TRACE: Temporal Rule-Anchored Chain-of-Evidence on Knowledge Graphs for Interpretable Stock Movement Prediction
Authors: Qianggang Ding, Haochen Shi, Luis Castejón Lozano, Miguel Conner, Juan Abia, Luis Gallego-Ledesma, Joshua Fellowes, Gerard Conangla Planes, +2 more
Organizations: DIRO & Institut Courtois, Université de Montréal, Montreal, QC, Canada · Mila – Québec AI Institute, Montreal, QC, Canada · Aily Labs, Madrid, Spain · Aily Labs, Milan, Italy · Canada CIFAR AI Chair
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 UP/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.
Figures & tables
Figure 1. Structure of the temporal financial knowledge graph showing entity types, relation categories, and temporal validity intervals.
Figure 2. Overview of our knowledge graph reasoning framework showing the integration of automated rule mining and temporal graph reasoning.
Figure 3. Example of interpretable reasoning paths showing how our knowledge graph reasoning framework connects stocks to textual evidence.
Method
Accuracy
Precision
Recall
F1-Score
Historical Momentum
0.517
0.524
0.709
0.603
XGBoost
0.529
0.535
0.678
0.598
News Sentiment (Qwen)
0.524
0.529
0.636
0.594
T-GNN
0.528
0.534
0.685
0.600
Mars
0.539
0.544
0.664
0.598
StockAgent
0.539
0.545
0.648
0.592
Table 1. Performance comparison for 1-day stock movement prediction. Best results in bold .
Configuration
Performance Metrics
Accuracy
Precision
Recall
Full Model
0.551
0.557
0.715
- Temporal Constraints
0.539
0.544
0.534
- Rule Mining
0.542
0.547
0.537
- Multi-hop Reasoning
0.532
0.537
0.527
- Path Aggregation
0.548
0.553
0.543
Table 2. Detailed ablation study showing the contribution of individual components to overall performance.
Method
Total Return
Sharpe
Ann. Vol
Max DD
Win Rate
Equal-Weight All
16.7%
1.17
14.1%
-12.9%
52.8%
Historical Momentum
8.7%
0.62
15.4%
-17.3%
54.4%
XGBoost
27.0%
1.31
19.9%
-14.8%
53.2%
News Sentiment (Qwen)
9.2%
0.67
14.8%
-12.2%
52.8%
T-GNN
3.5%
0.33
13.3%
-12.5%
51.2%
Mars
14.2%
0.98
14.7%
-17.3%
54.8%
Table 3. Top-10 buy&hold portfolio metrics (2023-01-01 to 2024-01-01). Each method forms a long-only, equal-weight basket of its Top-10 tickers without rebalancing. Best values in bold .
Figure 4. Cumulative returns of Top-10 buy&hold portfolios for all methods (2023-01-01 to 2024-01-01). Our approach dominates over the full horizon and exceeds the equal-weight benchmark.
Figure 5. Counterfactual Analysis with different deletion ratios for Mask-Tex and Mask-Edge strategies, respectively.
Department of Industrial Data Engineering, Hanyang University, Republic of Korea · School of Interdisciplinary Industrial Studies, Hanyang University, Republic of Korea