Beyond Price Correlations: AI Maps Cross-Border Market Relationships

New benchmark seeks to advance cross-market return prediction through standardized corporate disclosures

Researchers at National University of Singapore (NUS) have introduced CrossAlpha, a public benchmark designed to facilitate the study of how textual information from annual reports in one equity market can help predict stock returns in another, addressing longstanding challenges in cross-border financial analysis. The effort focuses on five interconnected markets—the United States, Japan, Taiwan, South Korea, and Hong Kong—chosen for their significant listed-firm universes and deep supply-chain linkages in technology and manufacturing. By standardizing heterogeneous corporate filings, constructing dense semantic firm graphs, and aligning them with daily price data spanning January 2015 to May 2026, the benchmark aims to provide a reusable foundation for evaluating cross-market factors grounded in economic relationships disclosed in public reports.
The motivation stems from the recognition that annual reports provide a textual channel to capture economic linkages before they fully manifest in stock price co-movements. A shock to a U.S. firm like NVIDIA, for instance, can ripple through to Asian semiconductor companies via relationships described in their filings regarding customers, suppliers, technologies, and end markets. Existing resources, however, fall short of supporting comprehensive cross-market evaluation. Single-market text-based peer methods overlook global supply chains. At the same time, cross-market asset-pricing studies often rely on backward-looking price correlations or statistical links rather than forward-looking disclosure-derived signals. Financial natural language processing benchmarks typically emphasize document-level understanding without testing whether extracted links generate economically meaningful return predictions.
CrossAlpha tackles three core difficulties in building such a resource. First, filings vary widely across languages and regulatory regimes, from SEC EDGAR in the U.S. to EDINET in Japan, MOPS in Taiwan, DART in South Korea, and HKEX in Hong Kong. Second, raw textual similarity can be distorted by common reporting styles and industry jargon that do not reflect genuine firm-specific exposures. Third, evaluations must respect trading-time constraints to avoid look-ahead bias arising from time-zone differences and mismatched calendars.
To overcome these, the benchmark employs a three-stage pipeline. Disclosure Distillation uses a large language model, specifically GPT-4.1, to convert raw annual reports into a standardized ten-category English business schema. These categories—covering products and technologies, customers and supply chains, geographic and financial exposures, and business model, strategy, and competition—were selected because each aligns with mandatory disclosure requirements across all five jurisdictions and draws on prior evidence of return predictability. The process produces comparable representations focused on substantive business content rather than format-specific elements.
Next, Residual Schema Graph Construction embeds each schema category using OpenAI’s text-embedding-3-large model. To mitigate bias from shared reporting conventions, principal component analysis whitening is applied to suppress dominant common components, yielding residual encodings that emphasize firm-specific differences. These are aggregated into a dense directed graph of approximately 19 million firm-pair similarity scores across 3,587 firms and more than 10,700 firm-year reports. The resulting structure connects source and target markets in a way that supports both static peer analysis and dynamic applications.
Finally, Timing-Aligned Evaluation pairs the graph with daily open, high, low, close, and volume data, as well as historical shares outstanding, for market-capitalization adjustments. The protocol enforces strict lags, such as using peer returns through month t-1 to rank target stocks for month t portfolios, ensuring feasibility in real-world cross-market execution. This setup enables testing of monthly peer-momentum factors and higher-frequency event-driven spillovers.
Experiments conducted on the benchmark demonstrate notable predictive power for cross-market text-derived links. In a U.S.-to-Japan setting, for example, disclosure-based peers achieved an information coefficient information ratio (ICIR) of 0.39, compared to 0.07 to 0.18 for domestic text, GICS sector, and return-correlation baselines. The long-short portfolio, formed by ranking target stocks into quintiles and going long the top and short the bottom, delivered positive risk-adjusted returns after transaction costs, with cumulative performance outperforming alternatives over the sample period.
Broader tests across targets reveal that cross-market signals frequently surpass domestic text baselines. For Taiwan as a target, domestic peers performed strongly, consistent with its concentrated semiconductor cluster, yet cross-market sources added value in most other cases. A source-selection analysis produced an “information geography” matrix showing directional and target-specific flows: the optimal non-domestic source varied—Korea for the U.S., the U.S. for Japan and Hong Kong, Hong Kong for Taiwan, and Taiwan for Korea—rather than defaulting uniformly to the U.S. market. Reversing source and target pairs often altered both the magnitude and sign of predictability, underscoring asymmetric information transmission.
The benchmark also supports event-conditioned daily trading evaluations. Following large positive idiosyncratic moves in source stocks (market-relative returns exceeding 3 percent), graph-retrieved neighbors generated abnormal subsequent one-day returns under a t+2 open-to-open execution protocol that accounts for trading lags. Incorporating a GPT-5 agent as a filter—labeling top graph neighbors as co-movers, inverse, or unrelated, and retaining only confident co-movers—further improved outcomes. At a basket size of 20, the agent-enhanced approach increased the annualized Sharpe ratio from 0.81 to 1.50 and substantially boosted cumulative returns. Case examples, such as an AI demand shock at NVIDIA, illustrate how the agent can exclude semantically similar but economically mismatched firms like Nintendo, whose console-focused exposures did not align with data-center dynamics.
These results suggest that annual-report-derived links capture nuanced, firm-level relationships missed by coarser industry classifications or price-based measures. The graph’s ability to surface cross-sector connections, such as automotive firms linked to technology and mobility platforms, highlights its potential for revealing supply-chain and competitive dynamics embedded in disclosures. Human evaluations of the distillation process indicated high accuracy overall (89 percent fully accurate across sampled filings). However, performance was somewhat lower for Korean-language documents, reflecting known challenges in multilingual financial text processing.
Nevertheless, the benchmark comes with important limitations that temper its immediate applicability for live trading. It is explicitly positioned as a research tool for studying disclosure-grounded return prediction in liquid-listed universes, not as a complete investment system. The annual-report graph is characterized as a medium-horizon signal, and the daily event tests use a fixed execution lag rather than exploring intraday possibilities. Reliance on LLMs for standardization and filtering introduces dependencies on model capabilities and knowledge cutoffs. However, the release includes standardized texts, embeddings, graphs, prompts, and outputs to support auditing and replication.
Construction and refresh costs are documented to encourage ongoing maintenance, and users are reminded that historical performance metrics do not constitute investment advice or guarantees of future results. All data are derived from public filings and market sources, with no incorporation of non-public information, and redistribution policies for the underlying portals are referenced to ensure compliance.
The release of CrossAlpha, including schema records, residual encodings, the full directed graph, aligned price panels, and evaluation code, fills a gap at the intersection of financial natural language processing and quantitative asset pricing. Prior single-market efforts, such as those deriving peer networks from 10-K filings or topic models within one jurisdiction, provided valuable foundations but lacked the multi-market scope and return-grounded protocols needed to assess global spillovers systematically. Cross-market studies using supply chains or graph neural networks have advanced return modeling but often stopped short of offering public, reusable disclosure text and aligned benchmarks.
By making these materials openly available, the creators enable further work on topics ranging from supply-chain shock propagation to the role of textual disclosures in international return comovement. The information geography findings, in particular, invite deeper investigation into why certain markets serve as leading sources for specific targets and how these patterns evolve with geopolitical or technological shifts. The agent-filtering results point toward hybrid human-AI workflows that could refine graph-based signals in practical settings, while also underscoring the need for careful validation to avoid overreliance on model reasoning.
Critically, while the benchmark demonstrates clear improvements over baselines in controlled backtests, questions remain about out-of-sample robustness in live markets where relationships may change and transaction costs, liquidity, and regulatory frictions could erode gains. The fixed FY2024 graph used for historical evaluation tests structural relevance but does not fully simulate a dynamic, point-in-time strategy that would require periodic graph updates. Sensitivity analyses in the accompanying materials address some hyperparameters and embedding providers, yet broader tests across different market regimes or alternative LLM architectures would strengthen confidence.
The directional nature of predictability highlighted in the information geography matrix also raises implications for market efficiency and information diffusion. If disclosures in one jurisdiction systematically lead returns elsewhere, this could reflect differences in reporting timeliness, analyst coverage, or investor attention across borders. Such asymmetries merit continued scrutiny, particularly as global supply chains face increasing fragmentation and policy-driven realignments.
Overall, CrossAlpha represents a methodical step toward bridging textual analysis with cross-border asset pricing. Its emphasis on standardization, bias mitigation through residual methods, and rigorous timing alignment provides a template that future benchmarks might extend to additional markets, disclosure types, or modalities such as earnings calls and news. As financial markets grow more interconnected, tools that systematically extract and evaluate economic linkages from public reports could enhance understanding of risk transmission and improve the informational efficiency of global capital allocation. Whether the benchmark spurs widespread adoption in research and practice will depend on its usability, the reproducibility of its findings, and ongoing efforts to address its acknowledged constraints.

IndraStra Global is now available on
Apple News, Google News, Feedly, Flipboard, and  WhatsApp Channel

DISCLAIMER: This content is for informational purposes only and should not be considered investment advice. Readers should conduct their own research or consult a qualified financial advisor before making any investment decisions.

COPYRIGHT: This article is published under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License. https://creativecommons.org/licenses/by-nc-nd/4.0/

REPUBLISH: Republish our articles online or in print for free if you follow these guidelines. https://www.indrastra.com/p/republish-us.html