Paper: arXiv 2601.02677

Authors: Gongao Zhang, Haijiang Zeng, Lu Jiang

Abstract

Financial institutions and regulators require systems that integrate heterogeneous data to assess risks from stock fluctuations to systemic vulnerabilities. Existing approaches often treat these tasks in isolation, failing to capture cross-scale dependencies. We propose Uni-FinLLM, a unified multimodal large language model that uses a shared Transformer backbone and modular task heads to jointly process financial text, numerical time series, fundamentals, and visual data. Through cross-modal attention and multi-task optimization, it learns a coherent representation for micro-, meso-, and macro-level predictions. Evaluated on stock forecasting, credit-risk assessment, and systemic-risk detection, Uni-FinLLM significantly outperforms baselines. It raises stock directional accuracy to 67.4% (from 61.7%), credit-risk accuracy to 84.1% (from 79.6%), and macro early-warning accuracy to 82.3%. Results validate that a unified multimodal LLM can jointly model asset behavior and systemic vulnerabilities, offering a scalable decision-support engine for finance.

Complexity vs Empirical Score

  • Math Complexity: 7.0/10
  • Empirical Rigor: 8.5/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents advanced Transformer architecture and cross-modal fusion with mathematical complexity evident in the model design and optimization, while demonstrating strong empirical rigor through specific performance metrics across multiple datasets and tasks, indicating backtest-ready implementation.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Unified Multimodal Financial Analysis"] --> B{"Methodology"}
  B --> C["Data Integration<br>Text, Time Series, Fundamentals, Visuals"]
  C --> D["Core Model<br>Uni-FinLLM: Shared Transformer + Modular Heads"]
  D --> E["Processing<br>Cross-Modal Attention & Multi-Task Optimization"]
  E --> F{"Outcomes & Improvements"}
  F --> G["Micro-Level: Stock Directional Accuracy 67.4%"]
  F --> H["Macro-Level: Systemic Risk Accuracy 82.3%"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes Uni-FinLLM, a unified multimodal large language model with a shared Transformer backbone and modular task heads for joint micro-level stock prediction and macro-level systemic risk assessment.

Source sentence
We propose Uni-FinLLM, a unified multimodal large language model that employs a shared Transformer-based backbone and modular task heads to jointly process textual financial news, numerical market time series, corporate fundamentals, and visual representations of financial dynamics.

Assumes 1

  • All datasets are aligned on a shared temporal axis, normalized, and matched to firm-level identifiers or national economic codes for multimodal reasoning. data
    Source sentence
    To support multimodal reasoning, all datasets were aligned on a shared temporal axis, normalized, and matched to firm-level identifiers or national economic codes.

Data

  • Historical equity data · CRSP, NYSE Trade and Quote (TAQ), Yahoo Finance market-prices commercial
    2006–2024 · more than 4,000 U.S. listed firms
    Source sentence
    For micro-level forecasting, we collected historical equity data from widely used market databases such as CRSP, NYSE Trade and Quote (TAQ), and Yahoo Finance.
  • Financial news and corporate disclosures · Reuters, Bloomberg, SEC EDGAR text-news commercial
    Source sentence
    For firm-level textual information, financial news and corporate disclosures were obtained from Reuters, Bloomberg, and SEC EDGAR, with around 2.8 million text samples processed using financial domain-specific tokenization.
  • Macroeconomic, banking, and regulatory variables · Federal Reserve FRED, IMF IFS, BIS, OECD macro public
    1995–2024
    Source sentence
    For macro-level systemic risk assessment, we assembled a comprehensive set of macroeconomic, banking, and regulatory variables from Federal Reserve FRED, IMF IFS, BIS global banking statistics, and OECD economic indicators. ... The

Recognised sources: Bloomberg SEC EDGAR Yahoo Finance

Costs

Not applicable The paper reports no trading or execution result.

Finds 3

  • Uni-FinLLM achieves 67.4% directional accuracy, 10.9% MAPE, and 64.3% hit ratio in micro-level stock prediction, outperforming baselines. performance
    Source sentence
    It achieves 67.4% directional accuracy, reducing the error to a substantially lower 10.9 MAPE while reaching a 64.3% hit ratio.
  • Uni-FinLLM reaches 84.1% accuracy, 83.6% F1-score, and 0.892 ROC-AUC in credit-risk prediction, the best among compared models. performance
    Source sentence
    Uni-FinLLM delivers the best overall performance, reaching 84.1% accuracy, an F1-score of 83.6%, ROC-AUC of 0.892,
  • Macro early‑warning accuracy improves to 82.3% compared with prior baselines. performance
    Source sentence
    It raises stock directional accuracy to 67.4% (from 61.7%), credit-risk accuracy to 84.1% (from 79.6%), and macro early-warning accuracy to 82.3%.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 43 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.