# Graph Neural Networks for Modeling Inter-Asset Correlations

The global financial ecosystem operates as a highly coupled, non-linear network of institutions, markets, and sovereign entities. Traditional predictive models for systematic trading have historically relied on time-series analysis of isolated variables or static, linear correlation matrices. However, these paradigms frequently fail to capture the dynamic relational interdependencies and topological shifts inherent in modern interconnected markets, particularly during periods of structural breaks or systemic crises [cite: 1, 2]. Graph Neural Networks (GNNs) have emerged as a foundational computational architecture for cross-market financial prediction, systemic risk modeling, and portfolio optimization. By representing financial assets as nodes and their complex interrelationships as edges, GNNs operate directly on non-Euclidean graph structures to capture both spatial dependencies and temporal dynamics [cite: 2, 3]. 

## Theoretical Foundations of Network Modeling

The transition from univariate or simple multivariate statistical forecasting to graph-based modeling represents a fundamental paradigm change in quantitative finance. Asset prices do not evolve in isolation; rather, they are continuously influenced by the diffusion of information across supply chains, industry sectors, corporate ownership networks, and macroeconomic regimes [cite: 4, 5]. 

Classical econometric approaches, such as the Autoregressive Integrated Moving Average (ARIMA) and Generalized Autoregressive Conditional Heteroskedasticity (GARCH) models, have historically served as the bedrock of financial forecasting. While ARIMA is effective for stationary series and GARCH successfully captures volatility clustering, both models are constrained by linear assumptions and struggle with the extreme non-stationarity and abrupt fluctuations characteristic of financial markets [cite: 1, 6, 7]. These models fundamentally treat financial data as Euclidean sequences, processing the history of a single asset without explicitly modeling cross-asset contagion or higher-order structural relationships [cite: 2].

Subsequent deep learning architectures, such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks, improved predictive accuracy by extracting non-linear temporal patterns [cite: 3, 8]. However, LSTMs primarily process data as isolated sequences and fail to natively model the cross-asset correlations and linkage effects that drive the market [cite: 3, 9]. When turbulence impacts one security, it propagates rapidly through sector peers, correlated positions, and broader market relationships [cite: 10]. Univariate models cannot directly capture this propagation, meaning they fail to condition an asset's forecast on the simultaneous behavior of its surrounding topological neighborhood. GNNs address this limitation by propagating information across the network, enabling each node to be aware of its structural context and allowing individual asset predictions to benefit from signals embedded within related securities [cite: 11, 12].

## Graph Construction Mechanisms

The empirical efficacy of any Graph Neural Network is highly dependent on the quality, density, and formulation of its underlying graph. In financial applications, graphs are constructed using several distinct methodologies to encode explicit and implicit relationships between assets, defining the exact topology over which message-passing algorithms operate.

Correlation-based graphs represent the most common dynamic construction. This mechanism tracks live volatility and price co-movement by computing the rolling Pearson return correlation between assets. Nodes are typically connected by an undirected edge if their correlation coefficient exceeds a specified threshold, such as an absolute value of 0.30 or 0.70 depending on the desired sparsity [cite: 3, 10]. This graph is highly dynamic and is often recomputed on a weekly or daily basis as the rolling window advances. Consequently, its density fluctuates significantly based on overarching market regimes. Empirical observations indicate that these graphs remain relatively sparse during calm macroeconomic periods but approach full connectivity during systemic shocks, reflecting the phenomenon where cross-asset correlations converge unpredictably during crises [cite: 6, 10]. While highly responsive to market sentiment, correlation graphs risk encoding spurious statistical noise as structural relationships.

To mitigate the noise inherent in price-derived networks, researchers utilize knowledge-based and sector graphs. These topologies connect assets sharing fundamental economic relationships, such as Global Industry Classification Standard (GICS) sector membership [cite: 10]. Sector graphs are remarkably stable, updated perhaps only once a year to reflect historical reclassifications, and they encode long-term fundamental relationships rather than short-term price co-movements [cite: 10]. Similarly, Granger-causality graphs offer a directed, semi-static alternative. In this construction, a directed edge connects two stocks if the historical returns of the source stock contain statistically significant predictive information about the target stock's returns, beyond what is contained in the target's own history [cite: 10]. 

The integration of alternative data has facilitated the construction of highly granular supply chain and corporate ownership networks. Supply chain datasets, such as those maintained by Bloomberg covering over 100,000 global companies, map customer-supplier relationships as directed edges [cite: 5]. This specific topology allows models to exploit lead-lag momentum effects, operating on the economic premise that an earnings shock to a customer company will predictably, albeit gradually, impact its suppliers [cite: 5, 13]. Furthermore, ownership ties and board overlaps provide additional layers of interconnectedness that are critical for predicting corporate default risk and modeling upstream contagion paths that stand-alone financial metrics completely obscure [cite: 14]. 

Recent advancements have also introduced dynamic graph generation via Natural Language Processing (NLP) and Large Language Models (LLMs). By analyzing daily financial news, platforms can identify co-mentions of corporate entities and extract implied relationships that extend far beyond predefined supply chains or shared business services [cite: 11, 15]. These LLM-inferred networks dynamically update edge weights based on real-time sentiment analysis, enabling the GNN to process breaking news, geopolitical events, and sentiment shifts as structural modifications to the market's topology [cite: 11].

[image delta #1, 0 bytes]





## Spatio-Temporal Graph Neural Network Architectures

To effectively process the multidimensional nature of financial data, GNNs must handle both the spatial topology of the asset network and the temporal evolution of individual asset prices. This dual requirement has driven the development of Spatio-Temporal Graph Neural Networks (ST-GNNs), which fuse convolutional or recurrent temporal layers with graph-based spatial aggregation modules.

### Capturing Temporal and Relational Dynamics

Modern forecasting architectures, such as the Financial Spatio-Temporal Graph Attention Network (FSTGAT) and the Spatio-Temporal Graph Attention Network (STGAT), utilize a hierarchical sandwich structure—typically alternating between time-convolution, graph-attention, and subsequent time-convolution layers [cite: 1, 3]. This cascading block design extracts underlying spatio-temporal patterns in the initial layers before capturing higher-order market dynamics closer to the output.

The temporal dimension is rigorously managed to prevent data leakage. Traditional predictive models can inadvertently incorporate future data points in a sequence, invalidating backtesting results. To resolve this, advanced ST-GNNs employ Gated Causal Convolution. Causal convolution strictly adheres to temporal causality, ensuring that the output at a given time step is calculated exclusively using historical and current inputs [cite: 1, 3]. This temporal extraction is frequently paired with a Gated Linear Unit (GLU). The GLU utilizes a sigmoid-driven gating signal to regulate information flow; by taking the Hadamard product of the causal convolution output and the gating coefficient, the model selectively filters relevant temporal features while actively suppressing transient market noise [cite: 1, 3].

Concurrently, the spatial dimension relies on attention mechanisms, primarily the Graph Attention Network (GAT). Advanced iterations, such as GATv2, are deployed to learn dynamic inter-stock relationships [cite: 3]. GATv2 addresses the "masking bias" prevalent in earlier attention networks by redesigning the attention weight calculation, rendering it more expressive and ensuring a fairer assessment across all connected nodes [cite: 3]. This allows the network to assign varying importance weights to neighboring assets based on their current hidden states, enabling the model to distinguish between genuine, enduring industry-driven relationships and temporary sentiment spillover [cite: 3]. Furthermore, explicit edge attributes—such as industry classification priors or specific Pearson correlation coefficients—are often injected directly into the spatial layer to guide the attention mechanism with domain knowledge [cite: 3]. 

To capture the macro-structure of the market, models increasingly implement Multiple-Input-Multiple-Output (MIMO) architectures. A MIMO framework groups stocks by their respective industry sectors, facilitating the simultaneous learning of intra-group synergistic effects (how stocks within the same sector move together) and inter-group influence spillovers (how a shock in the energy sector impacts the financial sector) [cite: 3]. Empirical evaluations conducted on specific sectors, such as the New York Stock Exchange (NYSE) commercial banking and metals sectors, demonstrate that these spatio-temporal architectures significantly outperform traditional benchmarks. In high-volatility scenarios, where standard models fail, architectures like FSTGAT have been shown to reduce Root Mean Square Error (RMSE) prediction metrics by 45% to 69% compared to ensemble learners like XGBoost and recurrent networks like LSTM [cite: 3].

### Managing Non-Stationarity and Regime Shifts

A persistent and formidable challenge in financial time-series forecasting is non-stationarity. Financial markets are subject to both temporal distribution shifts and spectral variability. During regime shifts—such as sudden transitions from accommodative to restrictive monetary policy, or the onset of global pandemics—the underlying data-generating processes alter dramatically. Consequently, models trained heavily on historical data structures frequently suffer from catastrophic performance degradation in production.

To address spatio-temporal distribution shifts within dynamic graphs, researchers have proposed disentangled attention networks. Architectures such as the Disentangled Intervention-based Dynamic graph Attention network (DIDA) are designed specifically to discover and utilize invariant patterns [cite: 16]. Invariant patterns represent specific graph structures and node features whose predictive abilities remain statistically stable across varying market distributions. By employing principles of causal inference theory alongside spatio-temporal intervention mechanisms, the DIDA model is mathematically regularized to focus strictly on these invariant patterns, rather than fitting to variant patterns that correlate with the target variable only during specific historical regimes [cite: 16].

Non-stationarity also manifests heavily in the frequency domain, where fundamental trend and seasonal patterns experience spectral shifts over time. To combat this, dual-branch frameworks have been introduced that operate simultaneously in the temporal and frequency domains. These models utilize non-stationary Mixture of Experts (MoE) filters to dynamically extract and remove non-stationary temporal patterns, establishing a stationary residual for dependency modeling. Complementing this, the frequency branch applies differencing to highlight components with the most significant spectral changes, ensuring the overall architecture remains robust even when market cyclicality breaks down [cite: 17].

## Comparative Assessment Against Traditional Models

Systematic trading models and risk management systems rely entirely on the accuracy of their dependency estimations. GNNs offer substantial theoretical and empirical advantages over traditional statistical methods by relaxing rigid linear assumptions and providing immense structural depth.

### Linear Correlation and Dimensionality Reduction

Pearson correlation remains the industry standard for measuring the linear dependence between asset returns. However, it is fundamentally flawed when applied to complex financial ecosystems; it fails to capture non-linear relationships, higher-order dependencies, and structural economic connections (such as supply chains) that do not immediately manifest in price co-movements [cite: 18, 19]. Similarly, Principal Component Analysis (PCA) is frequently utilized by quantitative analysts to reduce dimensionality and identify latent market factors. While efficient, PCA assumes a stationary, linear combination of assets, rendering it brittle during structural breaks [cite: 20].

Graph Neural Networks bypass these limitations by dynamically weighting asset relationships through learned attention mechanisms rather than static historical covariance. Unlike a static matrix, a GAT can amplify critical edges during specific market shocks—for instance, algorithmically strengthening the modeled linkage between energy producers and regional financial institutions during an unexpected oil price collapse—while suppressing irrelevant background noise [cite: 20].

### Copula Methods and the Emergence of CopulaGNNs

In the domain of quantitative risk management, Copulas serve as the premier statistical tool for dependency modeling. Copulas function by decoupling the marginal distributions of individual assets from their overarching dependence structure [cite: 18, 19]. This extreme flexibility allows copula models to capture complex, non-linear dependencies, including asymmetric tail dependence. Asymmetric tail dependence refers to the empirical reality that financial assets tend to correlate much more strongly during severe market crashes than they do during expansive bull markets—a phenomenon that linear correlation completely fails to capture, often leading to severe underestimates of portfolio risk [cite: 18, 19].

While copulas are mathematically rigorous and highly effective for modeling tail dependence, they fundamentally lack the deep feature extraction and representation learning capabilities inherent to neural networks. Conversely, standard GNNs excel at representation learning—constructing robust node features based on network topology—but they often fail to effectively utilize the correlational information inherent in the graph regarding node outcomes [cite: 21].

To bridge this substantial methodological gap, recent research has introduced the Copula Graph Neural Network (CopulaGNN). This architecture utilizes standard GNN models as a base but integrates the mathematical principles of copulas to explicitly describe the dependence among multivariate random variables [cite: 21]. By leveraging approximations based on the generalized quantile transform for discrete random variables, CopulaGNNs smooth step functions into piece-wise linear functions, making complex probability mass functions computationally tractable [cite: 22]. This integration ensures that the model utilizes both the representational role of graphs (feature construction) and the correlational role (outcome dependence), providing a unified framework that significantly outperforms standard GNNs in complex regression tasks [cite: 21, 22].

| Feature Dimension | Pearson Correlation / PCA | Copula Models | Standard Graph Neural Networks (GNNs) | CopulaGNN Architectures |
| :--- | :--- | :--- | :--- | :--- |
| **Primary Function** | Linear dependence measurement and dimensionality reduction. | Modeling joint probability distributions and isolating tail dependence. | Deep representation learning and spatial-temporal feature extraction. | Fusing representation learning with rigorous outcome dependence probability modeling. |
| **Linearity Assumption** | Strictly assumes linear relationships between variables. | Highly flexible; can model complex, non-linear, and asymmetric dependencies. | Captures highly non-linear, complex multi-agent interactions via neural layers. | Captures highly non-linear dependencies via combined neural and copula algorithms. |
| **Topology Awareness** | Ignored; treats assets as a flat, unstructured matrix. | Ignored; focuses purely on statistical distributions independent of network structure. | High; explicitly models and aggregates data over the structural network of assets. | High; models structural networks while simultaneously preserving statistical dependence features. |
| **Handling Extreme Events** | Poor; significantly underestimates fat tails and systemic risks. | Excellent; specifically designed for tail risk and extreme value scenarios. | Strong; adapts dynamically via learned attention mechanisms during shocks. | Exceptional; combines dynamic graph attention with copula-driven tail probability modeling. |

## Applications in Portfolio Optimization

The integration of GNNs into portfolio optimization algorithms directly addresses two of the most persistent barriers to robust portfolio construction: modeling dynamic asset correlations during market turbulence, and mitigating the transaction costs that heavily erode systematic trading profits [cite: 6, 20].

### Dynamic Weighting and Asset Allocation

Traditional optimization frameworks, such as the Markowitz mean-variance model, rely heavily on historical covariance matrices to allocate capital. Because these matrices are backward-looking and assume normality, they frequently break down during periods of high volatility, leading to suboptimal allocations precisely when risk management is most critical. GNNs, particularly when deployed within a Deep Reinforcement Learning (DRL) framework, dynamically optimize asset allocation by modeling evolving non-linear dependencies. 

A systematic review of deep learning solutions in portfolio optimization between 2018 and 2025 demonstrates that GNN-driven portfolios consistently achieve 15% to 30% higher Sharpe ratios than traditional methods [cite: 6, 20, 23]. This outperformance is driven by the network's ability to dynamically weight asset relationships, effectively anticipating shifts in correlation before they are fully priced into the historical covariance matrix [cite: 6, 20]. 

### Transaction Cost Integration

A critical operational friction in quantitative trading is the impact of transaction costs. Brokerage fees, bid-ask spreads, and execution slippage collectively erode between 10% and 60% of theoretical returns, particularly in high-frequency strategies where constant rebalancing is required [cite: 20, 24]. Traditional optimization models typically treat these costs as post-hoc constraints or oversimplified linear penalties, fundamentally misrepresenting how costs accumulate non-linearly during volatile periods when liquidity evaporates [cite: 20].

Recent GNN frameworks embed transaction cost optimization directly into their computational graphs. One prominent approach involves end-to-end learning systems equipped with differentiable quadratic cost layers. These layers mathematically backpropagate slippage gradients directly into the portfolio allocation weights. Consequently, the architecture organically learns to delay trades or freeze low-conviction positions during turbulent, illiquid market conditions. By trading only when expected returns exceed dynamically forecasted cost thresholds, these models reduce rebalancing frequency and overall portfolio turnover by 20% to 40% compared to mean-variance optimization, without compromising the target risk exposure [cite: 6, 20].

[image delta #2, 0 bytes]

 

In highly scaled portfolios, GNN models utilize custom reduce-weight mechanisms and L1 (LASSO) regularization to avoid generating thousands of tiny, fractional weights across assets. By penalizing small positions, the portfolio is concentrated into fewer, high-conviction trades, reducing the natural churn of the market [cite: 25]. Alternatively, hybrid methodologies utilize GNNs to forecast future transaction costs between pairs of financial assets based on historical price movements and liquidity metrics. These predicted costs are subsequently applied as edge weights in a financial asset graph. Classical pathfinding algorithms, such as Dijkstra's algorithm, are then deployed over this graph to identify the mathematically least costly path for reallocating capital across the entire portfolio network [cite: 24, 26].



## Systemic Risk and Market Contagion Forecasting

Financial stability is increasingly understood by regulators and quantitative analysts as a topological property of a network, rather than merely an assessment of the isolated solvency of individual institutions. Traditional risk assessment frameworks routinely fall short during "Black Swan" events—extremely rare, high-impact occurrences characterized by heavy-tailed statistical distributions and sudden volatility regime shifts [cite: 27]. 

During severe crises, standard Gaussian assumptions fail catastrophically. Liquidity evaporates, and previously uncorrelated assets experience severe contagion as market participants are forced into indiscriminate liquidation. While Extreme Value Theory (EVT) provides tools like the Generalized Pareto Distribution to model tail losses, it often struggles with non-stationarity during active crises [cite: 27]. GNNs offer a powerful alternative mechanism by focusing on topological anomaly detection. By continuously monitoring the financial network, GNNs can identify sudden changes in the connectivity of a market that lack an obvious fundamental economic cause. These topological anomalies often serve as the earliest warning signs of an impending systemic liquidity event or a hidden contagion pathway [cite: 28]. 

For macro-prudential surveillance, highly specific graph models have been developed. The Spatial-Temporal Graph Attention Network (ST-GAT) framework was created to detect early warning signs of bank distress within the United States interbank system. By modeling bilateral exposures reconstructed from publicly available FDIC Call Reports across 8,103 institutions between 2010 and 2024, the GNN predicts distress by learning the long-run structural vulnerabilities encoded directly in the network's topology [cite: 29]. Permutation importance analysis on this model revealed that Return on Assets (ROA) and Non-Performing Loan (NPL) Ratios were the dominant node features driving the graph's attention weights, accurately aligning with post-mortem analyses of recent regional banking crises [cite: 29]. 

Furthermore, architectures like the Trading Graph Neural Network (TGNN) structurally estimate the impact of specific asset features, dealer characteristics, and inter-dealer relationships on asset pricing dynamics [cite: 30, 31]. Integrating concepts from the Simulated Method of Moments (SMM) with graph learning, TGNNs iteratively apply contraction mapping to model the specific bargaining power and holding costs of nodes within a dealer network [cite: 30, 31]. This economic grounding allows the model to map exactly how distress in one critical node (e.g., a primary dealer) propagates through interbank liquidity channels, offering superior prediction accuracy over reduced-form methods that rely solely on simple network centrality measures [cite: 31].

## Asset Class Specific Microstructures

The application of GNNs in systematic trading cannot be uniform; models must actively account for the distinct microstructural characteristics of different asset classes, as the fundamental nature of inter-asset correlations varies dramatically across equities, commodities, and cryptocurrencies.

### Equities

Global equity markets, including the S&P 500, the STOXX Europe 600, and Japan's Nikkei 225, exhibit deep liquidity and are heavily driven by corporate fundamentals, macroeconomic indicators, and institutional order flow. The structural composition of these indices heavily impacts their correlation profiles. The STOXX Europe 600, for example, features a distinct defensive sector tilt—with a significantly higher weighting in healthcare (14.3%) and consumer staples compared to U.S. and Asian indices—which allows it to deliver steadier performance during periods of economic uncertainty [cite: 32]. In contrast, the Nikkei 225 is a price-weighted index comprising Japan's top 225 blue-chip companies, requiring models to adjust for stock splits and differing baseline valuations [cite: 33, 34, 35]. 

In these mature markets, GNNs heavily leverage well-documented structural links, including verified supply chains, overlapping corporate boards, and rigid sector classifications, to forecast price movements [cite: 10, 14, 32]. Comparative studies integrating LSTM with GNNs on major stock indices have consistently demonstrated substantial reductions in Mean Squared Error (MSE), confirming the necessity of merging temporal sequence analysis with structural relationship data for effective equity trading [cite: 8, 9, 36].

### Commodities

The commodities market possesses entirely unique microstructural traits driven by physical delivery constraints, localized supply shocks, and geopolitical disruptions [cite: 37]. Forecasting models in this space must synthesize broad macroeconomic data with the physical realities of the global supply chain. Advanced hybrid frameworks are currently applying agentic generative AI and deep learning to commodity forecasting. 

Recent studies utilizing World Bank data covering 1960 to 2023 have developed architectures that utilize dual-stream LSTMs with attention mechanisms to fuse structured time-series price data with semantically embedded, fact-checked news summaries [cite: 38, 39]. These models are highly effective at detecting early warning signs of commodity price shocks, achieving a mean Area Under the ROC Curve (AUC) of 0.94, substantially outperforming traditional baselines like logistic regression [cite: 38]. The explicit inclusion of news sentiment embeddings highlights the capability of modern GNN pipelines to successfully synthesize unstructured geopolitical context with strict numeric pricing sequences.

### Cryptocurrencies

The cryptocurrency market presents a novel modeling environment due to its decentralized nature, extreme volatility, and fragmented price discovery across varying, often unregulated exchange venues [cite: 37]. Historically, aggregate cryptocurrency returns exhibited exceptionally low correlations with traditional asset classes like equities and government bonds, operating as a largely independent ecosystem [cite: 40]. 

However, recent statistical data indicates a profound regime shift. Post-2020, the correlation between aggregate cryptocurrencies and U.S. equities has increased sharply, reaching approximately 37%, suggesting that crypto assets are increasingly trading akin to high-beta, large-cap growth stocks rather than uncorrelated hedges [cite: 40, 41]. Because cryptocurrencies are fundamentally "zero-carry" assets—lacking intrinsic cash flows, dividends, or interest yields to anchor their valuation—they trade heavily on technical indicators, momentum, and investor sentiment [cite: 41]. Consequently, GNNs applied to cryptocurrency markets frequently construct highly dynamic edges based on shifting social media sentiment indices and underlying network metrics (such as blockchain hash rates), allowing models to rapidly adapt to violent changes in cross-asset spillovers [cite: 42, 43].

## Operational Limitations and Regulatory Constraints

Despite their overwhelming theoretical and empirical advantages in modeling market dependencies, the deployment of Graph Neural Networks in live, systematic trading environments remains constrained by significant operational, computational, and regulatory hurdles.

### Computational Latency and Distributed Processing

Financial markets demand ultra-low latency execution. However, graph-based interpretability and deep message-passing algorithms exhibit extremely high computational complexity. Processing dense graphs—where the number of edges grows quadratically with the number of nodes—poses a severe operational bottleneck. Achieving real-time deployment that requires sub-50-millisecond execution remains an ongoing engineering challenge for large-scale GNN architectures [cite: 6, 20, 23].

To mitigate these latency issues, system-level optimizations are strictly required. Frameworks such as ROC (a distributed multi-GPU system) have been developed to optimize graph partitioning and memory management across hardware. This allows the acceleration of GNN training and inference on large-scale graphs by simultaneously utilizing the compute resources of multiple nodes, enabling exact GNN computation without resorting to accuracy-degrading sampling techniques [cite: 44]. Additionally, researchers propose utilizing lightweight graph architectures achieved via neural pruning to enhance execution speeds for high-frequency algorithmic trading [cite: 6, 23]. Beyond latency, the environmental sustainability of high-compute graph processing is an escalating concern; training complex, dynamic GNNs requires an order of magnitude more energy than traditional time-series models due to the massive volume of matrix multiplications involved in data aggregation steps [cite: 28].

### Explainability and the Regulatory Environment

The inherent "black box" nature of deep neural networks creates a profound transparency deficit, which is fundamentally incompatible with highly regulated global financial environments. Strict legal mandates, such as the European Union's AI Act and numerous consumer protection frameworks, increasingly codify the "right to explanation." These regulations require financial institutions to mathematically demonstrate that their predictive models are fair, interpretable, and devoid of discriminatory bias or unseen systemic risk [cite: 45, 46, 47]. A lack of explainability prevents human risk managers and external regulators from assessing compliance and anticipating model failure during crises [cite: 46].

Consequently, the integration of Explainable Artificial Intelligence (XAI) with GNNs is no longer an academic pursuit, but a strict structural necessity for deployment [cite: 45, 47]. Post-hoc explanatory methods, such as GNNExplainer and GraphLIME, attempt to provide local approximations of the model's decision-making by analyzing specific target node neighborhoods [cite: 45]. However, there is an industry-wide paradigm shift toward developing intrinsic, self-explaining architectures that generate predictions and their underlying explanations concurrently.

Mechanistic interpretability is also gaining significant traction in institutional finance. By revealing the internal circuits, attention weights, and specific neuron activations that drive predictions, mechanistic interpretability allows quantitative risk managers to trace a GNN's output back to specific topological features or raw data inputs, making the model fully auditable [cite: 48]. 

Furthermore, GNN-driven risk assessment introduces entirely novel systemic vulnerabilities, particularly concerning data poisoning. Because GNN predictions rely heavily on the integrity of relational topologies, malicious market actors can execute sophisticated adversarial attacks by fabricating artificial relationships. This could involve executing high-frequency wash-trades to create fake statistical correlations, or establishing networks of shell entities to manipulate the supply-chain graphs scraped by the models [cite: 2, 28]. Ensuring adversarial robustness through advanced techniques like Federated Learning and Differential Privacy is vital for maintaining the long-term integrity of graph-based financial trading systems [cite: 45].

## Conclusion

Graph Neural Networks represent a transformative advancement in the modeling of inter-asset correlations for systematic trading and risk management. By abandoning the restrictive assumptions of Euclidean, linear time-series analysis, GNNs natively process the complex, multi-layered topologies that define global financial markets. Through the application of advanced spatio-temporal architectures, dynamic causal attention mechanisms, and the integration of alternative data streams, GNNs achieve superior predictive accuracy and highly robust portfolio optimization. They successfully mitigate the detrimental impacts of non-stationarity, transaction costs, and structural regime shifts that cause traditional econometric models to fail.

However, realizing the full potential of GNNs in production environments requires overcoming substantial engineering and compliance challenges. The severe computational intensity of dynamic graph updating necessitates sophisticated distributed hardware frameworks and lightweight, pruned architectures to meet the strict latency requirements of systematic trading. More critically, the opacity of deep graph learning models mandates the continued acceleration of mechanistic interpretability and explainable AI to satisfy rigorous global financial regulations and guard against adversarial network manipulation. As these technical and regulatory frameworks mature, GNNs are positioned to supersede traditional dependency models, becoming the fundamental infrastructure for alpha generation, dynamic portfolio allocation, and systemic risk surveillance in quantitative finance.

## Sources

1. [www.preprints.org/manuscript/202507.1736](https://www.preprints.org/manuscript/202507.1736)
2. [www.mdpi.com/2073-8994/17/8/1344](https://www.mdpi.com/2073-8994/17/8/1344)
3. [proceedings.neurips.cc/paper_files/paper/2022/file/2857242c9e97de339ce642e75b15ff24-Paper-Conference.pdf](https://proceedings.neurips.cc/paper_files/paper/2022/file/2857242c9e97de339ce642e75b15ff24-Paper-Conference.pdf)
4. [ojs.aaai.org/index.php/AAAI/article/view/39585/43546](https://ojs.aaai.org/index.php/AAAI/article/view/39585/43546)
5. [arxiv.org/abs/2410.22377](https://arxiv.org/abs/2410.22377)
6. [www.qfintec.com/post/unlocking-market-interdependencies-with-graph-neural-networks-gnns](https://www.qfintec.com/post/unlocking-market-interdependencies-with-graph-neural-networks-gnns)
7. [www.bloomberg.com/professional/insights/data/supply-chain-momentum-strategies-with-graph-neural-networks/](https://www.bloomberg.com/professional/insights/data/supply-chain-momentum-strategies-with-graph-neural-networks/)
8. [arxiv.org/html/2306.03763](https://arxiv.org/html/2306.03763)
9. [scholar-press.com/uploads/papers/gwo9GeOSc2X5DJcbqL7ZvIExbm54DXEJM9kk6jIp.pdf](http://scholar-press.com/uploads/papers/gwo9GeOSc2X5DJcbqL7ZvIExbm54DXEJM9kk6jIp.pdf)
10. [www.mdpi.com/2504-2289/8/6/65](https://www.mdpi.com/2504-2289/8/6/65)
11. [www.google.com/search?q=time+in+Belo+Horizonte,+BR](https://www.google.com/search?q=time+in+Belo+Horizonte,+BR)
12. [isipress.org/index.php/IJAIR/article/view/91](https://isipress.org/index.php/IJAIR/article/view/91)
13. [medium.com/@huihuijiang0619/graph-neural-networks-for-stock-market-prediction-a-heterogeneous-graph-approach-with-multi-agent-18aedf30d06b](https://medium.com/@huihuijiang0619/graph-neural-networks-for-stock-market-prediction-a-heterogeneous-graph-approach-with-multi-agent-18aedf30d06b)
14. [arxiv.org/abs/2504.07923](https://arxiv.org/abs/2504.07923)
15. [papers.academic-conferences.org/index.php/icair/article/download/4294/3996/15803](https://papers.academic-conferences.org/index.php/icair/article/download/4294/3996/15803)
16. [www.ijcai.org/proceedings/2020/0626.pdf](https://www.ijcai.org/proceedings/2020/0626.pdf)
17. [journals.plos.org/plosone/article?id=10.1371/journal.pone.0346324](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0346324)
18. [arxiv.org/html/2504.07923v1](https://arxiv.org/html/2504.07923v1)
19. [arxiv.org/pdf/2505.02566](https://arxiv.org/pdf/2505.02566)
20. [par.nsf.gov/servlets/purl/10477853](https://par.nsf.gov/servlets/purl/10477853)
21. [openreview.net/pdf?id=NKvEGvaZLK](https://openreview.net/pdf?id=NKvEGvaZLK)
22. [openreview.net/forum?id=NKvEGvaZLK](https://openreview.net/forum?id=NKvEGvaZLK)
23. [arxiv.org/html/2605.19278v1](https://arxiv.org/html/2605.19278v1)
24. [theory.stanford.edu/~aiken/publications/papers/mlsys20.pdf](https://theory.stanford.edu/~aiken/publications/papers/mlsys20.pdf)
25. [www.mdpi.com/1911-8074/18/9/506](https://www.mdpi.com/1911-8074/18/9/506)
26. [f1000research.com/articles/15-274/pdf](https://f1000research.com/articles/15-274/pdf)
27. [pmc.ncbi.nlm.nih.gov/articles/PMC11098797/](https://pmc.ncbi.nlm.nih.gov/articles/PMC11098797/)
28. [arxiv.org/html/2510.14435v1](https://arxiv.org/html/2510.14435v1)
29. [advisor.morganstanley.com/seven-brz-group/documents/field/s/se/seven-brz-group/MSSBNA20251001576762_%281%29.pdf](https://advisor.morganstanley.com/seven-brz-group/documents/field/s/se/seven-brz-group/MSSBNA20251001576762_%281%29.pdf)
30. [www.scienpress.com/Upload/JAFB/Vol%2014_6_6.pdf](https://www.scienpress.com/Upload/JAFB/Vol%2014_6_6.pdf)
31. [journaljamcs.com/index.php/JAMCS/article/view/1952](https://journaljamcs.com/index.php/JAMCS/article/view/1952)
32. [arxiv.org/html/2502.15813v1](https://arxiv.org/html/2502.15813v1)
33. [drpress.org/ojs/index.php/HSET/article/view/6649/6444](https://drpress.org/ojs/index.php/HSET/article/view/6649/6444)
34. [personales.upv.es/thinkmind/dl/conferences/iciw/iciw_2024/iciw_2024_1_30_20016.pdf](https://personales.upv.es/thinkmind/dl/conferences/iciw/iciw_2024/iciw_2024_1_30_20016.pdf)
35. [www.atlantis-press.com/proceedings/icdeba-24/126008524](https://www.atlantis-press.com/proceedings/icdeba-24/126008524)
36. [cdn.arenafi.org/papers/arxiv/2410.01864v1.pdf](https://cdn.arenafi.org/papers/arxiv/2410.01864v1.pdf)
37. [openreview.net/forum?id=NKvEGvaZLK](https://openreview.net/forum?id=NKvEGvaZLK)
38. [openreview.net/pdf/1642ffb5db2aed311cb819e8593f578a35e45a15.pdf](https://openreview.net/pdf/1642ffb5db2aed311cb819e8593f578a35e45a15.pdf)
39. [doras.dcu.ie/30314/1/Thesis%20Ran%20Li%202024.pdf](https://doras.dcu.ie/30314/1/Thesis%20Ran%20Li%202024.pdf)
40. [arxiv.org/html/2407.15532v1](https://arxiv.org/html/2407.15532v1)
41. [www.google.com/search?q=time+in+China](https://www.google.com/search?q=time+in+China)
42. [isipress.org/index.php/IJAIR/article/view/91](https://isipress.org/index.php/IJAIR/article/view/91)
43. [www.researchgate.net/publication/393873789_Corporate_default_risk_prediction_model_based_on_graph_neural_networks](https://www.researchgate.net/publication/393873789_Corporate_default_risk_prediction_model_based_on_graph_neural_networks)
44. [arxiv.org/html/2605.21192v1](https://arxiv.org/html/2605.21192v1)
45. [ceur-ws.org/Vol-4035/Paper6.pdf](https://ceur-ws.org/Vol-4035/Paper6.pdf)
46. [www.mdpi.com/2078-2489/15/12/743](https://www.mdpi.com/2078-2489/15/12/743)
47. [arxiv.org/html/2508.06497](https://arxiv.org/html/2508.06497)
48. [pmc.ncbi.nlm.nih.gov/articles/PMC9857912/](https://pmc.ncbi.nlm.nih.gov/articles/PMC9857912/)
49. [www.aaup.edu/sites/default/files/2024-12/document_1.pdf](https://www.aaup.edu/sites/default/files/2024-12/document_1.pdf)
50. [www.mdpi.com/2227-9091/11/2/27](https://www.mdpi.com/2227-9091/11/2/27)
51. [arno.uvt.nl/show.cgi?fid=187649](http://arno.uvt.nl/show.cgi?fid=187649)
52. [openreview.net/forum?id=NKvEGvaZLK](https://openreview.net/forum?id=NKvEGvaZLK)
53. [arxiv.org/abs/2504.07923](https://arxiv.org/abs/2504.07923)
54. [www.mdpi.com/2073-8994/17/8/1344](https://www.mdpi.com/2073-8994/17/8/1344)
55. [arxiv.org/html/2605.19278v1](https://arxiv.org/html/2605.19278v1)
56. [www.ashpress.org/index.php/jcts/article/download/85/60](https://www.ashpress.org/index.php/jcts/article/download/85/60)
57. [www.isipress.org/index.php/IJAIR/article/download/84/111](https://www.isipress.org/index.php/IJAIR/article/download/84/111)
58. [www.mdpi.com/2227-7390/14/2/289](https://www.mdpi.com/2227-7390/14/2/289)
59. [www.researchgate.net/publication/401392167_Graph_Neural_Networks_for_Systemic_Financial_Risk_Forecasting_Modeling_Cross-Market_Contagion_Between_Banking_Systems_and_Cryptocurrency_Markets](https://www.researchgate.net/publication/401392167_Graph_Neural_Networks_for_Systemic_Financial_Risk_Forecasting_Modeling_Cross-Market_Contagion_Between_Banking_Systems_and_Cryptocurrency_Markets)
60. [fessorpro.com/blog/black_swan](https://fessorpro.com/blog/black_swan)
61. [www.rcresearcharchive.com/index.php/Journal/article/download/755/740](https://www.rcresearcharchive.com/index.php/Journal/article/download/755/740)
62. [www.bis.org/fsi/fsipapers24.pdf](https://www.bis.org/fsi/fsipapers24.pdf)
63. [aijourn.com/the-end-of-the-black-box-why-explainable-ai-is-winning-in-financial-services/](https://aijourn.com/the-end-of-the-black-box-why-explainable-ai-is-winning-in-financial-services/)
64. [arxiv.org/abs/2604.14232](https://arxiv.org/abs/2604.14232)
65. [www.nvidia.com/en-us/on-demand/session/gtc26-s82175/](https://www.nvidia.com/en-us/on-demand/session/gtc26-s82175/)
66. [arxiv.org/html/2410.16858v1](https://arxiv.org/html/2410.16858v1)
67. [pmc.ncbi.nlm.nih.gov/articles/PMC10614642/](https://pmc.ncbi.nlm.nih.gov/articles/PMC10614642/)
68. [ideas.repec.org/p/arx/papers/2410.16858.html](https://ideas.repec.org/p/arx/papers/2410.16858.html)
69. [www.researchgate.net/publication/395378283_Using_Dynamic_Graph_Neural_Networks_for_Temporal_Analysis_in_Financial_Markets](https://www.researchgate.net/publication/395378283_Using_Dynamic_Graph_Neural_Networks_for_Temporal_Analysis_in_Financial_Markets)
70. [openreview.net/forum?id=KsWRLyIAKP](https://openreview.net/forum?id=KsWRLyIAKP)
71. [www.researchgate.net/publication/379693335_A_COMPARATIVE_STUDY_OF_COPULA_MODELS_FOR_ANALYZING_TAIL_DEPENDENCE_IN_QUANTITATIVE_RISK_MANAGEMENT](https://www.researchgate.net/publication/379693335_A_COMPARATIVE_STUDY_OF_COPULA_MODELS_FOR_ANALYZING_TAIL_DEPENDENCE_IN_QUANTITATIVE_RISK_MANAGEMENT)
72. [www.diva-portal.org/smash/get/diva2:1966077/FULLTEXT01.pdf](https://www.diva-portal.org/smash/get/diva2:1966077/FULLTEXT01.pdf)
73. [eureka.patsnap.com/report-using-graph-neural-networks-for-predictive-financial-analytics](https://eureka.patsnap.com/report-using-graph-neural-networks-for-predictive-financial-analytics)
74. [arxiv.org/html/2407.15532v1](https://arxiv.org/html/2407.15532v1)
75. [arxiv.org/html/2605.21192v1](https://arxiv.org/html/2605.21192v1)
76. [arxiv.org/abs/2010.02089](https://arxiv.org/abs/2010.02089)
77. [openreview.net/pdf?id=XI-OJ5yyse](https://openreview.net/pdf?id=XI-OJ5yyse)
78. [medium.com/@ds-faxi-yuan/graph-neural-networks-gnn-comparison-between-cnn-and-gnn-5c97fdfb3e31](https://medium.com/@ds-faxi-yuan/graph-neural-networks-gnn-comparison-between-cnn-and-gnn-5c97fdfb3e31)
79. [ijchml.com/index.php/ijchml/article/view/91](https://ijchml.com/index.php/ijchml/article/view/91)
80. [www.mdpi.com/2227-7390/11/11/2515](https://www.mdpi.com/2227-7390/11/11/2515)
81. [tradingeconomics.com/japan/stock-market](https://tradingeconomics.com/japan/stock-market)
82. [www.tradingview.com/symbols/TVC-NI225/](https://www.tradingview.com/symbols/TVC-NI225/)
83. [indexes.nikkei.co.jp/en/nkave](https://indexes.nikkei.co.jp/en/nkave)
84. [simplywall.st/markets/jp](https://simplywall.st/markets/jp)
85. [groww.in/indices/global-indices/nikkei](https://groww.in/indices/global-indices/nikkei)
86. [arxiv.org/html/2502.15813v1](https://arxiv.org/html/2502.15813v1)
87. [arxiv.org/pdf/2502.15813](https://arxiv.org/pdf/2502.15813)
88. [medium.com/stanford-cs224w/temporal-graph-learning-for-stock-prediction-58429696f482](https://medium.com/stanford-cs224w/temporal-graph-learning-for-stock-prediction-58429696f482)
89. [www.goldmansachs.com/insights/articles/economic-growth-is-forecast-to-drive-a-modest-rise-in-europe](https://www.goldmansachs.com/insights/articles/economic-growth-is-forecast-to-drive-a-modest-rise-in-europe)
90. [www.visualcapitalist.com/sp/charted-getting-to-know-the-stoxx-europe-600-stx03/](https://www.visualcapitalist.com/sp/charted-getting-to-know-the-stoxx-europe-600-stx03/)
91. [openreview.net/pdf/1642ffb5db2aed311cb819e8593f578a35e45a15.pdf](https://openreview.net/pdf/1642ffb5db2aed311cb819e8593f578a35e45a15.pdf)
92. [openreview.net/forum?id=NKvEGvaZLK](https://openreview.net/forum?id=NKvEGvaZLK)
93. [ojs.aaai.org/index.php/AAAI/article/view/39734/43695](https://ojs.aaai.org/index.php/AAAI/article/view/39734/43695)
94. [epub.jku.at/download/pdf/11475795.pdf](https://epub.jku.at/download/pdf/11475795.pdf)
95. [www.researchgate.net/publication/388667021_Cost-Efficient_Asset_Allocation_Graph-Based_Machine_Learning_for_Dynamic_Portfolio_Rebalancing](https://www.researchgate.net/publication/388667021_Cost-Efficient_Asset_Allocation_Graph-Based_Machine_Learning_for_Dynamic_Portfolio_Rebalancing)

**Sources:**
1. [preprints.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE1OFqsKTFPJO_ECoxGgOh8v_CarVOTElKLAWt0OzOFDCodHwQYbK13ADe2XkU7u40JbgW5Y9uaGneVrfHUTCbxS_ZO5yc8Bb26473fcaVwjL1gskyNS7qOufTGQUQAMbtgJMgGRQ==)
2. [isipress.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHvnz5d3PjZ9kPGsSSJW23a3zx67IKQ_eJgGAgcMHf-mwO5X6y_E1BrYO-blgUSu4upe7M0yQrgVe7obJpUGgIgCP87FIfXTf0zVSOJ4dqLXLN1XkIoGPdtVnL0s4AO0Zg4AMQGRXZ1Rlw=)
3. [mdpi.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFrFBWs9fScXQBEvqCrXiM7HRjYXhYHGCnc7XfoxrskEa4PAdT1QTivXptQWly_opYbFpB8clIh-bwY6JE5LB0JiM1H8U9SWbqHScwkPDbArW3E9gGF4KqrhAyVRX8=)
4. [qfintec.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFabUpdjLbzv58dZ_nmbVWh77r8ueGleKA49sxtUOdiLMDW54kXMRq3gpfy15RYR0oXwZEONTFW7yQOSPRp4UKlGtmsoRGcJTftUUbFLq_4pm6upw4k24xQuvbHSq0yRa80TpfgTMaEtZfnTQV6alNAQ6yXjdLL9lTjlZEKdtYJVtiDLfBLeI9iOB0--R3zcqhqp1ky)
5. [bloomberg.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFe95IdNM6YbgZYxtrAYK2_cWFaYJxHXNzUhhgK-8Z4naMD81t63qzMg55ebjHGshJb65pnH9gPapjQhs0dN492u4GAX7kPyMDiDa-ru-7O_tje9MdO02HJ0CU8EKfwnlLjYzRqpuIc4FIqjc4qyNZXXkcLcR8Kdmn7w0MK1ju4wv0W7ioAsSW851h9hMUdoS04jfJDh_UhZyb27pW9An2xD3nAUVhB)
6. [openreview.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQH8uLum6bupq1Ge_pP-isoHh9XghRukjKg_fUI9xNP5JHRLX3G_IinbEQ7aSOfT5N9-OtYMqvr1VELeSN0KUP05sE6yudF5mFHMIjct96-QbmN2oKLRstD4IcC3OgQ=)
7. [repec.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGU39V2kjs23sxAZsg-8uz94zEWOm0p1iCVrvXBF8muzURAKPrQYplUmp7Al3rYKWRwPz-h51REqmu9Tvy6FAezQUcld-FhDrbD3w4BsWVootZncji3rFoV5-wKSI34SQ5e1tQTPWwQyL8=)
8. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGiuMNh7eZcNd4sgY1mLJnwmqeVmBMhof0ddLggZDPNfKYNORUJbIFrl5OrYYAZXhNK0HSxmmiYVzCamkDCycKX15SCdI7uNm4Pg8Kg90_FAv9QG2JpT6nG)
9. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHf8APSJJxZOQ4pHXNq6kc_argFaK1SU6H5waE7Q9zehx1aoKvV5iyEPig9WUyGwAyh-bbGQmeXIO0Cvn_7n8ZkoYipqROECaf_0nxWdY9OgneAvNTZ)
10. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGmY1iOq0_PGIxCub-yBY-vRapteyAbm6R8PVVdUo8jfAg9AJmSiCNo2XUnjy49ZrFSRhVmyYn6Px8sqWenl82TqEiZW7m7u5RCLCPrOQl56eNOVVXem6yp)
11. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFOcdkXv-CDjuNC4OPNcMYb7ZYOWMZVowUoqEGlMIEot5l5yOX39GA5wFK7ev83JjnU7B244onfgQn6OWtyWpzac2R4gDAsufYQGpjEsOmPWINF3p0fnw==)
12. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGeNhl1VHb3NUoTKSclNzP15QENULOGuIWqdhyiJRPF3FODIPr2J1_wgyMeAGIiuBoCE2FUfBs3ruwfJg-OlHd7OEr6O6NbmygrzDWkXUhf1NwDUU1sBbftojvLQFUTslolbF3FGPBx9Y4D9ETiOXOzNyujO7zTuw7Icj7W6wzG3iKEdcHbTcyYAQznERy-1YPS0hawcl1_GDLatXEmDEKigblKxl85wyxCeBat9dnt0qIywVVh89ZUViZ1b4y2-9d2OomU0UEeFw==)
13. [scholar-press.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGUljxD0WtIQHm0cyF-C1Hfe4gUnCQ-EyczQdoCrn2pUYWohju5qm83TXy7uiBXCjB-Cvix1uatJKzwggRJwMEkVx6i2TQhFQlUrd5wSFNodTblizPLrQWY9YXtux_1wM17xxxHG6vMnj7cq0hU7kH9LV6H4JBGe5YVbaop0ohm5wW9TUHWpJlmrg==)
14. [researchgate.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQH3TxmYXOducA2WEc0jNxnGnP4LQerco92Vwo3Go3t8qDuvOPbOjYGZZPUhbKo25e1CQJikY3D3Hs-M67k25k4G2db8dBcylpPbFNkSOQ4yT1h0TQdPELprCqMQejqyV6Hml5O6h8dj-y0uA_b0F6Sbj6MPpn1gycxIirlvCwBlw55FG45E3Oo1WIUz1rQeWS13OG8oMKD8HN0QQVHUYnzw90tu3PiLXALNZ55neQg=)
15. [ijcai.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFgpRp7EyqieUsIBZOJlYpS6CxEFaPMP8IiExIZXrB2ATG5UEUZ2aLOI9W7gHpZFVbhTeI3VOKBqcuRQD-3Tzy1VTrf1-gCm9brEQWVFLIMCbmrVCncX_U1HTlv7i6ns77Y4nIj)
16. [neurips.cc](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFoq1Kc59sTY-6VRz_feTvhd7KgQjzLeG_UYbHrFQ00WvtalsRa6QVQ-6JRUxJNUNJ79icf-ZR_6_THV9qbOm53HKk2vRUW8w0idcnxrRer5ahztTVLMBoY9s58yd3GdYYJLo3im6LTm58X4q6_fGft3IoK46M_ZXsNQRK_6JRGySRjRxXtVtc_84U96TfPb5CzbL-mgROOG7nd0uoDIUEDQeYS8MY=)
17. [aaai.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEUzweyDSsACIkBAxcAkoau54ZUvO7rLmDrIaZO3pZjO4cg6rZUm78lpp78CVWPFqqelfBQm8Lk3Gw0xPyTEizeujM5rQXHufSLx6XqVdrJnvqt84R9Sc1Wa066ajkHgQBHE0gRs59MhzqtKipVx4EydQ==)
18. [mdpi.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGaHA1F4E4CMbT0QZHEO9SrU09Npf6hwx3EqUd93eYVAfy1D5zQyAhua_bDJSA5pYyfabNmxLJGLM5xUUa5CUQG7Q3afM4lapbL6DT_WLMQXcisL6G6lsKhHuynTg==)
19. [researchgate.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFC7iULgHSq26_08ZYSpnvLUlVHy-f_NzvY11tRVreCxwtn3Hw2iIe0wuf6Duh_ceSWa4XPXQSMJFioiYOcnpukkgFsNEMyhG9tDgUl9CEkDJrO9CfiJj8pVDf1upXzq840LPUGsH92nQ7ckoHqXPngWmUCWuclGPNBAcyTz3eA6mrONeTCvqqDZfjA22H6NfLUiCoJAHioKKaIdkvCuA7NSCS6AanH7k4aJWCPMQtzEG-twBJM-uEsy1drRNI2tj5UN5MDlEPvWyK5Lg==)
20. [openreview.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE44xOWwHHTRClN63DKzmkd_PuegbxYpjYqZ-PB7L7Wt63_9ai1nHU72kDvyV6mBPLhW-ey-zCTmaeg4OCKWjBd42NO1YMu-BWPO71w1FJAc8mXN0QwxirRlO8uvLpTkhqY3v0M2eRdFAFIxHTLSLoIlfLDQlyCS8AMCw1e)
21. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEgnvMo4ShVbpHMsQ6jywEYurgNvCCBYMSFl1A0T-AO2rSLqSesR1S1gpdj8NUKeETuKEMNwYhNQ5YazM-4SA-Y_9QwYB7SmK9bs6hPbkPbBF2ib9jW)
22. [openreview.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEixjdu-pfV-Ambt6nBM8yaWmeBDrTGVq2HwWxUIcxPjh23PK65DxLViQbeyR3p-ash0y6wJjIaDbOLcWnxkTsU_wEQhuHz-jQywJenHX7YlU2ro7ChwYUOJMktj-E=)
23. [openreview.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF-dT0aQrgJghfcUetcdd4dpFXUFZVEuS_K27rKuT6ygyl2rCDIAOG2BhOtktTD3KogUH4jvAfiHT9Qsw9j9pCqMdnytyQTPvRBriYefnqe9x6-l46zR4VsotpZ0F16jQ==)
24. [arenafi.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEFSrOJ4OXPnMPIO_eu05aBxksSPiARzCaHpEqtrtiByYO_0hwrcs4PkA2g-_In7_58zQ6TKXJZSJHeUq7xpiRoXSSI5za54kIBMH5EYgTO9MQywGqFUM46sOLt3T9-LNcWNP5Erac2mqNI)
25. [Link](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQExVZmJ_mP5H6m8lmA7X-e5vX4RZUnEWG4FfLn8-EtyMgz3IM4TPnWXABbhr2msElQ1XZW1Z4VNooFihCUno9cQVRH33DRxCxQNTdDAP1AEDGBDPe-Cth-x)
26. [researchgate.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHluohvd57CMvumJgQzAF0MGBMt5BCx7USnGN6W4WiGOuVv2VAWGCHBisQB9wSWXHGn2LLlVVA8q6-2jcQAEb1JZ1JsuWbsU71lOq5CUf1zLafTGWZhS6Dg70JPLX2Kxt2LSYF9Hr27hpQiygahxuIAVXrbS-pdyb8xyF_ubDLnHS2Ty-5BYxVSCUemDzR19Ip327Ftali8muhjJFaZMAdwcv5gmb-IDc13SLMTCIRfYrgoeeMCqNvGADbWt6jielG3LAcQn0B2)
27. [fessorpro.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE6pyc76AXyysdWHkR-gRHTFKNNfM4TRMh-1J4HW8tVVZkuj9Bwl7FaiI3TIoE1Z8TS4NigDOxUUL999b8qJwY3VQZfRTnJXEcjB69XXKPrYF94WOTx5GddK8s=)
28. [isipress.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGOdJOMQQ_xT7ix9HHnFyI2KQyidvGjzicbEfB-NIFdaAhi0Fa3ZryT2i3az-Ardes7GAK73uW36-pMq4Pbwyf4MQXHGxGCkT9aQbz2kEwgbfTExV5kzDb3qC_gahVli_ePUNZciuOdvHtEojlaDeLkPPIiNrA=)
29. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGugLcjMdyksig6176zfDAGTJ_qU2t7gszRs7v-PkluyGIJTyrXANGfDrLntoVy4axVsj_Fhr_LTbmZLtpdaVgVI-XRduK3txRlICPERZ88MMrzgTsM)
30. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHQRcjrwKGgyvy34wt06GRC3JU9gwemT3eMJsssbRWYo10URbMeKf--xvAviYiSddEYu66DKs4lBrn0Qdb7dgKBVJL9_mR1MyOEtTSe9rbvBZiSOHOz)
31. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGADw9ovjgWxies5R3hjYswTpLuW9ae39CKpRlJTboI4p1r5wUvEBMYf31EBKyZ1fc8n1YKcq5tzAUxsMsy9d2sfxO2zA_9rvEjN8wnacquRtpnGJNETlvS)
32. [visualcapitalist.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE8sw9PQl01d8D0M575WZHmFAInXZwQO9fLy_eArSophXHnOYCNJk4pVs3QJ1lZofcGaE-0ZIYfr04WZtQlEdC-DMljQM-RIq8YdP1SMhi7HXIUghkw0yxx2Iy5s1VxAqG2ruGrY4t62WBhcjXnPQembW_Rt1bicdvtWhyikX4d1aXI5j2R-DfaQLf2rg==)
33. [tradingeconomics.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHIEMjKWB_ocB6YIkPkEpsr2T9jtHPUb3yq22u-tUp--oJVni1ECXbcA0DUo-WkdNsSTvNq4UMdoIws9MJikTNvdORfV_Bq2-GPCdCdXjg1wU-ZAZTovfUnPtr2mvPz7Uly2bsq)
34. [tradingview.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEM4aDHtbh9fO1WIYO3QIk5FkeLDok9i50Hvk5QIwuNpu-TK267givKsTmEPmYiYJH43CcvYW7ulUt2BabbzMcQSjIxrJBVQ1o6VEtkZYBJ_DGuuDpgC7ktRmvResocNnxsaRk=)
35. [groww.in](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFie5OmYNwBDt6IWFVY5WsBjZ-TQjwlbMTJBOo-Wn7NTAJzHsbo4M_FVCvVT_da9VXAm25wfoC8xvGV7QJxd4BI0UakrUE3DxeVv-YAk0-pSQTCnAiXBQK692VuZslLzavzHBg=)
36. [journaljamcs.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEXzkstJAcFi30B24TBf20bEAq_wRHx0fj4xd-NeRJ8_NBYIwn2jpvBNaXBQy-3F_2L9ck7pEJ_4zML13v8bliaukH5cqopwQx3KFgD1T9tnBr8P-f26t-IHDjvnDSJTY1frFDKmQ5y4G3eCj_qHFI=)
37. [f1000research.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGhFm2rRXCCwyKDi2a_AXKWVkhCa3yr9t7mj26Y9q3Y4o3uREjO9-JvoY7n-yUzKCvM7ydEP5gpQ2UpxFABhYmWV8Mn7L5D0pZ0uXaiD1EdxH6rwCDBDUEh3CvJZMZJIABCIg==)
38. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQG2D4h1Sa3lMXA6zhSJbdp046gLscEN5dLq3l-6otg76wBrUTtxUG9v4Y9v9qsCjMTPvp9O1l-hsu7gPvmP_9jETe6eURX_zI0wyiyOff6ddsmpo5RiYA==)
39. [nih.gov](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEXlw7E-tHevCNEOJPJyC142RndEjk3v12heaJCgEj1-JCmvcItFUDTPfvIvQylWxnIKHaNu6VgVoh0Xiw0NNsT4HeQD7f9awUtcLUPM7Lu5hVhpD0222zlgNvPmWLMXJg8f6UVfrU=)
40. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFwLgaI7MTfBFZdkEPej8YVuRHYYCYO1z41zGkHWlHSUv-eith9NrRD69CFTBOrGc6fIAkdOh_CafKlROjnOFfjTxEC8T1gRkJ-4BlclXO1poQ-acIUYnwG)
41. [morganstanley.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHTUs8sCsSQT7QOJ7-lktU3-5hKq20hu_vVRqgLIJSYXUKhZiB8aOU7u0DjoqYD4xqjUz0Kpr3Wca5Sd_iswFGARLPSftzIa9_aLe-3kGxxxYfaxlI4hZm3EYSRPlQuTid1AbfZju5P6IgnWp_XwWOT8TsdqFYWm-yqbVE6qw-JP4qy7hTMZ-mAe46MC_12VasfS7edafWwPt-DlO2rISq_L4QP2Kivc9lOtHVD)
42. [nih.gov](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF-Ek8SxTe5yOUAhUrDSQ-NbKxi25AcwoRxGIJEDhiMOCVxTeROSZxGLjWIc8VzzCohsPN6pD7MAmRai3zI0V9l5733VLKCqFpwvjNsSzn2peW_c58TNVFcOwXO2mxEKnmpjUGH0Dp3)
43. [scienpress.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQELY9Pm9W81RqgblJLYbRk6uqIz0XAqV6nUO58PXWcmfMIaOxQrawd8ePtoSUih1Unfo4Vffnu-BaPJQkRYbvpQNLzXz4mxe_AjobrpD2nkVfLMWhpQ9LExp3owm4hgEkbxzEBuIyV6OjTCPEQ=)
44. [stanford.edu](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGN_Be6zy3AwIuJ08G0IpoEHELeHSDpIl7pGfRhzwG_WzJf8tuPdKqXrPnwKIEaAAJLYHo6aT3QQN88D7G7TArVr4qmDxR6un7xsmKaaPHWa3OaHIP4UyerbzXBmCygXrPzB-EK3DpMqFavUYjCjRW0f2NcxD1Skw==)
45. [rcresearcharchive.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFrD2y0sa5SkEdjnqiYhoK2NlE95D61bVjjSbgn2H_oDIrfkRnCSKVOeeIfN9ORaidXDdIqFAjxGgGBc-E6KQ0zZoY9EigkHRDdK3CklZzPDMbzRYBiMHQ9S1gvVq_znRmoM0UprOW71GjqBmJrpessz0V9tSZRr1j0l16cQY84Xss=)
46. [bis.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGo6GkWwGzqioDRZHL2NKtcFfzNm36dIWyWJgSjrznB5NHYo3otuwBisdAraDUosKZSW2TL5kcxHokt2_3_G0XJBwO_tkfgYcxVYvOmvDut1wAzHmYUpi5GLgQ9qg==)
47. [aijourn.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE53FXU0GF_b_x8KngYccusl-BfKMXWkogYqF2u9SrKn_N5hrYA7v3WGClGdB5wbvFH2B4DNRo5sjvGcfo2QEbDqetpDAN7-dsUfAw5ssttsromJSy_PdX8qwnJkMpkhoagm3jjrdn2cr7Pv3DyMWY_1GKGFyd01-q6raKn-QVgU_HTL78Dra3u8wBzY_PKfGMToFABifU=)
48. [nvidia.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQH6Yv_L0qF5kuBI8MARl-La_VXme5s75BBaTjmRxNIqK4kFiIThUf984bMcfwVSECrBWYQAHkZWT8yKDywBVXW8SyDS4SP7F2NQyz3e3nhpdvmXIucEnzgG1rPOE4khlBErYM4uUEpqjndhiwfgwQFnaA==)