OmniGNN: Dynamic Multi-Relational GNN

• The paper introduces OmniGNN, a dynamic graph neural network that fuses metapath-based structural aggregation, graph attention, and Transformer temporal modeling to capture multifaceted financial signals.
• OmniGNN leverages sector nodes organized in a star-topology to propagate macroeconomic shocks effectively, significantly enhancing predictive accuracy during market crises.
• Empirical results demonstrate notable improvements over traditional models, with higher IC, IR, CR, and Precision@K metrics, underscoring its robustness in dynamic financial environments.

OmniGNN is an attention-based, multi-relational dynamic graph neural network (GNN) architecture designed for robust stock prediction in financial markets, with a distinct capability to integrate heterogeneous structural, temporal, and macroeconomic contexts. The model is characterized by its sector (industry) nodes functioning as global intermediaries, enabling efficient message propagation—particularly under macroeconomic shocks. OmniGNN advances the state-of-the-art by fusing metapath-based structural aggregation, graph attention networks, Transformer-based temporal modeling, and explicit macroeconomic context into a unified architecture, as detailed in "Structure Over Signal: A Globalized Approach to Multi-relational GNNs for Stock Prediction" (Li et al., 12 Oct 2025).

1. Architecture: Multi-relational and Dynamic Design

OmniGNN’s architectural core consists of three tightly integrated layers:

• Structural Layer: Encodes inter-stock relationships and sectorial organization using metapaths. Two primary metapaths are:
  • $\mathcal{SS}$: direct stock-to-stock relations (weighted by factors like sector similarity and shared institutional ownership).
  • $\mathcal{SIS}$: stock-to-industry-to-stock relations, leveraging industry nodes for higher-order connectivity.
• Graph Attention Mechanism: Employs an extension of Graph Attention Networks (GAT) with multi-head attention. Each stock node embedding is updated using a mechanism parameterized by node features $h_i \in \mathbb{R}^D$ and edge attributes $e_{ij}$:

  $$\beta_{(ij)} = a^\top (W h_i \mathbin\Vert W h_j \mathbin\Vert W e_{ij})$$

  $$\alpha_{(ij)} = \frac{\exp(\mathrm{LeakyReLU}(\beta_{(ij)}))}{\sum_{k\in\mathcal{N}_i} \exp(\mathrm{LeakyReLU}(\beta_{(ik)}))}$$

  The new embedding:

  $$h_i' = \frac{1}{H} \sum_{k=1}^H \sum_{j \in \mathcal{N}_i} \alpha_{(ij)}^k W^k h_j$$

  This framework allows for heterogeneous node/edge types and weights relations by contextual importance.

• Temporal Layer: Utilizes a Transformer module with Attention with Linear Biases (ALiBi), operating over sequences of node embeddings $$H_{v,t-\Delta t:t} \in \mathbb{R}^{\Delta t \times d_h}$$. The self-attention scores are given by:

  $$Z = \mathrm{softmax}\left(\frac{QK^\top}{\sqrt{d_k}} + m \cdot P + M \right) V$$

  $m \cdot P$ introduces a bias favoring recent information; $M$ enforces causality.

• Prediction Head: A linear layer per stock outputs the predicted excess return at time $t$:

  $f_i(\cdot) = W_i z_{it} + b_i$

  where $z_{it}$ is the temporally encoded embedding.

2. Macroeconomic Context and Sector Nodes

A central innovation of OmniGNN is the incorporation of the industry/sector node ("global node"). Each stock is linked to its sector node by multi-relational edges encoding supply-chain, regulatory, and market-specific information. The sector node forms a star-topology overlay, ensuring every stock is within two hops of this global node. This design enables instantaneous propagation of macroeconomic shocks across all stocks in an industry. Empirical ablations show that including the $\mathcal{SIS}$ metapath (sector node) significantly enhances performance, especially during market crises such as COVID-19.

3. Comparative Performance and Robustness Under Shocks

OmniGNN was systematically benchmarked against GAT and Transformer baselines on technology stock datasets (Jan 2019–Dec 2022). Performance metrics included Information Coefficient (IC), Information Ratio (IR), Cumulative Return (CR), and Precision@K. OmniGNN achieved:

• IC = 0.0673
• IR = 0.0767
• CR = 0.0218
• Precision@K = 0.5266

These metrics surpassed baseline alternatives (GAT, Transformer), with OmniGNN showing strong relative improvement in IC by approximately 2.22 and 1.99 margins and similar advantages in IR and CR. During the COVID-19 shock period, ablation analysis revealed that omitting the sector node metapath ($\mathcal{SIS}$) led to poor performance, while its inclusion recovered and improved IR from negative to 0.0819 and similarly enhanced cumulative returns.

Model	IC	IR	CR	Precision@K
OmniGNN	0.0673	0.0767	0.0218	0.5266
GAT	lower	lower	lower	lower
Transformer	lower	lower	lower	lower

These results confirm the model's superior robustness and predictive accuracy, especially when macroeconomic context is rapidly shifting.

4. Structural and Temporal Message Passing Mechanisms

The star topology introduced by the sector node in OmniGNN facilitates direct information transfer, bypassing the inefficiencies of multi-hop diffusion characteristic of conventional GNNs. The graph attention mechanism ensures that sector and stock-level signals are appropriately weighted in the aggregation process. Temporal modeling, via the Transformer with ALiBi, captures evolving dependencies in time series, biasing toward recent changes—critical for financial modeling where shocks may manifest abruptly.

5. Implications for Benchmarking, Error Analysis, and Extendability

Insights from GraphOmni (Xu et al., 17 Apr 2025), while not explicitly describing OmniGNN, provide methodological guidance for benchmarking and continuous evaluation. A plausible implication is that OmniGNN should utilize robust graph representations (adjacency lists, matrices) and adaptive prompting strategies, as LLM-based models display marked error variability depending on serialization and prompt style. Reinforcement learning-inspired factor selection, as introduced in GraphOmni’s RL-Scale experiments, could be used within OmniGNN to optimize message passing and reasoning under varying scenarios.

Error analysis in related LLM benchmarks underscores the necessity for algorithmic awareness (e.g., distinguishing “diameter” as the maximum shortest path, and appropriately counting cycles or triangles), which can be incorporated into OmniGNN for enhanced reliability. Benchmark datasets and heatmaps from GraphOmni establish a detailed performance baseline that can be used to evaluate OmniGNN in future extensions beyond financial applications.

6. Future Directions in Generalizable Graph Reasoning

The architecture and empirical results suggest OmniGNN's framework is extensible to other domains requiring multi-relational and context-aware information propagation, such as supply chain risk modeling or infectious disease spread in networked populations. The design principle of embedding global nodes for fast shock dissemination can generalize to other hierarchical, multiplex networks. Leveraging hybrid representations and adaptive error-sensitive learning strategies could further enhance OmniGNN's ability to tackle complex graph-theoretic tasks posed in natural language or structured data.

OmniGNN represents a rigorous synthesis of graph neural networks, attention mechanisms, and temporal prediction that jointly prioritize structure over raw signal, resulting in improved predictive fidelity and resilience to systemic perturbations.