Event-Type Propagation Graphs: A Lag-Aware Layer for Predictive Financial Analysis

Yuxuan Li^1,* Bohan Zhang^1,* Guo Yu¹ Ye Gao¹ Yixi Zhou^1,2 Haipeng Zhang^1,†

¹ ShanghaiTech University² Hong Kong Baptist University

* Equal contribution† Corresponding author

Paper PDF ↓Code ↗Project repository ↗Citation ↓

CausalLagStock grounds financial news into structured events, learns a sparse global graph of event-type strengths and preferred trading-day delays, and reuses that graph in stock movement prediction and event monitoring.

News events and evidence feed strength and lag matrices, fused with price features for Neural ODE movement prediction — From grounded financial events to a reusable propagation layer for stock movement prediction.

01 / The research question

Abstract

Financial news affects stock movements through delayed event pathways: an inflation report may later shape interest-rate expectations, while litigation news may precede governance changes or analyst reassessment. Existing news-driven predictors usually compress each item into sentiment, text embeddings, or stock-level relations, leaving no reusable representation of which financial event types tend to precede later event- or price-relevant signals and at what trading-day delays. We propose CausalLagStock, a lag-aware event-type propagation framework for predictive financial analysis. It grounds raw news into structured event records with event type, affected ticker, event-impact magnitude, surprise, and novelty evidence; learns a sparse global event-type propagation graph whose directed edges encode lagged predictive associations and preferred delays; and injects the learned graph into downstream predictors, including a graph-enhanced Neural ODE. Experiments on two aligned financial-news stock movement datasets show improved accuracy and weighted-F1 under chronological evaluation. The learned graph also exposes delayed pathways such as inflation-related news preceding interest-rate signals and litigation-related news preceding governance changes, making it useful as an inspectable monitoring layer while balanced prediction of threshold-sensitive minority classes remains challenging.

02 / A reusable event-type layer

Ground. Learn. Predict.

Ground financial events

Schema-constrained LLM records identify event type, affected tickers and impact evidence. FinBERT embeddings, surprise and novelty connect text with market context.

Learn strength and delay

A shared graph over 20 event types encodes directed strength A and preferred trading-day lag T. Sparse lag-aware attention learns from chronological supervision at 1, 3, 5, 10 and 15 days.

III

Reuse the frozen graph

Graph-derived features support a transparent HGB classifier. A graph-enhanced Neural ODE combines delayed event propagation with price histories for next-day Up / Down / Flat prediction.

The graph summarizes temporally ordered predictive associations between financial event types, supporting inspectable monitoring and horizon-aware analysis.

03 / Chronological evaluation

Main results

Results from the paper on FNSPID and CMIN-US. Examples are grouped by timestamp and purged at label-end boundaries. Direction labels use ±0.5% return thresholds.

Selected metrics from the paper’s main results table (mean ± standard deviation).
Dataset	Model	Seeds	Accuracy	Macro-F1	Weighted-F1	MCC
CMIN	StockNet	3	0.3667 ± 0.0186	0.1795 ± 0.0055	0.1978 ± 0.0157	−0.0014 ± 0.0025
CMIN	Causality-driven Transformer	3	0.3448 ± 0.0037	0.3436 ± 0.0036	0.3435 ± 0.0037	0.0175 ± 0.0025
CMIN	CausalLagStock + Neural ODE	5	0.4246 ± 0.0298	0.3172 ± 0.0131	0.4009 ± 0.0122	0.0075 ± 0.0243
FNSPID	StockNet	3	0.3689 ± 0.0177	0.2236 ± 0.0522	0.2509 ± 0.0682	−0.0072 ± 0.0092
FNSPID	Causality-driven Transformer	3	0.3577 ± 0.0009	0.3559 ± 0.0008	0.3579 ± 0.0013	0.0368 ± 0.0008
FNSPID	CausalLagStock + Neural ODE	5	0.4939 ± 0.0131	0.3513 ± 0.0236	0.4698 ± 0.0192	0.0511 ± 0.0313

Results are reported over five seeds for CausalLagStock and three runs for the baselines.

Overall prediction

Accuracy and weighted-F1 improve on both datasets. On CMIN, the full graph reaches 0.4246 accuracy versus 0.4093 without the graph and 0.4096 with a randomized graph.

Remaining challenge

Macro-F1 and MCC gains are not uniform. The minority Flat class remains difficult. On CMIN, HGB offers higher Macro-F1 (0.3319) than the Neural ODE (0.3172), with lower accuracy (0.4089).

04 / Inspectable monitoring patterns

What follows an event—and when?

Learned macro-policy and company-governance pathways, highlighting inflation to interest rates and litigation to executive changes — Representative relations from the manuscript; preferred delays are measured in trading days.

M5 → M1

Inflation → interest rates

6.18 trading days

Strength A = 0.3447. Inflation-related news precedes later rate-related signals over a multi-day monitoring window.

C7 → C3

Litigation → executive changes

2.28 trading days

Strength A = 0.3786. Legal-risk news precedes governance-related signals over a shorter monitoring window.

These learned pathways expose monitoring windows following macroeconomic and company events.

05 / Cite this work

BibTeX

Citation for the accepted paper (provisional proceedings metadata).

@inproceedings{li2026causallagstock,
  title = {Event-Type Propagation Graphs: A Lag-Aware Layer for Predictive Financial Analysis},
  author = {Li, Yuxuan and Zhang, Bohan and Yu, Guo and Gao, Ye and Zhang, Haipeng and Zhou, Yixi},
  booktitle = {2026 IEEE International Conference on Data Mining (ICDM)},
  year = {2026},
  address = {Shenyang, China},
  url = {https://github.com/XanderZhou2022/ICDM2026_CausalLagStock/blob/main/assets/paper.pdf},
  note = {Accepted; provisional citation pending official proceedings}
}

Download citation.bib