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

The TRACE framework introduces an interpretable, end-to-end pipeline for stock movement prediction that combines rule-guided multi-hop exploration on knowledge graphs with LLM-driven, news-grounded evidence aggregation to achieve superior recall and F1 scores on the S&P 500 while providing auditable, human-readable explanations.

The Gist

Imagine you are trying to predict whether a specific stock will go UP or DOWN tomorrow.

Most traditional methods are like a weather forecaster who only looks at the temperature of the last hour. They look at the stock's past price or read a single news headline and make a guess. Sometimes they get it right, but they often miss the bigger picture because they don't understand how different companies are connected.

Other methods use powerful AI (Large Language Models) that can read everything, but they sometimes "hallucinate"—making up facts or connecting dots that don't actually exist, like a detective who invents clues to solve a case.

TRACE is a new system that acts like a super-sleuth detective who never makes things up and always follows a strict rulebook. Here is how it works, broken down into simple concepts:

1. The Giant Family Tree (The Knowledge Graph)

Imagine the entire stock market as a massive, living family tree.

• The Nodes: Instead of just people, the "family members" are companies, products, news articles, and events.
• The Connections: These family members are linked by relationships. For example, "Company A buys Company B," "Company C sues Company D," or "News Article E mentions Company F."
• The Time Machine: This isn't just a static tree; it's a time machine. It only shows you the connections that existed up until the moment you are making your prediction. It strictly forbids looking at tomorrow's news today (no cheating!).

2. The Rulebook (Symbolic Priors)

Before the detective starts investigating, they study a Rulebook written by experts and mined from history.

• Example Rule: "If a company invests in an AI startup, and that startup is in the AI sector, then the investor's stock usually goes UP."
• Instead of wandering aimlessly through the family tree, the detective only follows paths that match these proven rules. This stops them from wasting time on dead ends or irrelevant gossip.

3. The Investigation (Chain-of-Evidence)

When the system needs to predict the movement of a specific stock (let's say, Microsoft), it doesn't just guess. It goes on a hunt:

1. Start: It starts at Microsoft.
2. Follow the Clues: It looks for connections that match the Rulebook. Maybe it finds Microsoft acquired a gaming company.
3. Check the Evidence: It doesn't stop there. It follows the link to the News Article that announced the acquisition.
4. The "Grounding" Step: This is crucial. The system asks an AI: "Does this news article actually support the idea that Microsoft's stock will rise?" If the news is positive and matches the rule, the clue is "grounded" in reality. If the news is vague or negative, the clue is discarded.

4. The Verdict (The Decision)

After gathering a few strong, rule-based, news-backed clues, the system makes a decision.

• It doesn't just say "Up." It says: "UP, because Microsoft bought a gaming company (Rule), and the news says the market loves this deal (Evidence)."
• It provides a clear, readable path showing exactly why it made that choice. You can trace the logic from the final decision all the way back to the original news article.

Why is this better?

• No Hallucinations: Because it forces the AI to stick to the "Rulebook" and the actual "News Articles," it can't invent fake connections.
• It Understands Context: It knows that if a company in the AI sector has a problem, it might hurt their partners, not just the company itself. It sees the web of connections.
• Trustworthy: In finance, you need to know why a decision was made. TRACE gives you the "receipts" (the news and the rules) so you can audit the decision yourself.

The Result

When they tested this "Detective" on the S&P 500 (the top 500 US companies), it was more accurate than the other methods.

• It correctly predicted the direction of stocks 55.1% of the time (which is a big deal in finance).
• More importantly, it was much better at catching the "UP" moves (high Recall) without getting tricked by fake signals.

In short: TRACE is like giving a financial analyst a super-powered flashlight (the Knowledge Graph), a strict law book (the Rules), and a fact-checker (the AI), ensuring they only make predictions based on solid, traceable evidence.

Technical Summary

1. Problem Statement

The paper addresses the challenge of interpretable stock movement prediction (specifically predicting whether a stock will go UP or DOWN the next trading day). Existing approaches suffer from two main limitations:

1. Lack of Relational Modeling: Traditional time-series or text-based models often treat stocks independently, ignoring the complex web of inter-firm dependencies (e.g., supply chains, partnerships, mergers) that propagate risk and information.
2. Lack of Interpretability & Hallucination: While Large Language Models (LLMs) are powerful, they often hallucinate unsupported facts and provide "black box" predictions, which is risky in high-stakes financial domains.
3. Gap in Integration: Current financial forecasting models rarely integrate explicit graph reasoning with unstructured text (news) in a unified, temporally valid framework.

The goal is to create a system that not only predicts stock movements with high accuracy but also provides auditable, human-readable reasoning paths connecting the prediction to specific news articles and structured relations.

2. Methodology: The TRACE Framework

TRACE (Temporal Rule-Anchored Chain-of-Evidence) is an end-to-end pipeline that unifies symbolic relational priors, dynamic graph exploration, and LLM-guided decision-making on a Temporal Knowledge Graph (TKG).

A. Knowledge Graph Construction

The authors constructed a large-scale financial TKG from 2022–2023 S&P 500 data:

• Scale: ~42,000 nodes and ~174,000 edges.
• Entity Types: TextSource (news/filings), Event, Product, Financial, Company, and Person.
• Relations: Includes EXTRACTED_FROM (anchoring text to entities), INVESTED_IN, PARTNERED, ACQUIRED, CAUSED_INCREASE/DECLINE, etc.
• Temporal Validity: Strict "as-of" constraints ensure that only facts and edges valid up to the prediction time $t$ are accessible, preventing look-ahead bias.

B. Automated Rule Mining

Before prediction, the system mines relational motifs from historical graph snapshots to create a rule bank ($R^*$).

• Process: Uses breadth-first enumeration to find 2–4 hop paths.
• Rule Format: A body (sequence of relations) implies a head (prediction direction: UP/DOWN).
  • Example: INVESTED_IN(X, Z) ∧ IN_SECTOR(Z, "AI") ⇒ predict_direction(X, UP).
• Filtering: Rules are filtered by support and confidence (threshold $\tau_{mine} = 0.60$). These rules act as symbolic priors to constrain the search space during reasoning.

C. Rule-Guided Graph Exploration

For a target stock on day $t$, the system performs an iterative beam search on the TKG:

1. Initialization: Starts from the stock node, respecting the "as-of" time constraint.
2. Rule-Grounded Expansion: The search is restricted to relation sequences that are prefixes of mined rules. This prevents arbitrary graph walks and focuses on economically meaningful motifs.
3. LLM-Guided Pruning: An LLM (Qwen QwQ-32B) acts as a "Relation Selector" to evaluate candidate edges based on semantic relevance to the current context and news snippets, filtering out noise.
4. Path Scoring: Paths are ranked lexicographically based on:
   • Completion of high-confidence rules.
   • Rule body coverage.
   • Temporal recency of evidence.
   • Penalty for "hub" nodes (to avoid generic, high-degree connections).
5. Hypothesis Generation: When a path completes a high-confidence rule and reaches a TextSource node (news/filing), it forms a label-predictive hypothesis.

D. Verdict Prediction & Aggregation

• Aggregation: The system aggregates the highest confidence scores from all valid hypotheses (UP vs. DOWN).
• LLM Validation: A second LLM pass validates the consistency between the reasoning path and the actual text content of the news source.
• Output: A binary verdict (UP/DOWN) accompanied by the top-$M$ supporting reasoning paths, providing a transparent "Chain-of-Evidence."

3. Key Contributions

1. Unified Framework: Introduced TRACE, the first framework to integrate symbolic path mining, dynamic graph exploration, and LLM-guided decision-making specifically for temporal financial prediction.
2. Rule-Grounded Reasoning: Designed a mechanism where mined relational motifs constrain admissible paths, ensuring that reasoning chains are both economically relevant and temporally valid.
3. Interpretability: The system produces explicit, auditable reasoning paths connecting stock movements to specific news articles and causal events, bridging the gap between performance and transparency.
4. Large-Scale Dataset: Constructed and utilized a comprehensive financial TKG (42K nodes, 174K edges) derived from S&P 500 data.

4. Experimental Results

The model was evaluated on the S&P 500 (2022–2023) with a 1-day prediction horizon.

Predictive Performance

TRACE outperformed strong baselines including XGBoost, Temporal GNNs (T-GNN), LLM-only sentiment models, and agent-based frameworks.

• Accuracy: 55.1% (vs. 52.9% for XGBoost, 52.8% for T-GNN).
• Precision: 55.7%.
• Recall: 71.5% (Significant improvement over baselines, indicating better detection of positive movements).
• F1-Score: 60.8%.

Backtesting (Economic Value)

A Top-10 buy-and-hold portfolio strategy (no rebalancing) was tested from Jan 2023 to Jan 2024:

• Total Return: 41.7% (vs. 27.0% for XGBoost and 16.7% for the equal-weight benchmark).
• Sharpe Ratio: 2.00 (vs. 1.31 for XGBoost).
• Max Drawdown: -11.7% (Lower risk than most baselines).

Ablation & Analysis

• Component Impact: Multi-hop reasoning provided the largest gain (+2.1%), followed by temporal constraints (+1.4%) and rule mining (+1.1%).
• Interpretability: 85% of predictions were supported by at least one complete path; 72% of paths matched mined rule patterns.
• Counterfactual Analysis: Removing the textual evidence (Mask-Text) or critical relations (Mask-Edge) caused a monotonic decline in accuracy, proving that the predictions rely on the surfaced evidence chains rather than spurious correlations.

5. Significance

• Solving the "Black Box" Problem: TRACE demonstrates that high-performance financial prediction does not require sacrificing interpretability. It provides a "Chain-of-Evidence" that allows human analysts to audit why a prediction was made.
• Relational Dynamics: By explicitly modeling the propagation of events (e.g., an acquisition affecting suppliers and competitors) through a knowledge graph, the model captures market dynamics that isolated time-series models miss.
• Robustness: The combination of symbolic rules (to reduce search space and hallucination) with LLMs (for semantic understanding and pruning) creates a robust system that is sensitive to market signals without sacrificing selectivity.
• Practical Application: The significant outperformance in backtesting suggests that the "explainable" signals captured by TRACE are economically valuable and actionable for investment strategies.