TRACE: Temporal Rule-Anchored Chain-of-Evidence on Knowledge Graphs for Interpretable Stock Movement Prediction

TL;DR

TRACE combines knowledge graphs and LLMs for interpretable stock prediction, achieving 55.1% accuracy on S&P 500.

cs.CE 🔴 Advanced 2026-03-13 3 views
Qianggang Ding Haochen Shi Luis Castejón Lozano Miguel Conner Juan Abia Luis Gallego-Ledesma Joshua Fellowes Gerard Conangla Planes Adam Elwood Bang Liu
knowledge graph stock prediction rule mining multi-hop reasoning large language models

Key Findings

Methodology

TRACE uses knowledge graphs for stock prediction, integrating symbolic relational priors, dynamic graph exploration, and LLM decision-making. Key components include rule-guided multi-hop exploration and text-grounded evidence consolidation.

Key Results

  • On the S&P 500 benchmark, TRACE achieved 55.1% accuracy, 55.7% precision, 71.5% recall, and 60.8% F1 score, surpassing strong baselines.
  • Rule-guided exploration focuses on economically meaningful motifs rather than arbitrary paths, improving recall and F1 score.
  • Text-grounded consolidation selectively aggregates high-confidence hypotheses, avoiding uniform pooling of weak signals.

Significance

TRACE provides an interpretable stock prediction method by combining knowledge graphs and LLMs, addressing limitations in existing models regarding heterogeneous information source integration and interpretability.

Technical Contribution

TRACE combines symbolic relational priors with dynamic graph exploration, offering new theoretical guarantees and engineering possibilities, surpassing current SOTA methods.

Novelty

TRACE is the first to combine knowledge graphs with LLMs for stock prediction, providing a unified methodology distinct from existing embedding-based models.

Limitations

  • The model may perform poorly in extreme market fluctuations, as these scenarios may not fit predefined rules.
  • Requires extensive historical data for rule mining, potentially limiting application in emerging markets.

Future Work

Future work can explore applying TRACE to broader markets and improving real-time adaptability and responsiveness.

AI Executive Summary

TRACE is a stock prediction method combining knowledge graphs and large language models. Existing methods often fail to effectively integrate heterogeneous information sources and lack interpretability. TRACE provides a unified prediction framework through symbolic relational priors and dynamic graph exploration.

The method performs excellently on the S&P 500 benchmark, achieving 55.1% accuracy and providing interpretable prediction paths. Through rule-guided exploration and text-grounded consolidation, TRACE selectively aggregates high-confidence hypotheses, avoiding uniform pooling of weak signals.

Although TRACE excels in current benchmarks, it may perform poorly in extreme market fluctuations. Future work can focus on improving real-time adaptability and responsiveness, and applying TRACE to broader markets.

Deep Analysis

Background

Stock prediction has long been a central problem in computational finance and data mining. Traditional approaches rely on time-series models, while modern methods incorporate textual sentiment from news and social media. However, these methods typically treat heterogeneous information sources independently, lacking modeling of relational dependencies among companies, events, and products.

Core Problem

Existing models are limited in integrating heterogeneous information sources and often lack interpretability. Financial markets are inherently relational, with a single stock's behavior influenced by a complex web of interactions beyond its fundamentals or sentiment.

Innovation

TRACE offers a unified prediction framework by combining knowledge graphs and LLMs. Core innovations include symbolic relational priors and dynamic graph exploration, ensuring reasoning paths reflect economically meaningful dependencies.

Methodology

  • �� Symbolic relational priors: Extract relational motifs from historical knowledge graph snapshots.
  • �� Dynamic graph exploration: Design guided exploration procedure to traverse the financial knowledge graph.
  • �� Text-grounded consolidation: Score paths using a multi-factor function combining rule coverage, temporal recency, and structural diversity.

Experiments

Experiments were conducted on 2022-2023 S&P 500 data, constructing a large-scale financial knowledge graph with 42K nodes and 174K edges. The experiments demonstrate TRACE's advantages in accuracy and interpretability.

Results

TRACE achieved 55.1% accuracy on the S&P 500 benchmark, surpassing strong baselines. Through rule-guided exploration and text-grounded consolidation, TRACE selectively aggregates high-confidence hypotheses.

Applications

TRACE can be used for real-time stock prediction, helping investors make more informed decisions. Its interpretable paths can be used for auditing and compliance analysis.

Limitations & Outlook

TRACE may perform poorly in extreme market fluctuations, as these scenarios may not fit predefined rules. Requires extensive historical data for rule mining, potentially limiting application in emerging markets.

Plain Language Accessible to non-experts

Imagine a complex traffic network, TRACE is like a smart navigation system. It not only tells you how to reach your destination but also explains why it chose that route. By combining historical traffic data and real-time conditions, TRACE can predict the best route and provide detailed explanations.

ELI14 Explained like you're 14

Imagine you're playing a complex strategy game, TRACE is like a super helper. It not only tells you what move to make next but also explains why. By combining historical game data and current situations, TRACE helps you make smart decisions and win the game!

Glossary

Knowledge Graph

A structured data model representing entities and their relationships.

Used in TRACE to represent relationships among companies, events, and products.

Rule Mining

Extracting frequently occurring patterns or rules from data.

Used to extract relational motifs of financial interactions.

Multi-hop Reasoning

Reasoning through multiple intermediate nodes to connect start and target nodes.

Used to explore relationships between stocks and events in the knowledge graph.

Large Language Model

A deep learning model capable of understanding and generating natural language.

Used in TRACE for decision-making and text-grounded consolidation.

Text Grounding

Associating structured data with textual data to provide evidence support.

Used to link reasoning paths with news or reports.

Open Questions Unanswered questions from this research

  • 1 How to improve TRACE's performance in extreme market fluctuations? Existing rules may not apply, requiring new methods.
  • 2 How to reduce reliance on extensive historical data? Emerging markets have limited data, necessitating exploration of new data sources.

Applications

Immediate Applications

Real-time Stock Prediction

TRACE can be used for real-time stock prediction, helping investors make more informed decisions. Requires historical data and news as input.

Long-term Vision

Market Analysis Tool

TRACE can transform into a comprehensive market analysis tool, helping identify market trends and risks. Requires solving data dependency and model adaptability issues.

Abstract

We present a Temporal Rule-Anchored Chain-of-Evidence (TRACE) on knowledge graphs for interpretable stock movement prediction that unifies symbolic relational priors, dynamic graph exploration, and LLM-guided decision making in a single end-to-end pipeline. The approach performs rule-guided multi-hop exploration restricted to admissible relation sequences, grounds candidate reasoning chains in contemporaneous news, and aggregates fully grounded evidence into auditable \texttt{UP}/\texttt{DOWN} verdicts with human-readable paths connecting text and structure. On an S\&P~500 benchmark, the method achieves 55.1\% accuracy, 55.7\% precision, 71.5\% recall, and 60.8\% F1, surpassing strong baselines and improving recall and F1 over the best graph baseline under identical evaluation. The gains stem from (i) rule-guided exploration that focuses search on economically meaningful motifs rather than arbitrary walks, and (ii) text-grounded consolidation that selectively aggregates high-confidence, fully grounded hypotheses instead of uniformly pooling weak signals. Together, these choices yield higher sensitivity without sacrificing selectivity, delivering predictive lift with faithful, auditably interpretable explanations.

cs.CE cs.AI