RiskLabs: Predicting Financial Risk Using Large Language Model Based on Multi-Sources Data

In one sentence

Summary

The paper addresses the underexplored application of Large Language Models (LLMs) in financial risk prediction, specifically targeting volatility and Value at Risk (VaR). While LLMs have succeeded in text-based finance tasks, their direct use for numerical regression remains challenging. The authors position RiskLabs as a solution that integrates multimodal data—audio and text from Earnings Conference Calls (ECCs), daily news, and market time series—to provide a holistic view of market dynamics, moving beyond simple text summarization or sentiment analysis. The RiskLabs framework employs specialized encoders for each data modality. ECC audio is processed via Wav2vec2 and self-attention, while transcripts are encoded using SimCSE and an LLM-based hierarchical summarization module that extracts key insights via a question bank. News is summarized by an LLM, and time series (VIX) is encoded using a BiLSTM. These features are fused additively and fed into a multi-task learning head to predict volatility across multiple horizons (3, 7, 15, 30 days) and 1-day VaR. Experiments compare RiskLabs against classical methods (GARCH), neural networks (LSTM, HAN, MRDM), and direct LLM prompting (GPT-3.5). Results show RiskLabs outperforms all baselines in MSE for short-to-medium term volatility and achieves the most accurate VaR prediction (0.049 vs. target 0.05), significantly better than the historical method (0.016) and neural networks. Ablation studies confirm that adding LLM-derived analysis and time series data incrementally improves performance. A key finding is that direct LLM prompting for risk prediction is ineffective and potentially hazardous, establishing LLMs as powerful feature extractors rather than standalone predictors. Limitations include challenges with news quality and the need for dynamic windowing for long-term forecasts.

Research problem

The paper investigates the gap in applying Large Language Models (LLMs) to financial risk prediction, specifically forecasting stock volatility and Value at Risk (VaR). It addresses the challenge of using generative LLMs for precise numerical regression and the need to effectively integrate multimodal financial data (text, audio, time series) for robust risk assessment.

Core contributions

• Proposes RiskLabs, a multimodal framework integrating LLMs with audio, text, and time-series data for financial risk prediction.
• Demonstrates that LLMs serve as effective feature extractors for ECCs and news, significantly boosting deep learning models.
• Shows that direct LLM prompting for risk prediction is ineffective, positioning LLMs as assistants rather than standalone predictors.
• Provides empirical evidence that multimodal fusion improves VaR and volatility forecasting over classical and neural baselines.

Method

• ECC Audio Encoding: Wav2vec2 embeddings processed by Multi-Head Self-Attention and average pooling.
• ECC Text Encoding: SimCSE sentence embeddings processed by Multi-Head Self-Attention.
• ECC Analysis: LLM (GPT-4) performs hierarchical summarization and extracts key sentences via a question bank.
• News Encoding: LLM summarizes daily news to generate text features.
• Time-Series Encoding: BiLSTM processes 30-day VIX data to extract temporal features.
• Fusion: Additive fusion of all modality features followed by multi-task regression heads.

Results & findings

• RiskLabs achieves the lowest MSE for 3, 7, and 15-day volatility predictions compared to all baselines.
• VaR prediction of 0.049 is closest to the target 0.05, outperforming the historical method (0.016) and FCNN (0.044).
• Direct GPT-3.5 prompting resulted in high MSE (e.g., 2.198 for 3-day volatility), indicating poor direct regression capability.
• Ablation shows that adding LLM analysis and VIX data progressively improves model performance.
• The framework effectively handles multi-task learning for both volatility and VaR simultaneously.

Data & experiments

• Dataset includes Earnings Conference Call audio and transcripts, daily financial news, and VIX time series data.
• Experiments compare RiskLabs against GARCH, LSTM, HAN, MRDM, HTML, and direct GPT-3.5 prompting.
• Metrics include Mean Squared Error (MSE) for volatility at 3, 7, 15, and 30-day horizons, and VaR prediction accuracy.
• Ablation studies evaluate the contribution of audio, text, LLM analysis, and time series modules.

Limitations & caveats

• Direct LLM prompting for risk prediction is shown to be ineffective and potentially hazardous.
• Performance on 30-day forecasts lags behind some specialized models like HTML.
• News data quality varies, potentially introducing noise into the model.
• The framework requires significant computational resources for multimodal processing.
• Scalability of the LLM-based analysis module to broader datasets is noted as a challenge.

Curator notes

Discovery status: round3_promoted. Source links: FinArena: A Human-Agent Collaboration Framework for Financial Market Analysis and Forecasting; FinCon: A Synthesized LLM Multi-Agent System with Conceptual Verbal Reinforcement for Enhanced Financial Decision Making