Go back to projects

RiskLabs: LLM-based Financial Risk Prediction

#Artificial Intelligence #Large Language Models #Financial NLP #Multimodal Learning #Financial Risk Prediction

Overview

Financial risk prediction requires integrating heterogeneous sources of information, including market time-series data, financial news, earnings conference calls, and corporate disclosures. Traditional quantitative models often struggle to capture the rich semantic information contained in unstructured financial documents.

RiskLabs introduces a multimodal framework that combines Large Language Models (LLMs), deep learning, and financial data analysis to improve the prediction of market volatility and Value-at-Risk (VaR). Rather than relying solely on numerical market indicators, the framework leverages both textual and acoustic financial information to construct more informative representations for downstream prediction tasks.

Motivation

Financial markets generate massive amounts of structured and unstructured information every day. While numerical indicators remain essential, qualitative information contained in corporate disclosures and financial news often provides additional insights into future market behavior.

This project investigates how Large Language Models can bridge traditional quantitative finance and modern AI by integrating financial reasoning, multimodal learning, and retrieval-based knowledge extraction into financial risk prediction.

Methodology

The proposed framework integrates multiple financial information sources into a unified prediction pipeline:

  • Multimodal representation learning from financial news, earnings conference call transcripts, speech recordings, and historical market data.
  • Large Language Models for financial document understanding and structured information extraction.
  • Retrieval-Augmented Generation (RAG) for context-aware financial reasoning.
  • Deep learning models for downstream financial risk prediction, including market volatility and Value-at-Risk estimation.

Instead of treating LLMs as end-to-end predictors, the framework employs them as intelligent financial reasoning modules that complement conventional quantitative modeling techniques.

My Role

This project was developed through close collaboration among all co-authors. My role primarily involved prompt engineering, development of the LLM-based analytical workflow, financial data annotation, and scientific writing. Together with my collaborators, we designed a multimodal framework that integrates large language models, retrieval-based reasoning, and quantitative financial modeling for financial risk prediction.

Results

Experimental evaluation demonstrated that the proposed multimodal framework consistently outperformed conventional baselines across multiple financial risk prediction tasks. By integrating structured market information with LLM-assisted financial reasoning, the framework achieved more accurate predictions while providing richer representations of financial knowledge.

The project was accepted by the International Workshop on Multimodal Financial Foundation Models (MFFMs) at ACM International Conference on AI in Finance (ICAIF 2024) and is also available as an arXiv preprint.

Research Areas

Large Language Models · Financial AI · Financial NLP · Multimodal Learning · Retrieval-Augmented Generation · Quantitative Finance