . However, existing CoT-based summarization approaches remain largely generic, lacking mechanisms to adapt to the specific priorities and information needs of individual investors or advisors [ 13 ] [ 14 ] . Most current approaches employ uniform reasoning chains that do not differentiate between various types of financial information, treating earnings reports, merger announcements, and regulatory updates with the same generic processing framework [ 15 ] . Furthermore, these methods lack domain-specific knowledge integration, failing to recognize the hierarchical importance of quantitative metrics such as revenue growth, profit margins, and valuation multiples that are critical for investment decision-making [ 16 ] . The reasoning chains in existing CoT approaches also do not incorporate user-specific context, such as investment focus areas and risk tolerance, which are essential for generating actionable financial insights [ 17 ] . This generic approach results in summaries that, while logically coherent, fail to provide the targeted, quantitative, and contextually relevant information that financial professionals require for effective decision-making. To address this gap, we propose a personalized CoT summarization framework for financial news. Our approach generates concise, event-driven summaries that are tailored to user-defined keywords, ensuring that only the most relevant information are highlighted. These keyword-driven summaries function as an intermediate representation, allowing LLMs to produce narratives that are both contextually rich and directly aligned with investor interests. In doing so, the framework bridges the gap between raw financial reporting and actionable investor insights. Our experimental evaluation demonstrates the effectiveness of this approach: our enhanced summaries achieve a BLEU score of 0.1786 and ROUGE-L score of 0.4028, representing substantial improvements of 267% and 90% respectively over GPT-4o generated summaries (Table I ). Furthermore, our personalized keyword-based binary relevance classification outperforms commonly used ranking selection methods by 40% in accuracy (Table II ), demonstrating superior precision in identifying articles relevant to specific investor interests. These results validate that our specialized CoT framework not only produces higher-quality financial summaries but also delivers more accurate personalization for investment decision support. The contributions of this paper are twofold: 1. Framework Design – We introduce a CoT-based summarization pipeline that condenses financial news into concise, event-focused outputs. 2. Personalization Mechanism – We incorporate keyword-driven filtering to adapt summaries to the unique needs of financial advisors and investors. II Related Work II-A Financial News Summarization and Domain Adaptation Financial news summarization (FNS) has been studied as a way to digest market-moving information efficiently. Summaries of financial reports and news are expected to retain events and signals that could affect market behavior [ 18 ] . Early approaches ranged from extractive techniques, such as selecting and ranking key sentences to fully abstractive generation. Graph-based methods such as TextRank have been applied to extract key sentences from stock news [ 19 ] . While extractive methods gather critical information, they struggle with long-range dependencies and offer no personalization. Abstractive summarization with encoder–decoder models such as BART or PEGASUS can produce more coherent financial summaries [ 20 ] , but these models risk hallucinating unsupported details and require large domain-specific datasets, which are costly to create. Domain adaptation has proven effective in addressing these challenges. FinBERT, for example, is pre-trained on financial corpora to capture the field’s unique context [ 21 ] . However, it is important to note that FinBERT is not designed for summarization; rather, it functions as a domain-adapted encoder for classification or embedding tasks. This highlights effective summarization architectures are still required to produce coherent and relevant summaries. Prior work has also explored financial summarization for investment insights, e.g., generating news summaries to assist stock prediction [ 22 ] , or company-focused news headlines [ 23 ] . Beyond task-specific models, broader surveys and evaluations have examined the intersection of NLP and finance. Du et al. provide a comprehensive survey of NLP in finance, covering applications such as sentiment analysis, narrative processing, forecasting, and summarization, while also outlining persistent challenges related to domain adaptation, data quality, and specialized summarization [ 24 ] . Complementing this, Du et al. assess the reasoning capabilities of LLMs in financial sentiment analysis, showing that even advanced models like GPT-3.5 and PaLM-2 are particularly weak in numerical and comparative reasoning [ 25 ] . Similarly, Cambria et al. discuss the opportunities and challenges of applying LLMs in financial contexts, noting that issues of factuality, domain grounding, and personalization remain unresolved for decision-support applications [ 26 ] . Nonetheless, most existing financial summarization systems still produce one-size-fits-all outputs and lack mechanisms to adapt to an individual investor’s focus. Our work addresses this gap by combining domain-aware summarization with personalized filtering. II-B Personalized and Controllable Summarization Personalized summarization aims to tailor outputs to the reader’s

Read the complete text