A New Method: Dynamic Higher-Order Relations and Event-Driven Modeling for Stock Price Prediction

Paper link: https://www.ijcai.org/proceedings/2025/0673.pdf

Code link: https://github.com/kijeong22/ijcai2025-spf

Background

The global equity market valuation reached about $111 trillion in 2023, making the dependencies and temporal patterns among stock prices increasingly complex. Existing deep‑learning models struggle to capture extreme co‑movements and dynamically evolving relationships, often ignoring simultaneous modeling of price extremes and group‑level dependencies.

Problem Definition

Two key challenges are addressed:

Insufficient modeling of extreme co‑movements (e.g., the sharp simultaneous declines of many stocks in August‑September 2015).

Difficulty capturing dynamic inter‑stock relationships that vary across time and sectors.

Using normalized OHLCV data and technical indicators (MA, MACD, RSI), the task is to predict the return ratio r for a future horizon τ based on historical features X, by learning a function f that minimizes a risk loss L.

Method

3.1 Time‑Dynamic Representation Learning

Feature embedding: A simple fully‑connected layer maps input features into a high‑dimensional latent space, followed by sinusoidal positional encoding ( PE) to preserve temporal order.

In‑stock attention: Self‑attention with a time mask captures temporal dependencies. Multi‑head attention outputs pass through two fully‑connected layers with residual connections to obtain historical context.

3.2 Enhancing and Aggregating Temporal Representations

Modeling Hawkes intensity: A dedicated layer learns the Hawkes intensity function using sliced historical context H as predictor, estimating base intensity μ_{s_t}, excitation factor η_{s_t}, and decay rate γ_{s_t}.

The resulting intensity Λ_{s_t} is combined with the historical context H_s to produce an enhanced temporal representation.

Temporal information aggregation: Temporal attention aggregates information across all time steps, yielding a final embedding for each stock.

3.3 High‑Order Relation Representation Learning

Predefined hypergraph construction: A hypergraph G = (V, E, W) models market relations, where V are stocks, E are hyper‑edges representing industry groups ( E_{Ind}) and correlation‑based groups ( E_{Corr}), and W is a diagonal weight matrix.

Wavelet hypergraph convolution: Using the hypergraph Laplacian, the method diagonalizes it to obtain Fourier bases U, computes wavelet bases, and performs convolution to capture domain‑knowledge‑driven stock correlations.

Dynamic hypergraph structure learning: A low‑rank method builds and learns a dynamic hypergraph whose association matrix evolves over time; the number of dynamic hyper‑edges is k. Dynamic stock representations are obtained via hypergraph convolution.

3.4 Model Output

The final stock representation combines static, predefined, and dynamic embeddings via a weighted sum. A two‑layer MLP predicts the return ratio, with the loss defined as root‑mean‑square error (RMSE).

Experiments

4.1 Experimental Setup

Dataset: Real‑world S&P 500 data filtered to 463 stocks across 10 phases; each phase contains 24 months of training, 4 months of validation, and 8 months of testing.

Evaluation metrics: Prediction metrics – Information Coefficient (IC), Rank‑IC, Precision@K; Portfolio metrics – Return, Sharpe Ratio (SR), Maximum Drawdown (MDD).

Baselines: Transformer‑based models (DTML, MASTER), graph‑based models (RSR, RT‑GCN), and hypergraph‑based models (STHAN‑SR, ESTIMATE).

4.2 Experimental Results

Overall performance: Across most phases, the proposed model outperforms baselines on IC, Rank‑IC, and Prec@10, and achieves higher return, Sharpe ratio, and lower maximum drawdown, demonstrating its ability to capture sudden market moves and build stable portfolios.

Lower‑Tail Dependency Coefficient (LTDC): The model better approximates true LTDC values, successfully capturing relationships such as the near‑identical movements of GOOG and GOOGL and the strong correlation between V and MA, evidencing its capability to model extreme downturns.

Dynamic hypergraph analysis: Examination of the learned dynamic hypergraph shows it captures evolving inter‑stock relations; for example, AAPL and MSFT exhibit strong connections to hyper‑edges 2 and 10 in Oct 2020, weakened after the US election in Nov, while GOOG‑GOOGL and V‑MA maintain consistent hyper‑edge connections across multiple time steps.