Paper Number

ICIS2026-3160

Paper Type

Completed

Abstract

This study develops a risk-aligned financial decision support system for LLM-based stock prediction. Existing LLMs can process market information but often fail to accommodate heterogeneous investor risk preferences because preference-annotated financial data are scarce. We propose the Distillation–Attribution–Prediction (DAP) framework, which uses realized market outcomes as factual anchors and self-reflection to generate risk-preference annotations for training a lightweight LLM that produces predictions and interpretable rationales. Using Chinese industry ETF data, expert evaluation shows that the DAP-trained LLM’s generated reasoning is well aligned with different risk preferences, including both risk-averse and risk-seeking settings. The DAP-trained LLM also achieves the highest prediction accuracy (0.5913) and Matthews Correlation Coefficient (0.1725) among all evaluated models. Its preference-specific strategies outperform LLM and rule-based baselines and remain competitive with machine learning models across multiple market regimes. These findings demonstrate the value of behaviorally grounded LLM design for personalized financial decision support under uncertainty.

Comments

10_Computational_design

Share

COinS
 
Dec 14th, 12:00 AM

Can LLMs think like different investors? A self-reflective framework for risk-aligned stock prediction

This study develops a risk-aligned financial decision support system for LLM-based stock prediction. Existing LLMs can process market information but often fail to accommodate heterogeneous investor risk preferences because preference-annotated financial data are scarce. We propose the Distillation–Attribution–Prediction (DAP) framework, which uses realized market outcomes as factual anchors and self-reflection to generate risk-preference annotations for training a lightweight LLM that produces predictions and interpretable rationales. Using Chinese industry ETF data, expert evaluation shows that the DAP-trained LLM’s generated reasoning is well aligned with different risk preferences, including both risk-averse and risk-seeking settings. The DAP-trained LLM also achieves the highest prediction accuracy (0.5913) and Matthews Correlation Coefficient (0.1725) among all evaluated models. Its preference-specific strategies outperform LLM and rule-based baselines and remain competitive with machine learning models across multiple market regimes. These findings demonstrate the value of behaviorally grounded LLM design for personalized financial decision support under uncertainty.

When commenting on articles, please be friendly, welcoming, respectful and abide by the AIS eLibrary Discussion Thread Code of Conduct posted here.