Public Signals of Python‑Enabled AI in Finance: Disclosure Patterns and Outcome Claims in NYSE Institutions

Public Signals of Python‑Enabled AI in Finance: Disclosure Patterns and Outcome Claims in NYSE Institutions | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Public Signals of Python‑Enabled AI in Finance: Disclosure Patterns and Outcome Claims in NYSE Institutions Veliota Drakopoulou This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8884680/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Artificial intelligence (AI) is diffusing rapidly across financial services, but public disclosure of AI and software capability remains sparse and heterogeneous, even as Python has become the dominant language for analytics and model deployment. This research constructs a systematic map of Python-framed AI disclosure for a strict frame of 180 New York Stock Exchange (NYSE) financial institutions. A triangulated corpus of regulatory filings, employer portals, corporate communications, and sectoral press is assembled through January 2026, and each firm is coded into one of three disclosure states: explicit Python, indirect AI, or none. Named Python libraries are detected via sentence-level dictionary matching and contextual filters, then mapped to seven analytical dimensions (natural language processing, machine learning, deep learning, reinforcement learning, probabilistic modeling, optimization, and visualization). An Outcome Claims Index (OCI) flags quantified performance assertions (e.g., risk reduction, accuracy gains) and supports both frequentist and Bayesian inference when such claims are rarely observed. The results show that 76.7% of firms disclose explicit Python usage, 2.2% disclose only indirect AI references, and 21.1% disclose neither, with a visibility-weighted explicit share of 0.629. Disclosure patterns vary strongly by subsector, and a common backbone of pandas, NumPy, and scikit-learn coexists with toolkits tailored to text-intensive, tabular risk, payments, and market microstructure tasks. OCI values are effectively zero across subsectors, indicating that quantified outcome claims are rarely placed in the public record. The study delivers a subsector-resolved empirical catalog of Python adoption in finance and a replication-ready pipeline for measuring tool-level AI disclosure. Finance Artificial Intelligence and Machine Learning NYSE financial institutions Python disclosure libraries Outcome-Claims Index triangulation replication Figures Figure 1 1. Introduction Artificial intelligence (AI) is diffusing rapidly across financial services, where data-intensive business models, complex risk management tasks, and high-frequency markets create strong incentives for algorithmic decision support and automation. At the same time, Python has become the dominant language for analytics, modeling, and production pipelines in quantitative finance, trading infrastructure, and enterprise data platforms. Together, these trends have fostered a perception that listed financial institutions now operate with substantial embedded AI and Python capability. Yet public disclosure of such capability is sparse, heterogeneous, and largely unstandardized, which complicates empirical assessment of technology adoption and its economic implications. Existing disclosure regimes were not designed with AI or software capability in mind. Regulatory filings contain extensive financial and risk information but usually treat data and analytics infrastructure as background inputs rather than core strategic assets. Voluntary sustainability and governance reports often reference “digital transformation” or “innovation” at a high level, typically in marketing-oriented language. Sectoral press, employer portals, and job advertisements contain richer technical details, but coverage varies widely across firms and is rarely synthesized into a comparable view of capabilities at the level of programming languages, libraries, or analytical paradigms. As a result, the public record understates or distorts the technological frontier in financial institutions and obscures the boundary between signaling sophistication and documenting realized performance gains. The research literature mirrors this fragmentation. Studies of information technology and productivity tend to rely on investment data, surveys, or proprietary vendor usage metrics, which can offer depth but often lack transparency or replicability. Work on AI adoption in finance has focused on case studies, specific application domains such as credit scoring or fraud detection, or infrastructure-level metrics such as cloud usage. Separate strands of research track the Python ecosystem through open-source repositories and developer activity yet rarely connect such evidence to firm-level disclosure by listed financial institutions. Consequently, empirical knowledge about the prevalence, structure, and claimed outcomes of Python and AI capability at the firm level remains limited, despite strong policy and supervisory interest in model risk, operational resilience, and digital transformation. Research gap. Taken together, these observations point to a specific gap. First, we lack a systematic, firm-level map of tool-specific AI and Python adoption based solely on public disclosures, for a clearly defined financial universe. Second, we have little cross-subsector evidence on how Python-framed AI disclosures and associated tool stacks vary across distinct business models and institutional logics. Third, there is almost no quantitative evidence on whether public AI and Python statements are accompanied by explicit, quantified outcome claims (for example, risk reduction, accuracy gains, or efficiency improvements), or on how rare such claims truly are once uncertainty is accounted for. Existing studies are either abstract from programming languages and libraries or rely on data that cannot easily be replicated by other researchers. This paper addresses these gaps by constructing a systematic map of Python and AI disclosure for a strict frame of 180 New York Stock Exchange (NYSE) listed financial institutions. Public communication is treated as a noisy but informative signal of internal capability. A triangulated corpus combines regulatory filings, employer portals and job advertisements, corporate communications, and sectoral press over a collection window that extends through January 2026. Within this corpus, each firm is assigned to one of three disclosure states explicit Python, indirect AI, or none and named libraries are detected using sentence-level dictionary matching subject to contextual filters. Mentions are then mapped to seven analytical dimensions: natural language processing, machine learning, deep learning, reinforcement learning, probabilistic modeling, optimization, and visualization. Inverse-visibility weights mitigate the influence of firms that generate unusually rich or sparse public traces, and inferential tools include cross-tabulations, Pearson chi-squared tests with multiple-testing control, and compact robustness models with subsector fixed effects. The study asks three research questions: RQ1. Prevalence : How prevalent are explicit Python disclosures and indirect AI references among NYSE-listed financial institutions, once differences in public visibility and evidence channels are considered? RQ2. Structure and heterogeneity : How do Python-framed AI disclosures, associated library stacks, and analytical dimensions vary across financial subsectors characterized by different business models and institutional logics? RQ3. Outcome claims : To what extent do public disclosures contain quantified outcome claims about AI and Python systems, and what upper bounds can be placed on their true frequency in sectors and subsectors where no such claims are observed? A central conceptual distinction in answering these questions concerns adoption signals versus explicit claims about outcomes. To capture this distinction, the study introduces a sentence-level Outcome Claims Index (OCI) that flags whether a given item of evidence advances quantified performance assertions rather than simple capability statements or vague claims of improvement. The OCI supports both frequentist and Bayesian uncertainty summaries and provides a structured way to characterize the near absence of quantified outcome disclosure, even among firms that prominently advertise Python and AI usage. This distinction matters for regulators, investors, and counterparties that must understand not only whether institutions state that AI is in use, but also whether those institutions offer measurable evidence that AI materially alters risk, efficiency, or profitability. The empirical results indicate three broad patterns. First, explicit Python disclosure is widespread in the NYSE financial frame, although a non-trivial minority of institutions exhibits no observable Python or AI signal in the assembled corpus. Second, disclosure patterns vary strongly across subsectors, and a common backbone of general-purpose libraries coexists with more specialized toolkits for text-heavy processes, tabular risk modeling, payments networks, and market microstructure. Third, quantified outcome claims appear extremely rarely, and the associated bounds are consistent with an Outcome Claims Index that is effectively zero across subsectors. Explicit Python disclosure instead aligns with performance-adjacent proxies such as technology hiring intensity, technology and research and development expenditure, and overall visibility, which suggests that public communication currently reflects capability accumulation more than demonstrable performance change. Further contribution concerns transparency and replicability. A blinded replication archive accompanies the study and includes a harmonized roster of NYSE financial institutions, sentence-level evidence with Outcome Claims Index flags, and structured appendix workbooks. The replication worksheet documents the sampling frame, coding rules, visibility weights, and modeling choices in a form that enables independent reproduction of the core tables and figures, as well as straightforward extension to additional exchanges, sectors, or time periods. This design aligns the measurement of Python and AI disclosure with contemporary norms of open data and computational reproducibility in empirical finance and applied machine learning. The remainder of the paper is organized as follows. Section 2 situates the study within the literature on information technology, AI adoption, disclosure, and financial-sector digitalization. Section 3 describes the NYSE financial frame, data sources, coding procedures, and construction of the Outcome Claims Index and visibility weights. Section 4 presents the main descriptive and inferential results on disclosure states, library ecosystems, analytical dimensions, and sub-sectoral patterns. Section 5 reports robustness exercises based on logit and linear probability models and discusses the interpretation of potential selection and measurement biases. Section 6 concludes with implications for research, supervision, and disclosure practice, and outlines directions for future extensions of the replication worksheet and corpus 2. Literature Review 2.1 Python as infrastructure for financial AI Across contemporary finance, Python increasingly functions as the backbone that links data engineering, model development, and production deployment. Its appeal rests on a mature, interoperable open-source ecosystem: pandas and NumPy for data management and numerical linear algebra (McKinney, 2018 ), scikit-learn for baseline machine learning and model selection (Pedregosa et al., 2011 ), gradient-boosting frameworks such as XGBoost and LightGBM for high-accuracy tabular prediction at scale (Chen & Guestrin, 2016 ; Ke et al., 2017 ), and deep-learning stacks such as TensorFlow and PyTorch for representation learning and sequence models (Abadi et al., 2016 ; Paszke et al., 2019 ). Domain-specific toolkits extend this stack into core financial tasks: QuantLib supports fixed-income analytics and derivative pricing (Ametrano & Ballabio, 2013 ), while optimization and portfolio packages implement mean–variance and constraint-aware allocation routines. Taken together, these components create a vertical pathway from raw data to deployment and combined with abundant developer mindshare and open-source governance, have made Python the de facto lingua franca for AI-enabled analytics in finance. Existing work, however, typically documents Python’s capabilities or use cases rather than providing a field-level, firm-resolved map of where Python is explicitly acknowledged in the public record. This motivates a language-specific, disclosure-based perspective in the present study. 2.2 Application domains in finance and the role of Python Python libraries operationalize a wide spectrum of AI tasks that now feature prominently in financial research and practice. Credit and fraud analytics rely on supervised learning and boosting to balance accuracy and false-positive control in imbalanced settings (Chen & Guestrin, 2016 ; Ke et al., 2017 ). Time-series forecasting and volatility modeling employ LSTMs and related deep sequence models (Fischer & Krauss, 2018 ), while text mining of filings, earnings calls, and news uses NLP pipelines built on spaCy and transformer-based architectures (Loughran & McDonald, 2016 ). Market surveillance and systemic-risk diagnostics draw on network representations of propagation channels and correlated exposures (Haldane & May, 2011 ), often implemented with Python-based graph toolkits. This application-level literature shows that Python is not merely an implementation detail but a platform technology that standardizes data structures, model interfaces, and deployment idioms across diverse financial tasks. What it does not provide is a systematic, firm-level view of how such Python-enabled capabilities appear in public communications by listed institutions or how disclosure practices vary across subsectors. The present analysis takes those omissions as a starting point. 2.3 From bibliometrics to disclosures: measuring adoption in organizational settings A separate strand of work concerns how technology adoption is measured inside organizations. Traditional approaches such as bibliometrics, patent counts, case studies, or proprietary vendor telemetry are informative about innovation activity but seldom resolve who uses which concrete tools now within specific institutional niches such as listed financial institutions. They typically operate at the level of broad “AI” or “analytics” categories and do not distinguish between programming languages, libraries, or analytical paradigms at the firm level. Organizational research has long advocated triangulation, deliberately combining evidence sources with distinct logics (e.g., regulatory filings, corporate materials, sectoral press, and labor-market signals) to strengthen construct validity and reduce single-source bias (Denzin, 1978 ; Jick, 1979 ; Patton, 2015 ; Yin, 2018 ). Yet despite these recommendations, most empirical work on AI adoption in finance either relies on one dominant channel (such as surveys or filings) or blends heterogeneous samples across markets, making it difficult to know which population is being described. This leaves three related measurement gaps. First, there is no systematic, tool-specific map of AI and Python adoption based solely on public disclosures for a clearly defined population of listed financial institutions. Second, existing studies rarely tackle visibility bias directly: highly vocal institutions dominate public channels, inflating apparent adoption rates unless disclosure intensity is explicitly modelled. Third, the non-disclosure paradox firms adopting technologies internally yet avoiding naming stacks or quantifying impacts in public is usually treated as an anecdotal caveat rather than as measurable regularity. The methodological stance used here responds to these gaps by combining a strictly defined NYSE finance population with multi-channel evidence and explicit coding rules for disclosure status. In this design, triangulation is not just a robustness check but the core of the measurement strategy: it treats visibility bias as a parameter to be adjusted for and non-disclosure itself as informative variation across firms and subsectors. This positioning directly motivates RQ1 on the prevalence of explicit Python and indirect AI disclosure and RQ2 on the structure and subsector heterogeneity of tool stacks and disclosure granularity. 2.4 Signaling, institutional logic, and disclosure granularity Why do some firms explicitly say “Python”, whereas others only reference “AI” or “automation”? Signaling theory predicts that organizations disclose just enough technical detail to convey competence to target audiences while minimizing competitive and regulatory exposure (Spence, 1973 ). In finance, institutional logics shape which details are perceived as legitimate signals: market-facing subsectors (e.g., trading and market infrastructure) tend to foreground technical capability, while credit vehicles often emphasize prudential themes and operational reliability (Coffee, 2020 ). Within this framework, language-level statements (“Python”) function as relatively low-risk, high-recognition signals, easily understood by investors, staff, and regulators. Library-level naming (specific packages) is a higher-granularity signal that may be strategically withheld because it reveals more about internal architectures and vendor relationships. Case studies and qualitative analyses consistently find that language mentions are common while named libraries are rare, suggesting that granularity is an analytic variable reflecting audience, risk posture, and competitive context, not simply the degree of internal adoption. Existing empirical work, however, typically stops at conceptual arguments; it does not quantify how granularity and explicitness vary across subsectors in a large, clearly defined population. The present study addresses this by treating disclosure granularity itself as an outcome of interest and by embedding it in the cross-subsector diagnostics that feed into RQ2. 2.5 From adoption to outcomes: the measurement gap A persistent critique of the AI adoption literature is that it maps use but rarely measures impact. Studies typically report whether AI or advanced analytics are present, which functions they support, and sometimes how extensively they are deployed, but they provide far less evidence on quantified outcomes such as return on investment, error reductions, risk or loss improvements, or latency and efficiency gains. This problem is particularly acute in regulated sectors such as finance, where public claims about performance can invite supervisory scrutiny, litigation risk, or competitive disadvantage. As a result, firms frequently name AI, machine learning, or Python yet refrain from publishing specific metrics that would document the economic or risk consequences of those systems. Two methodological responses have emerged. One advocates direct outcome designs, for example, event studies around dated AI announcements, confidential surveys, or access to internal telemetry to estimate economic effects more cleanly (cf. productivity debates in digital transformation; Brynjolfsson & McAfee, 2017 ). The other emphasizes performance-adjacent indicators observable in public data, such as disclosure intensity, channel breadth, or technology hiring, as proxies for capability maturity while carefully distinguishing these from audited key performance indicators. Both approaches share a limitation: they seldom treat the absence of quantified outcome claims in public disclosures as data, and they rarely provide formal uncertainty statements about how rare such claims might be within specific subsectors. Consequently, there is a persistent measurement gap between adoption signals and documented outcomes. We lack both a systematic accounting of how often financial institutions publicly attach numbers to their AI and Python systems and a framework for turning subsector-level “zero events” into interpretable bounds on the underlying rate of outcome disclosure. The present study addresses this gap by introducing a sentence-level Outcome Claims Index (OCI) that explicitly distinguishes adoption statements from quantified performance claims and by pairing it with both frequentist and Bayesian interval estimates for sparse-event strata. In doing so, the analysis reframes the near absence of outcome claims as a measurable empirical pattern rather than as a mere limitation of the data. This perspective underpins RQ3, which asks to what extent public disclosures contain quantified outcome claims about AI and Python systems and what upper bounds can be placed on their true frequency when none are observed. 2.6 Statistical treatment of cross-subsector patterns Qualitative studies of digitalization in finance often note that disclosure and adoption are concentrated in certain subsectors, but they rarely subject these impressions to formal statistical tests. Where cross-sector comparisons are made, researchers typically rely on contingency tables, χ² tests of independence, and associated effect sizes such as Cramér’s V to assess association strength, sometimes supplemented by post-hoc contrasts and multiple-testing adjustments. Visibility bias is usually acknowledged but not quantitatively addressed, and multivariate models that control both subsector and simple proxies for visibility remain uncommon in the AI adoption literature. This methodological landscape suggests the value of combining visibility-aware cross-tabulations with parsimonious regression diagnostics. In the present study, these tools are used not to claim causal effects but to translate intuitions from signaling and institutional logics into falsifiable, replicable statistics for the finance context. By doing so, the analysis helps move from impressionistic claims about “innovative subsectors” towards documented patterns of disclosure and explicitness that can be compared across markets and over time. 2.7 Python versus other languages: a language-specific lens Tool-agnostic surveys tend to pool languages and platforms, but a language-specific lens can reveal meaningful asymmetries. Python’s dominance in data-science education, scientific computing, and MLOps integration makes it more observable in public artefacts than R, MATLAB, or SAS, particularly where cloud-native deployment and model orchestration are salient (McKinney, 2018 ; Pedregosa et al., 2011 ). At the same time, some subsectors retain specialized stacks such as SAS in actuarial workflows, for example, or Java/Scala in legacy streaming and trading infrastructure which underscores why stack granularity should be treated as a disclosure variable rather than a direct proxy for internal prevalence. Existing work documents these ecosystems but does not quantify, at the firm level, how often different languages are named in public communications or how language mentions co-vary with sector, business model, and disclosure granularity. A language-specific approach, coupled with explicit caveats about what disclosure can and cannot reveal about internal systems, therefore improves both transparency and replicability in adoption mapping. It also justifies the study’s focus on Python-framed disclosure while benchmarking against alternative languages. 2.8 Governance, ethics, and stability considerations Even when Python-enabled models deliver internal value, governance considerations such as model opacity and validation, fairness and bias, third-party concentration, and cyber-security exposure shape what firms choose to disclose. The governance literature emphasizes documentation of model purpose, data lineage, validation and challenge, post-deployment monitoring (drift and stability), and human-in-the-loop controls as necessary conditions for trustworthy AI in finance (Coffee, 2020 ). Systemic-risk work warns that common data and model dependencies can propagate correlated errors and amplify shocks (Haldane & May, 2011 ). These concerns help explain why public statements often remain at the language level and why quantitative outcome claims are scarce: transparency must be balanced against security, competition, and prudential discipline. In this environment, naming Python and AI can serve to signal modern capability without revealing detailed model performance or stack composition. This literature therefore not only motivates the OCI’s focus on outcome claims but also frames the interpretation of “zeros” in outcome disclosure as a strategic equilibrium rather than as mere data sparsity. 2.9 Synthesis and positioning Literature collectively points to three unresolved issues. First, while application-level work documents numerous AI use cases in finance and Python has emerged as a de facto platform for these systems, we still lack a systematic, firm-level map of tool-specific AI and Python adoption based solely on public disclosures for a clearly delimited financial universe. Second, prior studies have not quantified how disclosure granularity and stack composition vary across subsectors governed by different institutional logics, even though case evidence suggests substantial heterogeneity in how explicitly firms talk about their tools. Third, existing research largely stops at adoption signals; it does not measure how often public AI and Python statements are accompanied by explicit, quantified outcome claims, nor does it provide uncertainty bounds when such claims are absent. The present study responds to these gaps by (i) focusing on a strict NYSE finance universe to ensure population clarity; (ii) implementing a triangulated, multi-channel measurement design with explicit coding rules for disclosure status and stack granularity; and (iii) introducing a sentence-level Outcome Claims Index (OCI) that operationalizes outcome claims and treats zeros as informative through both frequentist and Bayesian interval estimates. This design directly supports the three research questions set out in the Introduction: the prevalence of Python-framed and indirect AI disclosure (RQ1), the structure and subsector heterogeneity of AI/Python tool stacks and disclosure granularity (RQ2), and the frequency and interpretation of quantified outcome claim in public materials (RQ3). In doing so, it complements tool-agnostic surveys and proprietary usage studies by offering a replication-ready, disclosure-based map of Python-enabled AI in finance and by establishing performance-adjacent indicators that future work can link to direct outcome measures. 3. Data and Methods 3.1 Population and Sampling Frame The empirical universe consists of one hundred eighty (180) New York Stock Exchange (NYSE) listed financial institutions drawn from a single-exchange frame to ensure homogeneity in disclosure obligations and to avoid cross-listing ambiguity. Each issuer is assigned to a harmonized subsector taxonomy reflecting dominant business models and regulatory logics: Diversified Banks, Regional Banks, Custody & Asset Servicing, Investment Banks, Asset Managers, Asset Managers (Alternatives), Insurance, Payments & Consumer Finance, Market Infrastructure & Data, BDC/Finance, Mortgage REIT/Finance, Fixed Income e-Trading, and Investment, Tech & Risk. The cross section is fixed at N = 180 following ticker deduplication and consolidation of merger and delisting events, yielding a one-row-per-issuer design suitable for transparent replication. Representative firms, counts, and inclusion rationales are reported in Appendix Table 3.1. This strict and fully enumerated NYSE frame provides the population clarity required for the prevalence and heterogeneity analyses in RQ1 and RQ2. 3.2 Evidence Sources and Acquisition Data collection follows a triangulated strategy that integrates four public evidence channels with complementary institutional logics: (i) legally accountable regulatory filings (e.g., 10‑K/20‑F, risk and MD&A discussion); (ii) employer portals and job postings; (iii) corporate communications and sectoral press; and (iv) product and engineering documentation. Triangulation increases construct validity by reducing single‑source bias and by exploiting the different incentives that shape each channel (Denzin, 1978; Jick, 1979; Patton, 2015; Yin, 2018). For each evidence item, the dataset records the firm identifier and ticker, subsector, source channel, URL or document ID, evidentiary sentence(s), and retrieval date. Where available, filing dates and press‑release timestamps are retained to facilitate future event‑study designs. The collection window closes in January 2026, producing a time‑stamped, replication‑ready cross‑section of public communication. 3.3 Coding Scheme and Disclosure Granularity Disclosure is coded at the firm level using mutually exclusive status labels derived from convergent evidence: explicit Python when the public text directly names “Python” and/or specific Python libraries. indirect AI when the text discloses AI/ML/automation without naming the language. none when no AI/ML reference appears in the collection window. Coding follows a hierarchical rule: if any channel contains an explicit Python reference, the firm is coded as explicit Python; if no explicit Python references are found but AI/ML/automation is mentioned, the firm is coded as indirect AI; only firms with no AI/ML reference in any channel are coded as none. Because technical specificity varies strategically across channels, the corpus also records library‑level mentions when they occur (e.g., pandas/NumPy for data handling, scikit‑learn for classical ML, XGBoost/LightGBM for boosting, TensorFlow/PyTorch for deep learning), with string normalization and canonical family mapping to ensure comparability (Abadi et al., 2016; Chen & Guestrin, 2016; Ke et al., 2017; McKinney, 2018; Paszke et al., 2019; Pedregosa et al., 2011). Consistent with signaling theory and institutional logics, language‑level statements are interpreted as capability signals, whereas library‑level naming is treated as disclosure granularity rather than a census of internal stacks (Spence, 1973; Coffee, 2020). 3.4 Python Library Detection and Subsector Stacks Subsector‑level AI and Python tool adoption is summarized in a five‑column exhibit that reports “Subsector”, “Representative Companies”, “Analytical Dimensions”, “Primary AI/Python Use Cases”, and “Representative Python/AI Stack”. This summary is based solely on in‑sample evidence from the corpus and is reported in Appendix Table 4. Named tools are identified via sentence‑level pattern matching against a predefined dictionary of libraries and platforms (e.g., pandas, NumPy, SciPy, scikit‑learn, XGBoost, LightGBM, TensorFlow, PyTorch, spaCy, transformer libraries, PyMC/Pyro, NetworkX, PyTorch Geometric, CVXPY, and vendor SDKs such as GS Quant and AladdinSDK). The procedure follows standard text‑as‑data and dictionary‑based coding approaches (Grimmer & Stewart, 2013; Gentzkow, Kelly, & Taddy, 2019) and widely used Python ML/NLP toolchains (Pedregosa et al., 2011; Abadi et al., 2016; Paszke et al., 2019; Vaswani et al., 2017). Matches are retained only when library tokens appear near finance‑relevant actions or contexts (e.g., underwriting, pricing, execution, deploy, production, fraud, surveillance, risk). Ambiguous cases are manually reviewed. Validated detections are aggregated to the subsector level and read alongside local context to infer (i) dominant analytical dimensions, (ii) primary AI/Python use cases, and (iii) characteristic stacks, in line with common typologies in the ML/AI literature (Bishop, 2006; Goodfellow, Bengio, & Courville, 2016). 3.5 Thematic Dimensions and Theme Library Mapping To impose a consistent structure on otherwise heterogeneous AI and Python disclosures, all in‑sample evidence is recoded into seven analytical dimensions: natural language processing (NLP), machine learning (ML), deep learning (DL), reinforcement learning (RL), probabilistic modeling, optimization, and visualization. These dimensions follow standard typologies in the ML literature and provide a mutually exclusive, collectively exhaustive partition of the main modeling and analytics workstreams observed in the corpus (Bishop, 2006; Goodfellow, Bengio, & Courville, 2016). Each dimension is then mapped to canonical Python tool families, operationalizing the link between thematic capability and concrete implementation. Supervised tabular ML is associated with scikit‑learn, XGBoost, and LightGBM; DL with TensorFlow and PyTorch; NLP with spaCy and transformer‑based libraries; probabilistic modeling with PyMC/Pyro and related Bayesian frameworks; and optimization and quantitative finance with tools such as CVXPY, PyPortfolioOpt, and QuantLib. This theme library mapping is applied uniformly across channels and subsectors and serves as the organizing device for Appendix Table 4 and the cross‑subsector diagnostics in Section 4. 3.6 Outcome‑Claims Index (OCI) and Inference for Zeros Each sentence in the corpus is automatically scanned for quantified performance claims (e.g., percentage improvements, model metrics, loss or risk reductions, latency or efficiency gains, or financial KPIs) using standard text‑as‑data techniques (Grimmer & Stewart, 2013; Gentzkow, Kelly, & Taddy, 2019). A binary indicator is assigned at the sentence level and then aggregated to firm‑ and subsector‑level shares to form the Outcome Claims Index (OCI). To address sparse‑event strata (x = 0), uncertainty is summarized using two complementary methods. First, the frequentist “rule of three” provides a 95% upper bound (≈ 3/n) as a conservative maximum rate approximation (Hanley & Lippman‑Hand, 1983). Second, Bayesian credible intervals are computed using Jeffreys prior, Beta (x + 0.5, n − x + 0.5), which yields well‑behaved finite‑sample intervals at the boundary (Agresti, 2019; Gelman et al., 2014). OCI point estimates and both interval sets are reported in Appendix Tables S4.5d–S4.5e and visualized in Appendix Figure S4.5e; full strata-level computations appear in the replication workbook sheet S4.5. 3.7 Language Benchmarking and Statistical Diagnostics Named programming language mentions (Python, R, Java, SAS, Scala, MATLAB) are tallied at firm and subsector levels to contextualize Python’s prominence in public disclosures. The resulting cross tabs appear in Appendix Table S4.6 and underpin Figure 4c, with underlying counts provided in the replication workbook. Firm-level disclosure status (explicit/indirect/none) is cross-tabulated by subsector, and inference uses Pearson χ² with Cramér’s V as an effect size and Benjamini–Hochberg correction on post hoc standardized residual contrasts. These diagnostics are reported in Appendix Tables S4.2–S4.4 and interpreted in Section 4, confirming systematic sub sectoral heterogeneity in disclosure modes for the NYSE frame. 3.8 Data Hygiene, Reliability, and Limitations Quality control proceeds in three passes: issuer harmonization (standardized names and tickers, subsector taxonomy), dual-source verification at the evidence-row level, and manual spot checks of named library matches and exemplars used in Appendix Table 4. These steps improve validity but do not eliminate visibility bias or channel asymmetries; both are documented and treated as part of the phenomenon under study, the politics of disclosure, rather than as purely random noise, and they are revisited in the limitations section. 3.9 Reproducibility and replication materials All empirical results are fully reproducible. A blinded replication archive (submitted as Additional file 1 ) accompanies the manuscript and contains: (i) a harmonized roster of the 180 NYSE financial institutions with tickers, standardized names, subsector assignments, and disclosure status; (ii) the sentence‑level evidence corpus with source channel, retrieval dates, and Outcome Claims Index flags; (iii) language‑benchmark counts and visibility‑weighted disclosure shares; (iv) the appendix workbook with all statistical tables and diagnostics; and (v) the scripts/notebooks and configuration file used to generate the derived variables, cross‑tabulations, figures, and robustness checks. The archive is organized so that the complete set of main‑text and appendix outputs can be rebuilt from a single configuration, facilitating independent verification and extension to other exchanges, sectors, or time periods. 3.10 Availability of data and materials The dataset(s) and code supporting the conclusions of this article are included within the article and its additional file(s). A blinded replication archive (Additional file 1) contains the harmonized firm‑level roster, the sentence‑level Python/AI disclosure corpus with Outcome Claims Index flags, language‑benchmark counts, the appendix workbook with all statistical tables, and the scripts/notebooks required to reproduce the analyses. 4. Data Analysis and Results This section presents the empirical results for the strict NYSE frame of 180 financial institutions and answers the three research questions introduced in Section 1. RQ1 concerns the prevalence of explicit Python and indirect AI disclosure. RQ2 examines cross‑subsector structure and heterogeneity in disclosure modes and Python‑framed stacks. RQ3 addresses the frequency and interpretation of quantified outcome claims, via the Outcome Claims Index (OCI). All results rely on the triangulated corpus, coding scheme, and statistical tools described in Section 3. Because the evidence is drawn from regulatory filings, employer portals and job postings, corporate materials, and sectoral press, all estimates should be read as disclosure frequencies what firms choose to place in the public record rather than as a census of internal tool use. This is consistent with the triangulated design in the literature review, which treats visible traces as noisy but informative signals of underlying capability (Denzin, 1978; Jick, 1979; Patton, 2015; Yin, 2018). All statistical tables referenced below are reported in the Appendix and in the harmonized workbook Tables_S4_all.xlsx within the replication archive. 4.0 Python library adoption and analytical scope To contextualize RQ2 on subsector heterogeneity, Appendix Table 4 summarizes AI and Python tool adoption across representative NYSE financial subsectors. The exhibit translates the theme–library mapping from Section 3.5 into concrete subsector profiles, linking analytical dimensions, primary AI/Python use cases, and representative stacks. Across subsectors, a common Python backbone is visible: pandas and NumPy for data handling and numerical routines; scikit‑learn plus gradient‑boosting frameworks such as XGBoost, LightGBM, and CatBoost for supervised learning; and deep‑learning frameworks such as TensorFlow and PyTorch for more complex representation and sequence models. Probabilistic programming tools such as PyMC or Pyro appear were unstructured data, complex loss distributions, or multi‑scenario analysis feature prominently. Asset managers and alternative asset managers make extensive use of NLP, ML, DL, probabilistic modeling, optimization, and visualization. In these subsectors, Python‑based stacks support factor and risk models, portfolio construction, performance attribution, fund flow and distribution analytics, and ESG or covenant text analysis, often in conjunction with portfolio‑optimization libraries and specialized platforms such as Aladdin SDKs. Business development companies and mortgage‑finance firms apply similar tools to middle‑market or mortgage credit, with an emphasis on underwriting, early‑warning signals, and prepayment or default modeling rather than broad multi‑asset allocation. Banks, insurers, and consumer‑finance firms adapt the same core stack to retail, commercial, and insurance workflows. Diversified and regional banks use NLP and machine learning for contract intelligence, chatbots, KYC and AML monitoring, stress testing, deposit and cash‑flow forecasting, and branch or customer profitability analysis, typically wrapped in dashboard frameworks such as Dash or Streamlit. Insurers deploy boosting and deep‑learning models for pricing, claims triage, lapse and surrender behavior, telematics, and catastrophe risk, supported by NLP for policy, medical, and claims text. Payments‑ and card‑centric firms use graph and sequence models for real‑time fraud detection, merchant risk, chargeback prediction, and personalization, often in streaming or near real‑time Python environments. Market infrastructure, data providers, investment banks, and electronic trading platforms show the most intensive use of deep learning, reinforcement learning, and optimization. In these subsectors, Python stacks support tick‑data analytics, transaction‑cost analysis, smart order routing, derivatives pricing and XVA, surveillance, and balance‑sheet or funding optimization. They combine the scientific‑Python stack with distributed computing tools, graph and sequence‑modeling frameworks, reinforcement‑learning libraries, and vendor Python APIs such as GS Quant or Aladdin SDKs. Appendix Table 4 indicates that NYSE‑listed financial institutions share a narrow but powerful set of foundational Python tools and differentiate sharply in analytical dimensions and higher‑level stacks. NLP and deep learning are most salient in text‑heavy, surveillance, and fraud‑oriented activities, while probabilistic modeling is most salient in credit, insurance, and mortgage‑risk contexts. These patterns implement, in empirical form, the literature review’s argument that Python functions as a field‑level platform technology whose concrete expression varies by subsector (Sections 2.1–2.2). 4.1 Descriptive coverage (RQ1) Turning directly to RQ1 (prevalence), the harmonized NYSE sample contains 180 financial institutions. Three disclosure states are coded at the firm level: explicit Python (the firm names Python or a Python library), indirect AI (AI/ML is disclosed but no language is named), and none. Unweighted counts show 138 firms (76.67%) with explicit Python mentions, 4 (2.22%) with indirect AI mentions, and 38 (21.11%) with no disclosure (Figure 4a; Appendix Table S4.1). These shares mirror the headline numbers reported in the abstract and provide a first answer to RQ1: most NYSE financial institutions publicly acknowledge Python in at least one channel. Because public‑facing evidence is uneven across firms and channels, a visibility‑adjusted estimate is computed by weighting each issuer by . The inverse‑visibility‑weighted explicit share equals 0.629, indicating that explicit Python disclosure remains dominant even after down‑weighting highly visible firms (Appendix Table S4.1, “shares”). This pattern is consistent with the triangulated design’s expectation that combining filings, corporate releases, technical press, and job postings will surface a wide footprint of adoption while still reflecting visibility bias (Denzin, 1978; Jick, 1979; Patton, 2015; Yin, 2018). 4.2 Subsector differences in disclosure (RQ2) RQ2 asks how disclosure modes and tool stacks vary across subsectors. Firm‑level disclosure status (explicit Python, indirect AI, or none) differs markedly by subsector. A Pearson χ² test on the status × subsector table rejects independence (χ² (40, N = 180) = 100.65, p < .001), with Cramér’s V = 0.529 indicating a moderate–large association (Appendix Table S4.4, Panel A; Agresti, 2019). The row‑share heatmap (Figure 4b) makes the pattern visually obvious. Investment Banks, Market Infrastructure & Data, and Payments report explicit library‑ and language‑level disclosures at much higher rates, while Mortgage REIT/Finance and BDC/Finance display low explicitness. Standardized residuals in Appendix Table S4.2 highlight the subsector–status pairs that contribute most to the χ² statistic, and Benjamini–Hochberg–adjusted pairwise contrasts (Appendix Table S4.4, Panel B) show the largest and most robust gaps, for example, between Investment Banks (high explicit disclosure) and Mortgage REIT/Finance (low explicit disclosure). These differences map onto distinct task profiles and institutional logics (Section 2.4). Trading, surveillance, and payments businesses are data‑ and text‑intensive, operate under strong latency and automation pressures, and compete in technical labor markets, conditions that favor explicit naming of modern toolchains. Credit‑ and yield‑focused vehicles, by contrast, face stronger disclosure tradeoffs and reputational or regulatory incentives that encourage vaguer AI language or silence. Subsector heterogeneity in explicitness thus reflects the signaling and governance tensions highlighted in the literature review and confirms that disclosure granularity is itself an analytic variable rather than a simple proxy for internal adoption. 4.3 Outcome Claims Index (OCI): interpreting the “zeros” (RQ3) RQ3 focuses on quantified outcome claims. To separate adoption from results, the corpus is scanned for sentences that report measurable outcomes such as performance improvements, model metrics (e.g., AUC/F1), loss or risk reductions, latency or efficiency gains, or financial KPIs. These hits form the Outcome Claims Index (OCI) described in Section 3.6. Across subsectors, OCI hits are essentially absent at both the row and firm levels (Appendix Tables S4.5a–S4.5c). This near‑zero pattern should not be read as evidence of “no effect” from AI or Python systems but as a disclosure norm: NYSE institutions rarely publish quantified results even when they disclose Python/AI adoption. To make “zero” informative, two complementary checks are reported. First, rule‑of‑three 95% upper bounds translate zero events into conservative limits on the true rate (≈ 3/n when x = 0; for example, with 30 evidence rows and no claims, the true frequency is < 10%; Appendix Table S4.5d). Second, Jeffreys‑prior credible intervals (Beta [x + 0.5, n − x + 0.5]) place posterior mass tightly near zero in every subsector; the forest plot (Appendix Table S4.5e and Appendix Figure S4.5e) displays posterior means and 95% intervals. Methodologically, pairing frequentist bounds with Bayesian intervals converts the OCI “zeros” into usable evidence (Agresti, 2019) and directly addresses the measurement gap identified in Section 2.5. The results answer RQ3: quantified outcome claims are rare across all subsectors, and the uncertainty bounds are tight enough to rule out anything but low underlying frequencies in the public record. 4.4 Language benchmarking To corroborate the Python‑specific lens developed in Sections 2.1–2.2 and the abstract, sentence‑level tallies compare named mentions of Python, R, Java, SAS, Scala, and MATLAB. Python mentions far exceed those of any other language across subsectors. The language‑by‑subsector heatmap (Figure 4c) shows Python present throughout, with other languages appearing only sporadically. This benchmark substantiates the study’s Python focus and aligns with the centrality of the open‑source stack in modern analytics (Pedregosa et al., 2011; Abadi et al., 2016; Paszke et al., 2019; McKinney, 2018). Full counts are reported in Appendix Table S4.6. 4.5 Performance‑adjacent proxies: mini‑models Given the scarcity of direct outcome claims (RQ3), Section 4.5 explores performance‑adjacent proxies that can shed light on capability accumulation without asserting causal effects. Associations are examined between explicit Python disclosure and proxies such as technology‑hiring intensity, technology and R&D operating‑expense ratios, and visibility (evidence volume). A logit specification (reported as odds ratios) and an OLS linear‑probability model with HC1 robust standard errors, both including subsector fixed effects, indicate that these proxies are positively associated with explicit Python disclosure (Appendix Tables S4.7A–S4.7B; Agresti, 2019). For interpretability, Table S4.7B reports marginal‑effect coefficients, and multiple‑comparison control follows Benjamini–Hochberg (1995). These mini‑models do not claim that Python disclosure causes better performance. Instead, they triangulate the interpretation consistent with the literature on signaling and capability building that explicit disclosure co‑moves with broader indicators of digital investment and visibility. 4.6 Robustness and sensitivity Two diagnostics support the stability of the main findings. First, inverse‑visibility weighting lowers the explicit share from 0.767 to 0.629 yet preserves the conclusion that explicit disclosure is widespread, indicating that results are not an artifact of a few highly visible issuers (Appendix Table S4.1, “shares”). Second, Benjamini–Hochberg multiple‑comparison control preserves the principal subsector gaps in explicit disclosure (Appendix Table S4.4, Panel B). Diagnostic cross‑tabs and standardized residuals trace the same over‑ and under‑indexing pattern reported in Figure 4b (Appendix Tables S4.2–S4.3; Agresti, 2019; Benjamini & Hochberg, 1995). Annual counts of evidence sentences show a clear recent increase in public artifacts related to AI/Python (Appendix Figure S4.6). While visibility is not capability, the time trend is consistent with a secular rise in the salience of AI/Python adoption across NYSE financial institutions. 4.8 Summary of findings The evidence from the NYSE frame provides a coherent answer to RQ1–RQ3. Explicit Python disclosure is common and remains dominant after visibility adjustment (RQ1). Disclosure varies systematically by subsector, with high data‑intensity businesses such as investment banking, market infrastructure, and payments exhibiting the highest explicitness, and mortgage REITs and business‑development companies relying more on generic AI language or silence (RQ2). Public quantification of outcomes is rare, but rule‑of‑three bounds and Jeffreys‑prior intervals show that this scarcity is itself informative, placing tight upper bounds on the true frequency of quantified claims even in subsectors that make numerous adoption statements (RQ3). Python’s dominance over other named languages is clear, both in the number of mentions and in subsector coverage, which validates the Python‑centered focus of the study. Performance‑adjacent proxies such as technology‑hiring intensity, technology and R&D expenditure, and public visibility move in the expected direction with explicit disclosure and are consistent with the interpretation of disclosure as a signal of capability accumulation rather than proof of superior realized outcomes. All findings are robust to visibility adjustment and multiple‑comparison control and are fully documented in the Appendix and replication package. 5. Discussion This section interprets the empirical patterns documented in Section 4 considering the research gap and questions articulated in the Introduction and Literature Review. The triangulated corpus integrates filings, employer portals, corporate materials, and sectoral press, so all inferences concern public disclosure frequencies rather than private internal use. Consistent with the methodological stance in Section 2.3, visible traces are treated as noisy but informative proxies for underlying capability, not as direct measurements of internal systems (Denzin, 1978; Jick, 1979; Patton, 2015; Yin, 2018). 5.1 Interpreting the core patterns (RQ1–RQ3) Three headline facts organize the interpretation. First, explicit Python disclosure is prevalent even after inverse‑visibility reweighting, which indicates a broad public adoption signal across NYSE financial institutions (RQ1). The visibility‑adjusted explicit share remains well above one‑half of the sample, so the dominant picture is that Python is openly named rather than hidden in most firms. Second, subsector heterogeneity is statistically and substantively meaningful (RQ2). Investment Banks, Market Infrastructure & Data firms, and Payment’s providers are systematically more likely to disclose explicit Python and named libraries, while Mortgage REIT/Finance and BDC/Finance are systematically less explicit, as documented in Section 4.2 and Appendix Table S4.4. These differences align with institutional logics: market‑facing businesses compete on speed, analytics, and technical credibility, whereas credit vehicles emphasize prudence and stability. Third, the Outcome Claims Index is effectively zero across subsectors, even where adoption signals are strong (RQ3). Python’s dominance over R, Java, SAS, Scala, and MATLAB is visible across subsectors (Figure 4c and Appendix Table S4.6). Together, these facts describe a field that converges on a common technical lingua franca while diverging in how and why adoption is disclosed—exactly the tension anticipated in the literature between platform convergence and heterogeneous signaling strategies (Sections 2.1–2.5). 5.2 The non‑disclosure paradox and the OCI “zeros” The near‑zero OCI values operationalize the non‑disclosure paradox highlighted in Section 2.5. Institutions routinely name tools and sometimes describe use cases, yet rarely publish quantified outcomes such as return on investment, changes in error rates, risk or loss reductions, or latency and efficiency gains. The study therefore converts zero counts into information via rule‑of‑three upper bounds and Jeffreys‑prior credible intervals, following guidance on sparse data and binomial inference (Agresti, 2019). Both procedures show that in public channels the frequency of quantified outcomes is small across subsectors, even where adoption is widely signaled. This pattern reconciles a tension visible in the broader literature. External narratives and survey‑based reports stress rapid AI adoption in finance, yet verifiable performance metrics and detailed model diagnostics remain largely internal or confidential. The empirical OCI “zeros” quantify that gap. Public texts validate that AI and Python systems exist but rarely document how those systems alter risk or efficiency in ways that can be audited by outsiders. In this sense, the study does not contradict evidence of real impact; it shows that public evidence of impact is thin, even where internal adoption may be deep. 5.3 Theoretical implications: signaling, institutional logic, and isomorphism The findings extend signaling accounts by showing that naming languages and libraries functions as audience‑specific credibility work rather than neutral technical description. In signaling models, more informative signals tend to be chosen by higher‑capability types whose marginal cost of signaling is lower than for lower‑capability types (Spence, 1973). In the NYSE financial frame, explicit naming of Python, scikit‑learn, XGBoost, LightGBM, transformer libraries, and proprietary SDKs anchors institutions to a specific technical trajectory and invites benchmarking by technically literate investors, staff, and regulators. Variation across subsectors aligns with the institutional‑logics perspective, which emphasizes that sectors are governed by distinct logics of action that shape what counts as legitimate behavior and talk (Thornton & Ocasio, 2008). Market‑facing logics in trading, market‑data, and payments privilege innovation signaling and speed, so explicit AI and Python disclosures that emphasize advanced architectures and automation are congruent with expectations of technical excellence and agility. Prudential logics in credit‑focused vehicles privilege control signaling, stability, and conservative underwriting, so muted or generic AI language better supports dominant narratives about risk discipline. The near absence of quantified outcome claims fits with neo‑institutional accounts of decoupling, in which organizations adopt structures and vocabularies to satisfy external expectations while keeping some distance from core technical practices (Meyer & Rowan, 1977). In this case, naming Python and AI serves as a credible signal about inputs, but disclosure about outputs remains thin. Technologies thus operate simultaneously as production tools and as symbols in organizational communication (Orlikowski & Iacono, 2001). At the same time, the widespread convergence on a shared Python backbone reflects isomorphic pressures in the organizational field (DiMaggio & Powell, 1983). The fact that almost every subsector relies on pandas, NumPy, scikit‑learn, and a short list of deep‑learning and boosting libraries, even though alternative ecosystems remain available, is consistent with coercive pressures from regulators and large counterparties, mimetic responses to uncertainty, and normative influences from professional and developer communities. Professional training, vendor ecosystems, and open‑source communities all present this stack as the canonical solution for financial analytics, and listed institutions reproduce that template in their own disclosures. The result is a layered structure: a shared base stack provides a generic solution to data and model engineering, while subsector‑specific modules (e.g., graph toolkits in payments, probabilistic loss modeling frameworks in credit and insurance) encode local business models and risk profiles. 5.4 Practical, regulatory, and governance implications For boards and executives, the findings support a shift from viewing AI and Python as a loose collection of tools toward a conception of governed stacks. In such a conception, explicit lifecycle controls including model inventorying, validation cadence, drift and performance monitoring, and decommissioning criteria are applied consistently across models that rely on the shared Python backbone. Reproducible pipelines, documented data lineage, and “red‑team” routines for adversarial robustness and stress scenarios become central governance artefacts rather than optional best practice. The observed dominance of a small number of libraries implies that investments in controls and documentation around those components can yield broad risk‑management benefits across business lines. For investors, explicit language and library mentions function as noisy but useful proxies for digital capability. Section 4.5 shows that explicit Python disclosure is positively associated with technology‑hiring intensity, technology and R&D operating expenditure, and evidence volume. These associations are descriptive, not causal, but they suggest that disclosure can be combined with labor‑market signals and platform partnerships to sharpen assessments of AI maturity and digital resilience. For regulators and supervisors, the near‑zero Outcome Claims Index implies a visibility gap in quantified effects. If public materials continue to omit performance, fairness, and resilience metrics, supervisory frameworks may need to rely more heavily on confidential reporting channels and direct examinations to understand the impact of AI on risk. Lightweight templates for outcome summaries, fairness diagnostics, and model‑family‑level validation metadata could increase public assurance without forcing exposure of proprietary details. Finally, given the concentration of language and library dependence on Python and a handful of tools, supervisory stress testing and operational‑resilience planning should incorporate dependency and common‑mode failure scenarios, such as vulnerabilities in widely used libraries or correlated outages in shared toolchains an extension of existing model‑risk guidance that emphasizes inventories, dependency mapping, and governance for advanced analytics. 5.5 Contributions to research The study makes three main contributions that directly address the gaps identified in Section 2 and formalized in RQ1–RQ3. First, it delivers systematic, cross‑subsector evidence on Python‑framed AI disclosure frequencies for a strict NYSE finance universe, using a triangulated corpus that combines filings, employer portals, corporate materials, and sectoral press. This approach advances empirical baselines for the finance‑and‑AI literature and operationalizes long‑standing calls for multi‑source triangulation in organizational research (Denzin, 1978; Jick, 1979; Patton, 2015; Yin, 2018). Second, it formalizes “zeros” as evidence by pairing rule‑of‑three upper bounds with Jeffreys‑prior credible intervals, which converts the lack of outcome claims into interpretable uncertainty statements and extends standard practice in sparse‑count inference in applied settings (Agresti, 2019). This directly addresses the measurement gap between adoption signals and outcomes emphasized in Section 2.5 and answered in RQ3. Third, it frames language and library choice as a field‑level signal and documents Python’s ubiquity and the sparse incidence of alternatives, thereby refining accounts of disclosure as signaling rather than simple description and complementing broader firm‑level studies of AI adoption that do not track tools at this level of detail (e.g., Acemoglu et al., 2022; McElheran et al., 2024). The resulting subsector‑resolved catalog and replication‑ready pipeline provide a template that can be extended to other exchanges, sectors, and time periods. 5.6 Limitations and directions for future research Several limitations qualify these interpretations and suggest avenues for future work. The design is NYSE‑only and purposive, which likely over‑represents large, complex, and globally active institutions and may not capture disclosure norms in smaller or privately held financial firms. All measures are constructed from public sources, so the analysis observes only visible communication and cannot cover internal models, vendor arrangements, or infrastructures that remain deliberately undisclosed. Institutions that choose not to reveal AI usage for competitive or regulatory reasons may be coded as having no disclosure even when internal adoption is advanced. Sentence‑level dictionary matching with contextual filters reduces but does not eliminate misclassification risk, especially when institutions use idiosyncratic brand names for internal platforms or when key information appears in images or code appendices that are not readily parsed. Future work can address these limitations by both broadening the frame and deepening the link between disclosure and outcomes. Extending the corpus to other exchanges and jurisdictions would reveal whether similar disclosure equilibria and subsector patterns arise under different regulatory regimes and institutional histories. Linking disclosure‑based measures of AI and Python capability to independent indicators of performance such as stress‑test outcomes, realized loss distributions, or efficiency metrics would allow a more precise assessment of whether explicit disclosure is informative about risk and return in ways that matter for prudential policy and asset allocation. Qualitative research, including interviews with model‑risk managers, technology leaders, and disclosure officers, could illuminate how internal debates about AI and Python translate into the observed mix of explicit naming, generic language, and silence. Such work would connect the quantitative disclosure maps presented here to the organizational processes and institutional pressures that produce them and would further integrate technology, governance, and disclosure perspectives in empirical finance. References Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean, J., … Zheng, X. (2016). TensorFlow: A system for large-scale machine learning. In 12th USENIX Symposium on Operating Systems Design and Implementation (OSDI 2016) (pp. 265–283). USENIX Association. Acemoglu, D., Autor, D., Hazell, J., & Restrepo, P. (2022). Artificial intelligence and jobs: Evidence from online vacancies. Journal of Labor Economics, 40 (S1), S293–S340. https://doi.org/10.1086/718327 Agresti, A. (2019). An introduction to categorical data analysis (3rd ed.). Wiley. Ahmed, S., Alshater, M. M., El Ammari, A., & Hammami, H. (2022). Artificial intelligence and machine learning in finance: A bibliometric review. Research in International Business and Finance, 61 , 101646. https://doi.org/10.1016/j.ribaf.2022.101646 Ametrano, F. M., & Ballabio, L. (2013). QuantLib: A free/open-source library for quantitative finance. Quantitative Finance, 13 (4), 597–606. https://doi.org/10.1080/14697688.2012.744553 Bahoo, S., Cucculelli, M., Goga, X., & Mondolo, J. (2024). Artificial intelligence in finance: A comprehensive review through bibliometric and content analysis. SN Business & Economics, 4 , 23. https://doi.org/10.1007/s43546-023-00618-x Bank of England, & Financial Conduct Authority. (2024). Artificial intelligence in UK financial services—2024 . Bank of England. https://www.bankofengland.co.uk/report/2024/artificial-intelligence-in-uk-financial-services-2024 Benjamini, Y., & Hochberg, Y. (1995). Controlling the false discovery rate: A practical and powerful approach to multiple testing. Journal of the Royal Statistical Society: Series B (Methodological), 57 (1), 289–300. https://doi.org/10.1111/j.2517-6161.1995.tb02031.x Bishop, C. M. (2006). Pattern recognition and machine learning . Springer. Board of Governors of the Federal Reserve System. (2011). Supervisory guidance on model risk management (SR 11‑7). Board of Governors of the Federal Reserve System. Brynjolfsson, E., & McAfee, A. (2017). Machine, platform, crowd: Harnessing our digital future . W. W. Norton. Chen, T., & Guestrin, C. (2016). XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 785–794). ACM. https://doi.org/10.1145/2939672.2939785 Coffee, J. C., Jr. (2020). Gatekeepers: The professions and corporate governance (2nd ed.). Oxford University Press. Corporate Finance Institute. (2025). How is Python used in finance? Applications and examples . Corporate Finance Institute. Crisanto, J. C., Leuterio, C. B., Prenio, J., & Yong, J. (2024). Regulating AI in the financial sector: Recent developments and main challenges (FSI Insights on Policy Implementation No. 63). Bank for International Settlements. Denzin, N. K. (1978). The research act: A theoretical introduction to sociological methods (2nd ed.). McGraw-Hill. DiMaggio, P. J., & Powell, W. W. (1983). The iron cage revisited: Institutional isomorphism and collective rationality in organizational fields. American Sociological Review, 48 (2), 147–160. https://doi.org/10.2307/2095101 FINRA. (2020). Artificial intelligence (AI) in the securities industry . Financial Industry Regulatory Authority. Fischer, T., & Krauss, C. (2018). Deep learning with long short-term memory networks for financial market predictions. European Journal of Operational Research, 270 (2), 654–669. https://doi.org/10.1016/j.ejor.2017.11.054 Gelman, A., Carlin, J. B., Stern, H. S., Dunson, D. B., Vehtari, A., & Rubin, D. B. (2014). Bayesian data analysis (3rd ed.). Chapman & Hall/CRC. https://doi.org/10.1201/b16018 Gentzkow, M., Kelly, B. T., & Taddy, M. (2019). Text as data. Journal of Economic Literature, 57 (3), 535–574. https://doi.org/10.1257/jel.20181020 Goodell, J. W., Kumar, S., Lim, W. M., & Pattnaik, D. (2021). Artificial intelligence and machine learning in finance: Identifying foundations, themes, and research clusters from bibliometric analysis. Journal of Behavioral and Experimental Finance, 32 , 100577. https://doi.org/10.1016/j.jbef.2021.100577 Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learning . MIT Press. Grimmer, J., & Stewart, B. M. (2013). Text as data: The promise and pitfalls of automatic content analysis methods for political texts. Political Analysis, 21 (3), 267–297. https://doi.org/10.1093/pan/mps028 Haldane, A. G., & May, R. M. (2011). Systemic risk in banking ecosystems. Nature, 469 (7330), 351–355. https://doi.org/10.1038/nature09659 Hanley, J. A., & Lippman-Hand, A. (1983). If nothing goes wrong, is everything all right? Interpreting zero numerators. JAMA, 249 (13), 1743–1745. https://doi.org/10.1001/jama.1983.03330370053031 Hilpisch, Y. (2018). Python for finance: Mastering data-driven finance (2nd ed.). O’Reilly Media. Ion Group. (2024). How global markets are adapting to AI and digitalization . ION. Jick, T. D. (1979). Mixing qualitative and quantitative methods: Triangulation in action. Administrative Science Quarterly, 24 (4), 602–611. https://doi.org/10.2307/2392366 Kelliher, C. (2022). Quantitative finance with Python: A practical guide to investment management, trading, and financial engineering . Chapman & Hall/CRC. Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., … Liu, T.-Y. (2017). LightGBM: A highly efficient gradient boosting decision tree. In Advances in Neural Information Processing Systems (Vol. 30, pp. 3146–3154). Curran Associates, Inc. Loughran, T., & McDonald, B. (2016). Textual analysis in accounting and finance: A survey. Journal of Accounting Research, 54 (4), 1187–1230. https://doi.org/10.1111/1475-679X.12123 McKinney, W. (2018). Python for data analysis: Data wrangling with pandas, NumPy, and IPython (2nd ed.). O’Reilly Media. Meyer, J. W., & Rowan, B. (1977). Institutionalized organizations: Formal structure as myth and ceremony. American Journal of Sociology, 83 (2), 340–363. https://doi.org/10.1086/226550 Nazareth, N., & Reddy, Y. V. R. (2023). Financial applications of machine learning: A literature review. Expert Systems with Applications, 219 , 119640. https://doi.org/10.1016/j.eswa.2023.119640 NVIDIA. (2024). State of AI in financial services: 2024 trends . NVIDIA. NVIDIA. (2025). State of AI in financial services: 2025 trends . NVIDIA. NYSE Arca. (2023). Safeguarding the market: NYSE Arca’s surveillance system . NYSE Arca. Orlikowski, W. J., & Iacono, C. S. (2001). Research commentary: Desperately seeking the “IT” in IT research—A call to theorizing the IT artifact. Information Systems Research, 12 (2), 121–134. https://doi.org/10.1287/isre.12.2.121 Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., … Chintala, S. (2019). PyTorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems (Vol. 32, pp. 8024–8035). Curran Associates, Inc. Patton, M. Q. (2015). Qualitative research & evaluation methods (4th ed.). Sage. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., … Duchesnay, É. (2011). Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12 , 2825–2830. Spence, M. (1973). Job market signaling. Quarterly Journal of Economics, 87 (3), 355–374. https://doi.org/10.2307/1882010 Thornton, P. H., & Ocasio, W. (2008). Institutional logics. In R. Greenwood, C. Oliver, K. Sahlin, & R. Suddaby (Eds.), The SAGE handbook of organizational institutionalism (pp. 99–129). Sage. https://doi.org/10.4135/9781849200387.n4 Tonello, M. (2025). AI risk disclosures in the S&P 500: Reputation, cybersecurity, and regulation. Harvard Law School Forum on Corporate Governance . U.S. Department of the Treasury. (2024). Uses, opportunities, and risks of artificial intelligence in financial services . U.S. Department of the Treasury. Uberti-Bona Marin, L. G., Rijsbosch, B., Spanakis, G., & Kollnig, K. (2025). Are companies taking AI risks seriously? A systematic analysis of companies’ AI risk disclosures in SEC 10‑K forms. In Proceedings of the SoGood Workshop at ECML PKDD 2025 . (Preprint available at arXiv:2508.19313). U.S. Government Accountability Office. (2025). Artificial intelligence: Use and oversight in financial services (GAO‑25‑107197). U.S. Government Accountability Office. U.S. Securities and Exchange Commission. (2023). Conflicts of interest associated with the use of predictive data analytics by broker-dealers and investment advisers (Release No. 34‑97990; IA‑6353; File No. S7‑12‑23). U.S. Securities and Exchange Commission. VanderPlas, J. (2022). Python data science handbook: Essential tools for working with data (2nd ed.). O’Reilly Media. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., … Polosukhin, I. (2017). Attention is all you need. In Advances in Neural Information Processing Systems (Vol. 30). Curran Associates, Inc. (Original work available at https://arxiv.org/abs/1706.03762) Weil, Gotshal & Manges LLP. (2023). SEC disclosures of artificial intelligence technologies: Survey results and examples from 2023 company filings . Weil, Gotshal & Manges LLP. Yin, R. K. (2018). Case study research and applications: Design and methods (6th ed.). Sage. Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8884680","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":591570806,"identity":"68459cf1-cc41-4752-a10d-2bac12b648a6","order_by":0,"name":"Veliota Drakopoulou","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA3UlEQVRIiWNgGAWjYFCCBAbGBhDNDsQfQFzitTAzMDDOIFkLMw8xWgyO5z78OKPijpzBYeZj0rZtdnn87A2MHz7m4NFy5rmx5IYzz4wNDrOlSee2JRdL9hxglpy5DbcWyRlpDJIP2w4nbjjMYwbUwpy44UYCGzMvfi3MP+FaLNvqCWvhl0hjk9wI08IIYhDUwvOMzXIG0C+Sh9mSLXvOHU+c2XOwGa9f2NjTmG/2AEOM73jzwRs/yqoT+9mbD374iEcLFByAUIxsYLKBoHqEFoY/xCgeBaNgFIyCkQYAmydUwqnLZn8AAAAASUVORK5CYII=","orcid":"https://orcid.org/0000-0002-1670-8033","institution":"Affiliation (1) Embry Riddle Aeronautical University; Affiliation (2) Higher Colleges of Technology","correspondingAuthor":true,"prefix":"","firstName":"Veliota","middleName":"","lastName":"Drakopoulou","suffix":""}],"badges":[],"createdAt":"2026-02-15 08:26:41","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-8884680/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8884680/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":102901136,"identity":"77413851-6616-43f0-b0f1-8ac646c9c776","added_by":"auto","created_at":"2026-02-18 08:10:34","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":143482,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eFigure 4a. Disclosure status counts (explicit Python / indirect AI / none).\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eFigure 4b. Disclosure status by subsector (row shares).\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eFigure 4c. Named language mentions by subsector (Python vs. R/Java/SAS/Scala/MATLAB).\u003c/em\u003e\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-8884680/v1/7d21efda9d32b0c9cebda9b6.png"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003ePublic Signals of Python‑Enabled AI in Finance: Disclosure Patterns and Outcome Claims in NYSE Institutions\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eArtificial intelligence (AI) is diffusing rapidly across financial services, where data-intensive business models, complex risk management tasks, and high-frequency markets create strong incentives for algorithmic decision support and automation. At the same time, Python has become the dominant language for analytics, modeling, and production pipelines in quantitative finance, trading infrastructure, and enterprise data platforms. Together, these trends have fostered a perception that listed financial institutions now operate with substantial embedded AI and Python capability. Yet public disclosure of such capability is sparse, heterogeneous, and largely unstandardized, which complicates empirical assessment of technology adoption and its economic implications.\u003c/p\u003e \u003cp\u003eExisting disclosure regimes were not designed with AI or software capability in mind. Regulatory filings contain extensive financial and risk information but usually treat data and analytics infrastructure as background inputs rather than core strategic assets. Voluntary sustainability and governance reports often reference \u0026ldquo;digital transformation\u0026rdquo; or \u0026ldquo;innovation\u0026rdquo; at a high level, typically in marketing-oriented language. Sectoral press, employer portals, and job advertisements contain richer technical details, but coverage varies widely across firms and is rarely synthesized into a comparable view of capabilities at the level of programming languages, libraries, or analytical paradigms. As a result, the public record understates or distorts the technological frontier in financial institutions and obscures the boundary between signaling sophistication and documenting realized performance gains.\u003c/p\u003e \u003cp\u003eThe research literature mirrors this fragmentation. Studies of information technology and productivity tend to rely on investment data, surveys, or proprietary vendor usage metrics, which can offer depth but often lack transparency or replicability. Work on AI adoption in finance has focused on case studies, specific application domains such as credit scoring or fraud detection, or infrastructure-level metrics such as cloud usage. Separate strands of research track the Python ecosystem through open-source repositories and developer activity yet rarely connect such evidence to firm-level disclosure by listed financial institutions. Consequently, empirical knowledge about the prevalence, structure, and \u003cem\u003eclaimed\u003c/em\u003e outcomes of Python and AI capability at the firm level remains limited, despite strong policy and supervisory interest in model risk, operational resilience, and digital transformation.\u003c/p\u003e \u003cp\u003e \u003cb\u003eResearch gap.\u003c/b\u003e Taken together, these observations point to a specific gap. First, we lack a systematic, firm-level map of tool-specific AI and Python adoption based solely on public disclosures, for a clearly defined financial universe. Second, we have little cross-subsector evidence on how Python-framed AI disclosures and associated tool stacks vary across distinct business models and institutional logics. Third, there is almost no quantitative evidence on whether public AI and Python statements are accompanied by explicit, quantified outcome claims (for example, risk reduction, accuracy gains, or efficiency improvements), or on how rare such claims truly are once uncertainty is accounted for. Existing studies are either abstract from programming languages and libraries or rely on data that cannot easily be replicated by other researchers.\u003c/p\u003e \u003cp\u003eThis paper addresses these gaps by constructing a systematic map of Python and AI disclosure for a strict frame of 180 New York Stock Exchange (NYSE) listed financial institutions. Public communication is treated as a noisy but informative signal of internal capability. A triangulated corpus combines regulatory filings, employer portals and job advertisements, corporate communications, and sectoral press over a collection window that extends through January 2026. Within this corpus, each firm is assigned to one of three disclosure states explicit Python, indirect AI, \u003cb\u003eor\u003c/b\u003e none and named libraries are detected using sentence-level dictionary matching subject to contextual filters. Mentions are then mapped to seven analytical dimensions: natural language processing, machine learning, deep learning, reinforcement learning, probabilistic modeling, optimization, and visualization. Inverse-visibility weights mitigate the influence of firms that generate unusually rich or sparse public traces, and inferential tools include cross-tabulations, Pearson chi-squared tests with multiple-testing control, and compact robustness models with subsector fixed effects. The study asks three research questions:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eRQ1. Prevalence\u003c/b\u003e: How prevalent are explicit Python disclosures and indirect AI references among NYSE-listed financial institutions, once differences in public visibility and evidence channels are considered?\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eRQ2. Structure and heterogeneity\u003c/b\u003e: How do Python-framed AI disclosures, associated library stacks, and analytical dimensions vary across financial subsectors characterized by different business models and institutional logics?\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eRQ3. Outcome claims\u003c/b\u003e: To what extent do public disclosures contain quantified outcome claims about AI and Python systems, and what upper bounds can be placed on their true frequency in sectors and subsectors where no such claims are observed?\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eA central conceptual distinction in answering these questions concerns adoption signals versus explicit claims about outcomes. To capture this distinction, the study introduces a sentence-level Outcome Claims Index (OCI) that flags whether a given item of evidence advances quantified performance assertions rather than simple capability statements or vague claims of improvement. The OCI supports both frequentist and Bayesian uncertainty summaries and provides a structured way to characterize the near absence of quantified outcome disclosure, even among firms that prominently advertise Python and AI usage. This distinction matters for regulators, investors, and counterparties that must understand not only whether institutions state that AI is in use, but also whether those institutions offer measurable evidence that AI materially alters risk, efficiency, or profitability.\u003c/p\u003e \u003cp\u003eThe empirical results indicate three broad patterns. First, explicit Python disclosure is widespread in the NYSE financial frame, although a non-trivial minority of institutions exhibits no observable Python or AI signal in the assembled corpus. Second, disclosure patterns vary strongly across subsectors, and a common backbone of general-purpose libraries coexists with more specialized toolkits for text-heavy processes, tabular risk modeling, payments networks, and market microstructure. Third, quantified outcome claims appear extremely rarely, and the associated bounds are consistent with an Outcome Claims Index that is effectively zero across subsectors. Explicit Python disclosure instead aligns with performance-adjacent proxies such as technology hiring intensity, technology and research and development expenditure, and overall visibility, which suggests that public communication currently reflects capability accumulation more than demonstrable performance change.\u003c/p\u003e \u003cp\u003eFurther contribution concerns transparency and replicability. A blinded replication archive accompanies the study and includes a harmonized roster of NYSE financial institutions, sentence-level evidence with Outcome Claims Index flags, and structured \u003cspan refid=\"Sec22\" class=\"InternalRef\"\u003eappendix\u003c/span\u003e workbooks. The replication worksheet documents the sampling frame, coding rules, visibility weights, and modeling choices in a form that enables independent reproduction of the core tables and figures, as well as straightforward extension to additional exchanges, sectors, or time periods. This design aligns the measurement of Python and AI disclosure with contemporary norms of open data and computational reproducibility in empirical finance and applied machine learning.\u003c/p\u003e \u003cp\u003eThe remainder of the paper is organized as follows. Section \u003cspan refid=\"Sec2\" class=\"InternalRef\"\u003e2\u003c/span\u003e situates the study within the literature on information technology, AI adoption, disclosure, and financial-sector digitalization. Section \u003cspan refid=\"Sec12\" class=\"InternalRef\"\u003e3\u003c/span\u003e describes the NYSE financial frame, data sources, coding procedures, and construction of the Outcome Claims Index and visibility weights. Section 4 presents the main descriptive and inferential results on disclosure states, library ecosystems, analytical dimensions, and sub-sectoral patterns. Section 5 reports robustness exercises based on logit and linear probability models and discusses the interpretation of potential selection and measurement biases. Section 6 concludes with implications for research, supervision, and disclosure practice, and outlines directions for future extensions of the replication worksheet and corpus\u003c/p\u003e"},{"header":"2. Literature Review","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Python as infrastructure for financial AI\u003c/h2\u003e \u003cp\u003eAcross contemporary finance, Python increasingly functions as the backbone that links data engineering, model development, and production deployment. Its appeal rests on a mature, interoperable open-source ecosystem: pandas and NumPy for data management and numerical linear algebra (McKinney, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2018\u003c/span\u003e), scikit-learn for baseline machine learning and model selection (Pedregosa et al., \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2011\u003c/span\u003e), gradient-boosting frameworks such as XGBoost and LightGBM for high-accuracy tabular prediction at scale (Chen \u0026amp; Guestrin, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Ke et al., \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2017\u003c/span\u003e), and deep-learning stacks such as TensorFlow and PyTorch for representation learning and sequence models (Abadi et al., \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Paszke et al., \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). Domain-specific toolkits extend this stack into core financial tasks: QuantLib supports fixed-income analytics and derivative pricing (Ametrano \u0026amp; Ballabio, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2013\u003c/span\u003e), while optimization and portfolio packages implement mean\u0026ndash;variance and constraint-aware allocation routines. Taken together, these components create a vertical pathway from raw data to deployment and combined with abundant developer mindshare and open-source governance, have made Python the de facto lingua franca for AI-enabled analytics in finance. Existing work, however, typically documents Python\u0026rsquo;s capabilities or use cases rather than providing a field-level, firm-resolved map of where Python is explicitly acknowledged in the public record. This motivates a language-specific, disclosure-based perspective in the present study.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Application domains in finance and the role of Python\u003c/h2\u003e \u003cp\u003ePython libraries operationalize a wide spectrum of AI tasks that now feature prominently in financial research and practice. Credit and fraud analytics rely on supervised learning and boosting to balance accuracy and false-positive control in imbalanced settings (Chen \u0026amp; Guestrin, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Ke et al., \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). Time-series forecasting and volatility modeling employ LSTMs and related deep sequence models (Fischer \u0026amp; Krauss, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2018\u003c/span\u003e), while text mining of filings, earnings calls, and news uses NLP pipelines built on spaCy and transformer-based architectures (Loughran \u0026amp; McDonald, \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). Market surveillance and systemic-risk diagnostics draw on network representations of propagation channels and correlated exposures (Haldane \u0026amp; May, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2011\u003c/span\u003e), often implemented with Python-based graph toolkits. This application-level literature shows that Python is not merely an implementation detail but a platform technology that standardizes data structures, model interfaces, and deployment idioms across diverse financial tasks. What it does not provide is a systematic, firm-level view of how such Python-enabled capabilities appear in public communications by listed institutions or how disclosure practices vary across subsectors. The present analysis takes those omissions as a starting point.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 From bibliometrics to disclosures: measuring adoption in organizational settings\u003c/h2\u003e \u003cp\u003eA separate strand of work concerns how technology adoption is measured inside organizations. Traditional approaches such as bibliometrics, patent counts, case studies, or proprietary vendor telemetry are informative about innovation activity but seldom resolve \u003cem\u003ewho uses which concrete tools now\u003c/em\u003e within specific institutional niches such as listed financial institutions. They typically operate at the level of broad \u0026ldquo;AI\u0026rdquo; or \u0026ldquo;analytics\u0026rdquo; categories and do not distinguish between programming languages, libraries, or analytical paradigms at the firm level. Organizational research has long advocated triangulation, deliberately combining evidence sources with distinct logics (e.g., regulatory filings, corporate materials, sectoral press, and labor-market signals) to strengthen construct validity and reduce single-source bias (Denzin, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e1978\u003c/span\u003e; Jick, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e1979\u003c/span\u003e; Patton, \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2015\u003c/span\u003e; Yin, \u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Yet despite these recommendations, most empirical work on AI adoption in finance either relies on one dominant channel (such as surveys or filings) or blends heterogeneous samples across markets, making it difficult to know which population is being described.\u003c/p\u003e \u003cp\u003eThis leaves three related measurement gaps. First, there is no systematic, tool-specific map of AI and Python adoption based solely on public disclosures for a clearly defined population of listed financial institutions. Second, existing studies rarely tackle visibility bias directly: highly vocal institutions dominate public channels, inflating apparent adoption rates unless disclosure intensity is explicitly modelled. Third, the non-disclosure paradox firms adopting technologies internally yet avoiding naming stacks or quantifying impacts in public is usually treated as an anecdotal caveat rather than as measurable regularity. The methodological stance used here responds to these gaps by combining a strictly defined NYSE finance population with multi-channel evidence and explicit coding rules for disclosure status. In this design, triangulation is not just a robustness check but the core of the measurement strategy: it treats visibility bias as a parameter to be adjusted for and non-disclosure itself as informative variation across firms and subsectors. This positioning directly motivates RQ1 on the prevalence of explicit Python and indirect AI disclosure and RQ2 on the structure and subsector heterogeneity of tool stacks and disclosure granularity.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Signaling, institutional logic, and disclosure granularity\u003c/h2\u003e \u003cp\u003eWhy do some firms explicitly say \u0026ldquo;Python\u0026rdquo;, whereas others only reference \u0026ldquo;AI\u0026rdquo; or \u0026ldquo;automation\u0026rdquo;? Signaling theory predicts that organizations disclose just enough technical detail to convey competence to target audiences while minimizing competitive and regulatory exposure (Spence, \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e1973\u003c/span\u003e). In finance, institutional logics shape which details are perceived as legitimate signals: market-facing subsectors (e.g., trading and market infrastructure) tend to foreground technical capability, while credit vehicles often emphasize prudential themes and operational reliability (Coffee, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Within this framework, language-level statements (\u0026ldquo;Python\u0026rdquo;) function as relatively low-risk, high-recognition signals, easily understood by investors, staff, and regulators. Library-level naming (specific packages) is a higher-granularity signal that may be strategically withheld because it reveals more about internal architectures and vendor relationships. Case studies and qualitative analyses consistently find that language mentions are common while named libraries are rare, suggesting that granularity is an analytic variable reflecting audience, risk posture, and competitive context, not simply the degree of internal adoption. Existing empirical work, however, typically stops at conceptual arguments; it does not quantify how granularity and explicitness vary across subsectors in a large, clearly defined population. The present study addresses this by treating disclosure granularity itself as an outcome of interest and by embedding it in the cross-subsector diagnostics that feed into RQ2.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.5 From adoption to outcomes: the measurement gap\u003c/h2\u003e \u003cp\u003eA persistent critique of the AI adoption literature is that it maps use but rarely measures impact. Studies typically report whether AI or advanced analytics are present, which functions they support, and sometimes how extensively they are deployed, but they provide far less evidence on quantified outcomes such as return on investment, error reductions, risk or loss improvements, or latency and efficiency gains. This problem is particularly acute in regulated sectors such as finance, where public claims about performance can invite supervisory scrutiny, litigation risk, or competitive disadvantage. As a result, firms frequently name AI, machine learning, or Python yet refrain from publishing specific metrics that would document the economic or risk consequences of those systems.\u003c/p\u003e \u003cp\u003eTwo methodological responses have emerged. One advocates direct outcome designs, for example, event studies around dated AI announcements, confidential surveys, or access to internal telemetry to estimate economic effects more cleanly (cf. productivity debates in digital transformation; Brynjolfsson \u0026amp; McAfee, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). The other emphasizes performance-adjacent indicators observable in public data, such as disclosure intensity, channel breadth, or technology hiring, as proxies for capability maturity while carefully distinguishing these from audited key performance indicators. Both approaches share a limitation: they seldom treat the absence of quantified outcome claims in public disclosures as data, and they rarely provide formal uncertainty statements about how rare such claims might be within specific subsectors. Consequently, there is a persistent measurement gap between adoption signals and documented outcomes. We lack both a systematic accounting of how often financial institutions publicly attach numbers to their AI and Python systems and a framework for turning subsector-level \u0026ldquo;zero events\u0026rdquo; into interpretable bounds on the underlying rate of outcome disclosure.\u003c/p\u003e \u003cp\u003eThe present study addresses this gap by introducing a sentence-level Outcome Claims Index (OCI) that explicitly distinguishes adoption statements from quantified performance claims and by pairing it with both frequentist and Bayesian interval estimates for sparse-event strata. In doing so, the analysis reframes the near absence of outcome claims as a measurable empirical pattern rather than as a mere limitation of the data. This perspective underpins RQ3, which asks to what extent public disclosures contain quantified outcome claims about AI and Python systems and what upper bounds can be placed on their true frequency when none are observed.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e2.6 Statistical treatment of cross-subsector patterns\u003c/h2\u003e \u003cp\u003eQualitative studies of digitalization in finance often note that disclosure and adoption are concentrated in certain subsectors, but they rarely subject these impressions to formal statistical tests. Where cross-sector comparisons are made, researchers typically rely on contingency tables, χ\u0026sup2; tests of independence, and associated effect sizes such as Cram\u0026eacute;r\u0026rsquo;s V to assess association strength, sometimes supplemented by post-hoc contrasts and multiple-testing adjustments. Visibility bias is usually acknowledged but not quantitatively addressed, and multivariate models that control both subsector and simple proxies for visibility remain uncommon in the AI adoption literature. This methodological landscape suggests the value of combining visibility-aware cross-tabulations with parsimonious regression diagnostics. In the present study, these tools are used not to claim causal effects but to translate intuitions from signaling and institutional logics into falsifiable, replicable statistics for the finance context. By doing so, the analysis helps move from impressionistic claims about \u0026ldquo;innovative subsectors\u0026rdquo; towards documented patterns of disclosure and explicitness that can be compared across markets and over time.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e2.7 Python versus other languages: a language-specific lens\u003c/h2\u003e \u003cp\u003eTool-agnostic surveys tend to pool languages and platforms, but a language-specific lens can reveal meaningful asymmetries. Python\u0026rsquo;s dominance in data-science education, scientific computing, and MLOps integration makes it more observable in public artefacts than R, MATLAB, or SAS, particularly where cloud-native deployment and model orchestration are salient (McKinney, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; Pedregosa et al., \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2011\u003c/span\u003e). At the same time, some subsectors retain specialized stacks such as SAS in actuarial workflows, for example, or Java/Scala in legacy streaming and trading infrastructure which underscores why stack granularity should be treated as a disclosure variable rather than a direct proxy for internal prevalence. Existing work documents these ecosystems but does not quantify, at the firm level, how often different languages are named in public communications or how language mentions co-vary with sector, business model, and disclosure granularity. A language-specific approach, coupled with explicit caveats about what disclosure can and cannot reveal about internal systems, therefore improves both transparency and replicability in adoption mapping. It also justifies the study\u0026rsquo;s focus on Python-framed disclosure while benchmarking against alternative languages.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e2.8 Governance, ethics, and stability considerations\u003c/h2\u003e \u003cp\u003eEven when Python-enabled models deliver internal value, governance considerations such as model opacity and validation, fairness and bias, third-party concentration, and cyber-security exposure shape what firms choose to disclose. The governance literature emphasizes documentation of model purpose, data lineage, validation and challenge, post-deployment monitoring (drift and stability), and human-in-the-loop controls as necessary conditions for trustworthy AI in finance (Coffee, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Systemic-risk work warns that common data and model dependencies can propagate correlated errors and amplify shocks (Haldane \u0026amp; May, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2011\u003c/span\u003e). These concerns help explain why public statements often remain at the language level and why quantitative outcome claims are scarce: transparency must be balanced against security, competition, and prudential discipline. In this environment, naming Python and AI can serve to signal modern capability without revealing detailed model performance or stack composition. This literature therefore not only motivates the OCI\u0026rsquo;s focus on outcome claims but also frames the interpretation of \u0026ldquo;zeros\u0026rdquo; in outcome disclosure as a strategic equilibrium rather than as mere data sparsity.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e2.9 Synthesis and positioning\u003c/h2\u003e \u003cp\u003eLiterature collectively points to three unresolved issues. First, while application-level work documents numerous AI use cases in finance and Python has emerged as a de facto platform for these systems, we still lack a systematic, firm-level map of tool-specific AI and Python adoption based solely on public disclosures for a clearly delimited financial universe. Second, prior studies have not quantified how disclosure granularity and stack composition vary across subsectors governed by different institutional logics, even though case evidence suggests substantial heterogeneity in how explicitly firms talk about their tools. Third, existing research largely stops at adoption signals; it does not measure how often public AI and Python statements are accompanied by explicit, quantified outcome claims, nor does it provide uncertainty bounds when such claims are absent.\u003c/p\u003e \u003cp\u003eThe present study responds to these gaps by (i) focusing on a strict NYSE finance universe to ensure population clarity; (ii) implementing a triangulated, multi-channel measurement design with explicit coding rules for disclosure status and stack granularity; and (iii) introducing a sentence-level Outcome Claims Index (OCI) that operationalizes outcome claims and treats zeros as informative through both frequentist and Bayesian interval estimates. This design directly supports the three research questions set out in the Introduction: the prevalence of Python-framed and indirect AI disclosure (RQ1), the structure and subsector heterogeneity of AI/Python tool stacks and disclosure granularity (RQ2), and the frequency and interpretation of quantified outcome claim in public materials (RQ3). In doing so, it complements tool-agnostic surveys and proprietary usage studies by offering a replication-ready, disclosure-based map of Python-enabled AI in finance and by establishing performance-adjacent indicators that future work can link to direct outcome measures.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Data and Methods","content":"\u003cp\u003e\u003cstrong\u003e3.1 Population and Sampling Frame\u003c/strong\u003e\u003cbr\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe empirical universe consists of one hundred eighty (180) New York Stock Exchange (NYSE) listed financial institutions drawn from a single-exchange frame to ensure homogeneity in disclosure obligations and to avoid cross-listing ambiguity. Each issuer is assigned to a harmonized subsector taxonomy reflecting dominant business models and regulatory logics: Diversified Banks, Regional Banks, Custody \u0026amp; Asset Servicing, Investment Banks, Asset Managers, Asset Managers (Alternatives), Insurance, Payments \u0026amp; Consumer Finance, Market Infrastructure \u0026amp; Data, BDC/Finance, Mortgage REIT/Finance, Fixed Income e-Trading, and Investment, Tech \u0026amp; Risk. \u0026nbsp;The cross section is fixed at N = 180 following ticker deduplication and consolidation of merger and delisting events, yielding a one-row-per-issuer design suitable for transparent replication. Representative firms, counts, and inclusion rationales are reported in Appendix Table 3.1.\u0026nbsp;This strict and fully enumerated NYSE frame provides the population clarity required for the prevalence and heterogeneity analyses in RQ1 and RQ2.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.2 Evidence Sources and Acquisition\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eData collection follows a triangulated strategy that integrates four public evidence channels with complementary institutional logics: (i) legally accountable regulatory filings (e.g., 10‑K/20‑F, risk and MD\u0026amp;A discussion); (ii) employer portals and job postings; (iii) corporate communications and sectoral press; and (iv) product and engineering documentation. Triangulation increases construct validity by reducing single‑source bias and by exploiting the different incentives that shape each channel (Denzin, 1978; Jick, 1979; Patton, 2015; Yin, 2018). For each evidence item, the dataset records the firm identifier and ticker, subsector, source channel, URL or document ID, evidentiary sentence(s), and retrieval date. Where available, filing dates and press‑release timestamps are retained to facilitate future event‑study designs. The collection window closes in January 2026, producing a time‑stamped, replication‑ready cross‑section of public communication.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.3 Coding Scheme and Disclosure Granularity\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDisclosure is coded at the firm level using mutually exclusive status labels derived from convergent evidence:\u003c/p\u003e\n\u003cul type=\"disc\"\u003e\n \u003cli\u003e\u003cstrong\u003eexplicit Python\u003c/strong\u003e when the public text directly names \u0026ldquo;Python\u0026rdquo; and/or specific Python libraries.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eindirect AI\u003c/strong\u003e when the text discloses AI/ML/automation without naming the language.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003enone\u003c/strong\u003e when no AI/ML reference appears in the collection window.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eCoding follows a hierarchical rule: if any channel contains an explicit Python reference, the firm is coded as explicit Python; if no explicit Python references are found but AI/ML/automation is mentioned, the firm is coded as indirect AI; only firms with no AI/ML reference in any channel are coded as none. Because technical specificity varies strategically across channels, the corpus also records library‑level mentions when they occur (e.g., pandas/NumPy for data handling, scikit‑learn for classical ML, XGBoost/LightGBM for boosting, TensorFlow/PyTorch for deep learning), with string normalization and canonical family mapping to ensure comparability (Abadi et al., 2016; Chen \u0026amp; Guestrin, 2016; Ke et al., 2017; McKinney, 2018; Paszke et al., 2019; Pedregosa et al., 2011). Consistent with signaling theory and institutional logics, language‑level statements are interpreted as capability signals, whereas library‑level naming is treated as disclosure granularity rather than a census of internal stacks (Spence, 1973; Coffee, 2020).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.4 Python Library Detection and Subsector Stacks\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eSubsector‑level AI and Python tool adoption is summarized in a five‑column exhibit that reports \u0026ldquo;Subsector\u0026rdquo;, \u0026ldquo;Representative Companies\u0026rdquo;, \u0026ldquo;Analytical Dimensions\u0026rdquo;, \u0026ldquo;Primary AI/Python Use Cases\u0026rdquo;, and \u0026ldquo;Representative Python/AI Stack\u0026rdquo;. This summary is based solely on in‑sample evidence from the corpus and is reported in Appendix Table 4. Named tools are identified via sentence‑level pattern matching against a predefined dictionary of libraries and platforms (e.g., pandas, NumPy, SciPy, scikit‑learn, XGBoost, LightGBM, TensorFlow, PyTorch, spaCy, transformer libraries, PyMC/Pyro, NetworkX, PyTorch Geometric, CVXPY, and vendor SDKs such as GS Quant and AladdinSDK). The procedure follows standard text‑as‑data and dictionary‑based coding approaches (Grimmer \u0026amp; Stewart, 2013; Gentzkow, Kelly, \u0026amp; Taddy, 2019) and widely used Python ML/NLP toolchains (Pedregosa et al., 2011; Abadi et al., 2016; Paszke et al., 2019; Vaswani et al., 2017).\u003c/p\u003e\n\u003cp\u003eMatches are retained only when library tokens appear near finance‑relevant actions or contexts (e.g., underwriting, pricing, execution, deploy, production, fraud, surveillance, risk). Ambiguous cases are manually reviewed. Validated detections are aggregated to the subsector level and read alongside local context to infer (i) dominant analytical dimensions, (ii) primary AI/Python use cases, and (iii) characteristic stacks, in line with common typologies in the ML/AI literature (Bishop, 2006; Goodfellow, Bengio, \u0026amp; Courville, 2016).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.5 Thematic Dimensions and Theme Library Mapping\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo impose a consistent structure on otherwise heterogeneous AI and Python disclosures, all in‑sample evidence is recoded into seven analytical dimensions: natural language processing (NLP), machine learning (ML), deep learning (DL), reinforcement learning (RL), probabilistic modeling, optimization, and visualization. These dimensions follow standard typologies in the ML literature and provide a mutually exclusive, collectively exhaustive partition of the main modeling and analytics workstreams observed in the corpus (Bishop, 2006; Goodfellow, Bengio, \u0026amp; Courville, 2016). Each dimension is then mapped to canonical Python tool families, operationalizing the link between thematic capability and concrete implementation. Supervised tabular ML is associated with scikit‑learn, XGBoost, and LightGBM; DL with TensorFlow and PyTorch; NLP with spaCy and transformer‑based libraries; probabilistic modeling with PyMC/Pyro and related Bayesian frameworks; and optimization and quantitative finance with tools such as CVXPY, PyPortfolioOpt, and QuantLib. This theme library mapping is applied uniformly across channels and subsectors and serves as the organizing device for\u0026nbsp;\u003cstrong\u003eAppendix Table 4\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eand the cross‑subsector diagnostics in Section 4.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.6 Outcome‑Claims Index (OCI) and Inference for Zeros\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eEach sentence in the corpus is automatically scanned for quantified performance claims (e.g., percentage improvements, model metrics, loss or risk reductions, latency or efficiency gains, or financial KPIs) using standard text‑as‑data techniques (Grimmer \u0026amp; Stewart, 2013; Gentzkow, Kelly, \u0026amp; Taddy, 2019). A binary indicator is assigned at the sentence level and then aggregated to firm‑ and subsector‑level shares to form the Outcome Claims Index (OCI). To address sparse‑event strata (x = 0), uncertainty is summarized using two complementary methods. First, the frequentist \u0026ldquo;rule of three\u0026rdquo; provides a 95% upper bound (\u0026asymp; 3/n) as a conservative maximum rate approximation (Hanley \u0026amp; Lippman‑Hand, 1983). Second, Bayesian credible intervals are computed using Jeffreys prior, Beta (x + 0.5, n \u0026minus; x + 0.5), which yields well‑behaved finite‑sample intervals at the boundary (Agresti, 2019; Gelman et al., 2014). OCI point estimates and both interval sets are reported in Appendix Tables S4.5d\u0026ndash;S4.5e and visualized in Appendix Figure S4.5e; full strata-level computations appear in the replication workbook sheet S4.5.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.7 Language Benchmarking and Statistical Diagnostics\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNamed programming language mentions (Python, R, Java, SAS, Scala, MATLAB) are tallied at firm and subsector levels to contextualize Python\u0026rsquo;s prominence in public disclosures. The resulting cross tabs appear in Appendix Table S4.6 and underpin Figure 4c, with underlying counts provided in the replication workbook. Firm-level disclosure status (explicit/indirect/none) is cross-tabulated by subsector, and inference uses Pearson \u0026chi;\u0026sup2; with Cram\u0026eacute;r\u0026rsquo;s V as an effect size and Benjamini\u0026ndash;Hochberg correction on post hoc standardized residual contrasts. These diagnostics are reported in Appendix Tables S4.2\u0026ndash;S4.4 and interpreted in Section 4, confirming systematic sub sectoral heterogeneity in disclosure modes for the NYSE frame.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.8 Data Hygiene, Reliability, and Limitations\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eQuality control proceeds in three passes: issuer harmonization (standardized names and tickers, subsector taxonomy), dual-source verification at the evidence-row level, and manual spot checks of named library matches and exemplars used in Appendix Table 4. These steps improve validity but do not eliminate visibility bias or channel asymmetries; both are documented and treated as part of the phenomenon under study, the politics of disclosure, rather than as purely random noise, and they are revisited in the limitations section.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.9 Reproducibility and replication materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll empirical results are fully reproducible. A blinded replication archive (submitted as \u003cstrong\u003eAdditional file\u0026nbsp;1\u003c/strong\u003e) accompanies the manuscript and contains: (i) a harmonized roster of the 180 NYSE financial institutions with tickers, standardized names, subsector assignments, and disclosure status; (ii) the sentence‑level evidence corpus with source channel, retrieval dates, and Outcome Claims Index flags; (iii) language‑benchmark counts and visibility‑weighted disclosure shares; (iv) the appendix workbook with all statistical tables and diagnostics; and (v) the scripts/notebooks and configuration file used to generate the derived variables, cross‑tabulations, figures, and robustness checks. The archive is organized so that the complete set of main‑text and appendix outputs can be rebuilt from a single configuration, facilitating independent verification and extension to other exchanges, sectors, or time periods.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.10 Availability of data and materials\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe dataset(s) and code supporting the conclusions of this article are included within the article and its additional file(s). A blinded replication archive (Additional file 1) contains the harmonized firm‑level roster, the sentence‑level Python/AI disclosure corpus with Outcome Claims Index flags, language‑benchmark counts, the appendix workbook with all statistical tables, and the scripts/notebooks required to reproduce the analyses.\u003c/p\u003e"},{"header":"4. Data Analysis and Results","content":"\u003cp\u003eThis section presents the empirical results for the strict NYSE frame of 180 financial institutions and answers the three research questions introduced in Section 1. RQ1 concerns the prevalence of explicit Python and indirect AI disclosure. RQ2 examines cross‑subsector structure and heterogeneity in disclosure modes and Python‑framed stacks. RQ3 addresses the frequency and interpretation of quantified outcome claims, via the Outcome Claims Index (OCI). All results rely on the triangulated corpus, coding scheme, and statistical tools described in Section 3.\u003c/p\u003e\n\u003cp\u003eBecause the evidence is drawn from regulatory filings, employer portals and job postings, corporate materials, and sectoral press, all estimates should be read as disclosure frequencies what firms choose to place in the public record rather than as a census of internal tool use. This is consistent with the triangulated design in the literature review, which treats visible traces as noisy but informative signals of underlying capability (Denzin, 1978; Jick, 1979; Patton, 2015; Yin, 2018). All statistical tables referenced below are reported in the Appendix and in the harmonized workbook \u003cem\u003eTables_S4_all.xlsx\u003c/em\u003e within the replication archive.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.0 Python library adoption and analytical scope\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo contextualize RQ2 on subsector heterogeneity, Appendix Table 4 summarizes AI and Python tool adoption across representative NYSE financial subsectors. The exhibit translates the theme\u0026ndash;library mapping from Section 3.5 into concrete subsector profiles, linking analytical dimensions, primary AI/Python use cases, and representative stacks. Across subsectors, a common Python backbone is visible: pandas and NumPy for data handling and numerical routines; scikit‑learn plus gradient‑boosting frameworks such as XGBoost, LightGBM, and CatBoost for supervised learning; and deep‑learning frameworks such as TensorFlow and PyTorch for more complex representation and sequence models. Probabilistic programming tools such as PyMC or Pyro appear were unstructured data, complex loss distributions, or multi‑scenario analysis feature prominently.\u003c/p\u003e\n\u003cp\u003eAsset managers and alternative asset managers make extensive use of NLP, ML, DL, probabilistic modeling, optimization, and visualization. In these subsectors, Python‑based stacks support factor and risk models, portfolio construction, performance attribution, fund flow and distribution analytics, and ESG or covenant text analysis, often in conjunction with portfolio‑optimization libraries and specialized platforms such as Aladdin SDKs. Business development companies and mortgage‑finance firms apply similar tools to middle‑market or mortgage credit, with an emphasis on underwriting, early‑warning signals, and prepayment or default modeling rather than broad multi‑asset allocation.\u003c/p\u003e\n\u003cp\u003eBanks, insurers, and consumer‑finance firms adapt the same core stack to retail, commercial, and insurance workflows. Diversified and regional banks use NLP and machine learning for contract intelligence, chatbots, KYC and AML monitoring, stress testing, deposit and cash‑flow forecasting, and branch or customer profitability analysis, typically wrapped in dashboard frameworks such as Dash or Streamlit. Insurers deploy boosting and deep‑learning models for pricing, claims triage, lapse and surrender behavior, telematics, and catastrophe risk, supported by NLP for policy, medical, and claims text. Payments‑ and card‑centric firms use graph and sequence models for real‑time fraud detection, merchant risk, chargeback prediction, and personalization, often in streaming or near real‑time Python environments.\u003c/p\u003e\n\u003cp\u003eMarket infrastructure, data providers, investment banks, and electronic trading platforms show the most intensive use of deep learning, reinforcement learning, and optimization. In these subsectors, Python stacks support tick‑data analytics, transaction‑cost analysis, smart order routing, derivatives pricing and XVA, surveillance, and balance‑sheet or funding optimization. They combine the scientific‑Python stack with distributed computing tools, graph and sequence‑modeling frameworks, reinforcement‑learning libraries, and vendor Python APIs such as GS Quant or Aladdin SDKs.\u003c/p\u003e\n\u003cp\u003eAppendix Table 4 indicates that NYSE‑listed financial institutions share a narrow but powerful set of foundational Python tools and differentiate sharply in analytical dimensions and higher‑level stacks. NLP and deep learning are most salient in text‑heavy, surveillance, and fraud‑oriented activities, while probabilistic modeling is most salient in credit, insurance, and mortgage‑risk contexts. These patterns implement, in empirical form, the literature review\u0026rsquo;s argument that Python functions as a field‑level platform technology whose concrete expression varies by subsector (Sections 2.1\u0026ndash;2.2).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.1 Descriptive coverage (RQ1)\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTurning directly to RQ1 (prevalence), the harmonized NYSE sample contains 180 financial institutions. Three disclosure states are coded at the firm level: explicit Python (the firm names Python or a Python library), indirect AI (AI/ML is disclosed but no language is named), and none. Unweighted counts show 138 firms (76.67%) with explicit Python mentions, 4 (2.22%) with indirect AI mentions, and 38 (21.11%) with no disclosure (Figure 4a; Appendix Table S4.1). These shares mirror the headline numbers reported in the abstract and provide a first answer to RQ1: most NYSE financial institutions publicly acknowledge Python in at least one channel.\u003c/p\u003e\n\u003cp\u003eBecause public‑facing evidence is uneven across firms and channels, a visibility‑adjusted estimate is computed by weighting each issuer \u003cimg width=\"6\" height=\"22\" src=\"https://myfiles.space/user_files/58895_8739fc6c57c1c19a/58895_custom_files/img177140155298.png\" alt=\"image\"\u003eby \u003cimg width=\"200\" height=\"24\" src=\"https://myfiles.space/user_files/58895_8739fc6c57c1c19a/58895_custom_files/img1771401552.png\" alt=\"image\"\u003e. The inverse‑visibility‑weighted explicit share equals 0.629, indicating that explicit Python disclosure remains dominant even after down‑weighting highly visible firms (Appendix Table S4.1, \u0026ldquo;shares\u0026rdquo;). This pattern is consistent with the triangulated design\u0026rsquo;s expectation that combining filings, corporate releases, technical press, and job postings will surface a wide footprint of adoption while still reflecting visibility bias (Denzin, 1978; Jick, 1979; Patton, 2015; Yin, 2018).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.2 Subsector differences in disclosure (RQ2)\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eRQ2 asks how disclosure modes and tool stacks vary across subsectors. Firm‑level disclosure status (explicit Python, indirect AI, or none) differs markedly by subsector. A Pearson \u0026chi;\u0026sup2; test on the status \u0026times; subsector table rejects independence (\u0026chi;\u0026sup2; (40, N = 180) = 100.65, p \u0026lt; .001), with Cram\u0026eacute;r\u0026rsquo;s V = 0.529 indicating a moderate\u0026ndash;large association (Appendix Table S4.4, Panel A; Agresti, 2019).\u003c/p\u003e\n\u003cp\u003eThe row‑share heatmap (Figure 4b) makes the pattern visually obvious. Investment Banks, Market Infrastructure \u0026amp; Data, and Payments report explicit library‑ and language‑level disclosures at much higher rates, while Mortgage REIT/Finance and BDC/Finance display low explicitness. Standardized residuals in Appendix Table S4.2 highlight the subsector\u0026ndash;status pairs that contribute most to the \u0026chi;\u0026sup2; statistic, and Benjamini\u0026ndash;Hochberg\u0026ndash;adjusted pairwise contrasts (Appendix Table S4.4, Panel B) show the largest and most robust gaps, for example, between Investment Banks (high explicit disclosure) and Mortgage REIT/Finance (low explicit disclosure).\u003c/p\u003e\n\u003cp\u003eThese differences map onto distinct task profiles and institutional logics (Section 2.4). Trading, surveillance, and payments businesses are data‑ and text‑intensive, operate under strong latency and automation pressures, and compete in technical labor markets, conditions that favor explicit naming of modern toolchains. Credit‑ and yield‑focused vehicles, by contrast, face stronger disclosure tradeoffs and reputational or regulatory incentives that encourage vaguer AI language or silence. Subsector heterogeneity in explicitness thus reflects the signaling and governance tensions highlighted in the literature review and confirms that disclosure granularity is itself an analytic variable rather than a simple proxy for internal adoption.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.3 Outcome Claims Index (OCI): interpreting the \u0026ldquo;zeros\u0026rdquo; (RQ3)\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eRQ3 focuses on quantified outcome claims. To separate adoption from results, the corpus is scanned for sentences that report measurable outcomes such as performance improvements, model metrics (e.g., AUC/F1), loss or risk reductions, latency or efficiency gains, or financial KPIs. These hits form the Outcome Claims Index (OCI) described in Section 3.6. Across subsectors, OCI hits are essentially absent at both the row and firm levels (Appendix Tables S4.5a\u0026ndash;S4.5c). This near‑zero pattern should not be read as evidence of \u0026ldquo;no effect\u0026rdquo; from AI or Python systems but as a disclosure norm: NYSE institutions rarely publish quantified results even when they disclose Python/AI adoption.\u003c/p\u003e\n\u003cp\u003eTo make \u0026ldquo;zero\u0026rdquo; informative, two complementary checks are reported. First, rule‑of‑three 95% upper bounds translate zero events into conservative limits on the true rate (\u0026asymp; 3/n when x = 0; for example, with 30 evidence rows and no claims, the true frequency is \u0026lt; 10%; Appendix Table S4.5d). Second, Jeffreys‑prior credible intervals (Beta [x + 0.5, n \u0026minus; x + 0.5]) place posterior mass tightly near zero in every subsector; the forest plot (Appendix Table S4.5e and Appendix Figure S4.5e) displays posterior means and 95% intervals. Methodologically, pairing frequentist bounds with Bayesian intervals converts the OCI \u0026ldquo;zeros\u0026rdquo; into usable evidence (Agresti, 2019) and directly addresses the measurement gap identified in Section 2.5. The results answer RQ3: quantified outcome claims are rare across all subsectors, and the uncertainty bounds are tight enough to rule out anything but low underlying frequencies in the public record.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.4 Language benchmarking\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo corroborate the Python‑specific lens developed in Sections 2.1\u0026ndash;2.2 and the abstract, sentence‑level tallies compare named mentions of Python, R, Java, SAS, Scala, and MATLAB. Python mentions far exceed those of any other language across subsectors. The language‑by‑subsector heatmap (Figure 4c) shows Python present throughout, with other languages appearing only sporadically. This benchmark substantiates the study\u0026rsquo;s Python focus and aligns with the centrality of the open‑source stack in modern analytics (Pedregosa et al., 2011; Abadi et al., 2016; Paszke et al., 2019; McKinney, 2018). Full counts are reported in Appendix Table S4.6.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.5 Performance‑adjacent proxies: mini‑models\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eGiven the scarcity of direct outcome claims (RQ3), Section 4.5 explores performance‑adjacent proxies that can shed light on capability accumulation without asserting causal effects. Associations are examined between explicit Python disclosure and proxies such as technology‑hiring intensity, technology and R\u0026amp;D operating‑expense ratios, and visibility (evidence volume). A logit specification (reported as odds ratios) and an OLS linear‑probability model with HC1 robust standard errors, both including subsector fixed effects, indicate that these proxies are positively associated with explicit Python disclosure (Appendix Tables S4.7A\u0026ndash;S4.7B; Agresti, 2019). For interpretability, Table S4.7B reports marginal‑effect coefficients, and multiple‑comparison control follows Benjamini\u0026ndash;Hochberg (1995). These mini‑models do not claim that Python disclosure causes better performance. Instead, they triangulate the interpretation consistent with the literature on signaling and capability building that explicit disclosure co‑moves with broader indicators of digital investment and visibility.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.6 Robustness and sensitivity\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTwo diagnostics support the stability of the main findings. First, inverse‑visibility weighting lowers the explicit share from 0.767 to 0.629 yet preserves the conclusion that explicit disclosure is widespread, indicating that results are not an artifact of a few highly visible issuers (Appendix Table S4.1, \u0026ldquo;shares\u0026rdquo;). Second, Benjamini\u0026ndash;Hochberg multiple‑comparison control preserves the principal subsector gaps in explicit disclosure (Appendix Table S4.4, Panel B). Diagnostic cross‑tabs and standardized residuals trace the same over‑ and under‑indexing pattern reported in Figure 4b (Appendix Tables S4.2\u0026ndash;S4.3; Agresti, 2019; Benjamini \u0026amp; Hochberg, 1995). Annual counts of evidence sentences show a clear recent increase in public artifacts related to AI/Python (Appendix Figure S4.6). While visibility is not capability, the time trend is consistent with a secular rise in the salience of AI/Python adoption across NYSE financial institutions.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.8 Summary of findings\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe evidence from the NYSE frame provides a coherent answer to RQ1\u0026ndash;RQ3. Explicit Python disclosure is common and remains dominant after visibility adjustment (RQ1). Disclosure varies systematically by subsector, with high data‑intensity businesses such as investment banking, market infrastructure, and payments exhibiting the highest explicitness, and mortgage REITs and business‑development companies relying more on generic AI language or silence (RQ2). Public quantification of outcomes is rare, but rule‑of‑three bounds and Jeffreys‑prior intervals show that this scarcity is itself informative, placing tight upper bounds on the true frequency of quantified claims even in subsectors that make numerous adoption statements (RQ3).\u003c/p\u003e\n\u003cp\u003ePython\u0026rsquo;s dominance over other named languages is clear, both in the number of mentions and in subsector coverage, which validates the Python‑centered focus of the study. Performance‑adjacent proxies such as technology‑hiring intensity, technology and R\u0026amp;D expenditure, and public visibility move in the expected direction with explicit disclosure and are consistent with the interpretation of disclosure as a signal of capability accumulation rather than proof of superior realized outcomes. All findings are robust to visibility adjustment and multiple‑comparison control and are fully documented in the Appendix and replication package.\u003c/p\u003e"},{"header":"5. Discussion","content":"\u003cp\u003eThis section interprets the empirical patterns documented in Section 4 considering the research gap and questions articulated in the Introduction and Literature Review. The triangulated corpus integrates filings, employer portals, corporate materials, and sectoral press, so all inferences concern public disclosure frequencies rather than private internal use. Consistent with the methodological stance in Section 2.3, visible traces are treated as noisy but informative proxies for underlying capability, not as direct measurements of internal systems (Denzin, 1978; Jick, 1979; Patton, 2015; Yin, 2018).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e5.1 Interpreting the core patterns (RQ1\u0026ndash;RQ3)\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThree headline facts organize the interpretation. First, explicit Python disclosure is prevalent even after inverse‑visibility reweighting, which indicates a broad public adoption signal across NYSE financial institutions (RQ1). The visibility‑adjusted explicit share remains well above one‑half of the sample, so the dominant picture is that Python is openly named rather than hidden in most firms.\u003c/p\u003e\n\u003cp\u003eSecond, subsector heterogeneity is statistically and substantively meaningful (RQ2). Investment Banks, Market Infrastructure \u0026amp; Data firms, and Payment\u0026rsquo;s providers are systematically more likely to disclose explicit Python and named libraries, while Mortgage REIT/Finance and BDC/Finance are systematically less explicit, as documented in Section 4.2 and Appendix Table S4.4. These differences align with institutional logics: market‑facing businesses compete on speed, analytics, and technical credibility, whereas credit vehicles emphasize prudence and stability.\u003c/p\u003e\n\u003cp\u003eThird, the Outcome Claims Index is effectively zero across subsectors, even where adoption signals are strong (RQ3). Python\u0026rsquo;s dominance over R, Java, SAS, Scala, and MATLAB is visible across subsectors (Figure 4c and Appendix Table S4.6). Together, these facts describe a field that converges on a common technical lingua franca while diverging in how and why adoption is disclosed\u0026mdash;exactly the tension anticipated in the literature between platform convergence and heterogeneous signaling strategies (Sections 2.1\u0026ndash;2.5).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e5.2 The non‑disclosure paradox and the OCI \u0026ldquo;zeros\u0026rdquo;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe near‑zero OCI values operationalize the non‑disclosure paradox highlighted in Section 2.5. Institutions routinely name tools and sometimes describe use cases, yet rarely publish quantified outcomes such as return on investment, changes in error rates, risk or loss reductions, or latency and efficiency gains. The study therefore converts zero counts into information via rule‑of‑three upper bounds and Jeffreys‑prior credible intervals, following guidance on sparse data and binomial inference (Agresti, 2019). Both procedures show that in public channels the frequency of quantified outcomes is small across subsectors, even where adoption is widely signaled.\u003c/p\u003e\n\u003cp\u003eThis pattern reconciles a tension visible in the broader literature. External narratives and survey‑based reports stress rapid AI adoption in finance, yet verifiable performance metrics and detailed model diagnostics remain largely internal or confidential. The empirical OCI \u0026ldquo;zeros\u0026rdquo; quantify that gap. Public texts validate that AI and Python systems exist but rarely document how those systems alter risk or efficiency in ways that can be audited by outsiders. In this sense, the study does not contradict evidence of real impact; it shows that public evidence of impact is thin, even where internal adoption may be deep.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e5.3 Theoretical implications: signaling, institutional logic, and isomorphism\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe findings extend signaling accounts by showing that naming languages and libraries functions as audience‑specific credibility work rather than neutral technical description. In signaling models, more informative signals tend to be chosen by higher‑capability types whose marginal cost of signaling is lower than for lower‑capability types (Spence, 1973). In the NYSE financial frame, explicit naming of Python, scikit‑learn, XGBoost, LightGBM, transformer libraries, and proprietary SDKs anchors institutions to a specific technical trajectory and invites benchmarking by technically literate investors, staff, and regulators.\u003c/p\u003e\n\u003cp\u003eVariation across subsectors aligns with the institutional‑logics perspective, which emphasizes that sectors are governed by distinct logics of action that shape what counts as legitimate behavior and talk (Thornton \u0026amp; Ocasio, 2008). Market‑facing logics in trading, market‑data, and payments privilege innovation signaling and speed, so explicit AI and Python disclosures that emphasize advanced architectures and automation are congruent with expectations of technical excellence and agility. Prudential logics in credit‑focused vehicles privilege control signaling, stability, and conservative underwriting, so muted or generic AI language better supports dominant narratives about risk discipline.\u003c/p\u003e\n\u003cp\u003eThe near absence of quantified outcome claims fits with neo‑institutional accounts of decoupling, in which organizations adopt structures and vocabularies to satisfy external expectations while keeping some distance from core technical practices (Meyer \u0026amp; Rowan, 1977). In this case, naming Python and AI serves as a credible signal about inputs, but disclosure about outputs remains thin. Technologies thus operate simultaneously as production tools and as symbols in organizational communication (Orlikowski \u0026amp; Iacono, 2001).\u003c/p\u003e\n\u003cp\u003eAt the same time, the widespread convergence on a shared Python backbone reflects isomorphic pressures in the organizational field (DiMaggio \u0026amp; Powell, 1983). The fact that almost every subsector relies on pandas, NumPy, scikit‑learn, and a short list of deep‑learning and boosting libraries, even though alternative ecosystems remain available, is consistent with coercive pressures from regulators and large counterparties, mimetic responses to uncertainty, and normative influences from professional and developer communities. Professional training, vendor ecosystems, and open‑source communities all present this stack as the canonical solution for financial analytics, and listed institutions reproduce that template in their own disclosures. The result is a layered structure: a shared base stack provides a generic solution to data and model engineering, while subsector‑specific modules (e.g., graph toolkits in payments, probabilistic loss modeling frameworks in credit and insurance) encode local business models and risk profiles.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e5.4 Practical, regulatory, and governance implications\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFor boards and executives, the findings support a shift from viewing AI and Python as a loose collection of tools toward a conception of governed stacks. In such a conception, explicit lifecycle controls including model inventorying, validation cadence, drift and performance monitoring, and decommissioning criteria are applied consistently across models that rely on the shared Python backbone. Reproducible pipelines, documented data lineage, and \u0026ldquo;red‑team\u0026rdquo; routines for adversarial robustness and stress scenarios become central governance artefacts rather than optional best practice. The observed dominance of a small number of libraries implies that investments in controls and documentation around those components can yield broad risk‑management benefits across business lines.\u003c/p\u003e\n\u003cp\u003eFor investors, explicit language and library mentions function as noisy but useful proxies for digital capability. Section 4.5 shows that explicit Python disclosure is positively associated with technology‑hiring intensity, technology and R\u0026amp;D operating expenditure, and evidence volume. These associations are descriptive, not causal, but they suggest that disclosure can be combined with labor‑market signals and platform partnerships to sharpen assessments of AI maturity and digital resilience.\u003c/p\u003e\n\u003cp\u003eFor regulators and supervisors, the near‑zero Outcome Claims Index implies a visibility gap in quantified effects. If public materials continue to omit performance, fairness, and resilience metrics, supervisory frameworks may need to rely more heavily on confidential reporting channels and direct examinations to understand the impact of AI on risk. Lightweight templates for outcome summaries, fairness diagnostics, and model‑family‑level validation metadata could increase public assurance without forcing exposure of proprietary details. Finally, given the concentration of language and library dependence on Python and a handful of tools, supervisory stress testing and operational‑resilience planning should incorporate dependency and common‑mode failure scenarios, such as vulnerabilities in widely used libraries or correlated outages in shared toolchains an extension of existing model‑risk guidance that emphasizes inventories, dependency mapping, and governance for advanced analytics.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e5.5 Contributions to research\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study makes three main contributions that directly address the gaps identified in Section 2 and formalized in RQ1\u0026ndash;RQ3.\u003c/p\u003e\n\u003cp\u003eFirst, it delivers systematic, cross‑subsector evidence on Python‑framed AI disclosure frequencies for a strict NYSE finance universe, using a triangulated corpus that combines filings, employer portals, corporate materials, and sectoral press. This approach advances empirical baselines for the finance‑and‑AI literature and operationalizes long‑standing calls for multi‑source triangulation in organizational research (Denzin, 1978; Jick, 1979; Patton, 2015; Yin, 2018).\u003c/p\u003e\n\u003cp\u003eSecond, it formalizes \u0026ldquo;zeros\u0026rdquo; as evidence by pairing rule‑of‑three upper bounds with Jeffreys‑prior credible intervals, which converts the lack of outcome claims into interpretable uncertainty statements and extends standard practice in sparse‑count inference in applied settings (Agresti, 2019). This directly addresses the measurement gap between adoption signals and outcomes emphasized in Section 2.5 and answered in RQ3.\u003c/p\u003e\n\u003cp\u003eThird, it frames language and library choice as a field‑level signal and documents Python\u0026rsquo;s ubiquity and the sparse incidence of alternatives, thereby refining accounts of disclosure as signaling rather than simple description and complementing broader firm‑level studies of AI adoption that do not track tools at this level of detail (e.g., Acemoglu et al., 2022; McElheran et al., 2024). The resulting subsector‑resolved catalog and replication‑ready pipeline provide a template that can be extended to other exchanges, sectors, and time periods.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e5.6 Limitations and directions for future research\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eSeveral limitations qualify these interpretations and suggest avenues for future work. The design is NYSE‑only and purposive, which likely over‑represents large, complex, and globally active institutions and may not capture disclosure norms in smaller or privately held financial firms. All measures are constructed from public sources, so the analysis observes only visible communication and cannot cover internal models, vendor arrangements, or infrastructures that remain deliberately undisclosed. Institutions that choose not to reveal AI usage for competitive or regulatory reasons may be coded as having no disclosure even when internal adoption is advanced. Sentence‑level dictionary matching with contextual filters reduces but does not eliminate misclassification risk, especially when institutions use idiosyncratic brand names for internal platforms or when key information appears in images or code appendices that are not readily parsed.\u003c/p\u003e\n\u003cp\u003eFuture work can address these limitations by both broadening the frame and deepening the link between disclosure and outcomes. Extending the corpus to other exchanges and jurisdictions would reveal whether similar disclosure equilibria and subsector patterns arise under different regulatory regimes and institutional histories. Linking disclosure‑based measures of AI and Python capability to independent indicators of performance such as stress‑test outcomes, realized loss distributions, or efficiency metrics would allow a more precise assessment of whether explicit disclosure is informative about risk and return in ways that matter for prudential policy and asset allocation. Qualitative research, including interviews with model‑risk managers, technology leaders, and disclosure officers, could illuminate how internal debates about AI and Python translate into the observed mix of explicit naming, generic language, and silence. Such work would connect the quantitative disclosure maps presented here to the organizational processes and institutional pressures that produce them and would further integrate technology, governance, and disclosure perspectives in empirical finance.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAbadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean, J., \u0026hellip; Zheng, X. (2016). TensorFlow: A system for large-scale machine learning. In \u003cem\u003e12th USENIX Symposium on Operating Systems Design and Implementation (OSDI 2016)\u003c/em\u003e (pp. 265\u0026ndash;283). USENIX Association.\u003c/li\u003e\n\u003cli\u003eAcemoglu, D., Autor, D., Hazell, J., \u0026amp; Restrepo, P. (2022). Artificial intelligence and jobs: Evidence from online vacancies. \u003cem\u003eJournal of Labor Economics, 40\u003c/em\u003e(S1), S293\u0026ndash;S340. https://doi.org/10.1086/718327\u003c/li\u003e\n\u003cli\u003eAgresti, A. (2019). \u003cem\u003eAn introduction to categorical data analysis\u003c/em\u003e (3rd ed.). Wiley.\u003c/li\u003e\n\u003cli\u003eAhmed, S., Alshater, M. M., El Ammari, A., \u0026amp; Hammami, H. (2022). Artificial intelligence and machine learning in finance: A bibliometric review. \u003cem\u003eResearch in International Business and Finance, 61\u003c/em\u003e, 101646. https://doi.org/10.1016/j.ribaf.2022.101646\u003c/li\u003e\n\u003cli\u003eAmetrano, F. M., \u0026amp; Ballabio, L. (2013). QuantLib: A free/open-source library for quantitative finance. \u003cem\u003eQuantitative Finance, 13\u003c/em\u003e(4), 597\u0026ndash;606. https://doi.org/10.1080/14697688.2012.744553\u003c/li\u003e\n\u003cli\u003eBahoo, S., Cucculelli, M., Goga, X., \u0026amp; Mondolo, J. (2024). Artificial intelligence in finance: A comprehensive review through bibliometric and content analysis. \u003cem\u003eSN Business \u0026amp; Economics, 4\u003c/em\u003e, 23. https://doi.org/10.1007/s43546-023-00618-x\u003c/li\u003e\n\u003cli\u003eBank of England, \u0026amp; Financial Conduct Authority. (2024). \u003cem\u003eArtificial intelligence in UK financial services\u0026mdash;2024\u003c/em\u003e. Bank of England. https://www.bankofengland.co.uk/report/2024/artificial-intelligence-in-uk-financial-services-2024\u003c/li\u003e\n\u003cli\u003eBenjamini, Y., \u0026amp; Hochberg, Y. (1995). Controlling the false discovery rate: A practical and powerful approach to multiple testing. \u003cem\u003eJournal of the Royal Statistical Society: Series B (Methodological), 57\u003c/em\u003e(1), 289\u0026ndash;300. https://doi.org/10.1111/j.2517-6161.1995.tb02031.x\u003c/li\u003e\n\u003cli\u003eBishop, C. M. (2006). \u003cem\u003ePattern recognition and machine learning\u003c/em\u003e. Springer.\u003c/li\u003e\n\u003cli\u003eBoard of Governors of the Federal Reserve System. (2011). \u003cem\u003eSupervisory guidance on model risk management\u003c/em\u003e (SR 11‑7). Board of Governors of the Federal Reserve System.\u003c/li\u003e\n\u003cli\u003eBrynjolfsson, E., \u0026amp; McAfee, A. (2017). \u003cem\u003eMachine, platform, crowd: Harnessing our digital future\u003c/em\u003e. W. W. Norton.\u003c/li\u003e\n\u003cli\u003eChen, T., \u0026amp; Guestrin, C. (2016). XGBoost: A scalable tree boosting system. In \u003cem\u003eProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining\u003c/em\u003e (pp. 785\u0026ndash;794). ACM. https://doi.org/10.1145/2939672.2939785\u003c/li\u003e\n\u003cli\u003eCoffee, J. C., Jr. (2020). \u003cem\u003eGatekeepers: The professions and corporate governance\u003c/em\u003e (2nd ed.). Oxford University Press.\u003c/li\u003e\n\u003cli\u003eCorporate Finance Institute. (2025). \u003cem\u003eHow is Python used in finance? Applications and examples\u003c/em\u003e. Corporate Finance Institute.\u003c/li\u003e\n\u003cli\u003eCrisanto, J. C., Leuterio, C. B., Prenio, J., \u0026amp; Yong, J. (2024). \u003cem\u003eRegulating AI in the financial sector: Recent developments and main challenges\u003c/em\u003e (FSI Insights on Policy Implementation No. 63). Bank for International Settlements.\u003c/li\u003e\n\u003cli\u003eDenzin, N. K. (1978). \u003cem\u003eThe research act: A theoretical introduction to sociological methods\u003c/em\u003e (2nd ed.). McGraw-Hill.\u003c/li\u003e\n\u003cli\u003eDiMaggio, P. J., \u0026amp; Powell, W. W. (1983). The iron cage revisited: Institutional isomorphism and collective rationality in organizational fields. \u003cem\u003eAmerican Sociological Review, 48\u003c/em\u003e(2), 147\u0026ndash;160. https://doi.org/10.2307/2095101\u003c/li\u003e\n\u003cli\u003eFINRA. (2020). \u003cem\u003eArtificial intelligence (AI) in the securities industry\u003c/em\u003e. Financial Industry Regulatory Authority.\u003c/li\u003e\n\u003cli\u003eFischer, T., \u0026amp; Krauss, C. (2018). Deep learning with long short-term memory networks for financial market predictions. \u003cem\u003eEuropean Journal of Operational Research, 270\u003c/em\u003e(2), 654\u0026ndash;669. https://doi.org/10.1016/j.ejor.2017.11.054\u003c/li\u003e\n\u003cli\u003eGelman, A., Carlin, J. B., Stern, H. S., Dunson, D. B., Vehtari, A., \u0026amp; Rubin, D. B. (2014). \u003cem\u003eBayesian data analysis\u003c/em\u003e (3rd ed.). Chapman \u0026amp; Hall/CRC. https://doi.org/10.1201/b16018\u003c/li\u003e\n\u003cli\u003eGentzkow, M., Kelly, B. T., \u0026amp; Taddy, M. (2019). Text as data. \u003cem\u003eJournal of Economic Literature, 57\u003c/em\u003e(3), 535\u0026ndash;574. https://doi.org/10.1257/jel.20181020\u003c/li\u003e\n\u003cli\u003eGoodell, J. W., Kumar, S., Lim, W. M., \u0026amp; Pattnaik, D. (2021). Artificial intelligence and machine learning in finance: Identifying foundations, themes, and research clusters from bibliometric analysis. \u003cem\u003eJournal of Behavioral and Experimental Finance, 32\u003c/em\u003e, 100577. https://doi.org/10.1016/j.jbef.2021.100577\u003c/li\u003e\n\u003cli\u003eGoodfellow, I., Bengio, Y., \u0026amp; Courville, A. (2016). \u003cem\u003eDeep learning\u003c/em\u003e. MIT Press.\u003c/li\u003e\n\u003cli\u003eGrimmer, J., \u0026amp; Stewart, B. M. (2013). Text as data: The promise and pitfalls of automatic content analysis methods for political texts. \u003cem\u003ePolitical Analysis, 21\u003c/em\u003e(3), 267\u0026ndash;297. https://doi.org/10.1093/pan/mps028\u003c/li\u003e\n\u003cli\u003eHaldane, A. G., \u0026amp; May, R. M. (2011). Systemic risk in banking ecosystems. \u003cem\u003eNature, 469\u003c/em\u003e(7330), 351\u0026ndash;355. https://doi.org/10.1038/nature09659\u003c/li\u003e\n\u003cli\u003eHanley, J. A., \u0026amp; Lippman-Hand, A. (1983). If nothing goes wrong, is everything all right? Interpreting zero numerators. \u003cem\u003eJAMA, 249\u003c/em\u003e(13), 1743\u0026ndash;1745. https://doi.org/10.1001/jama.1983.03330370053031\u003c/li\u003e\n\u003cli\u003eHilpisch, Y. (2018). \u003cem\u003ePython for finance: Mastering data-driven finance\u003c/em\u003e (2nd ed.). O\u0026rsquo;Reilly Media.\u003c/li\u003e\n\u003cli\u003eIon Group. (2024). \u003cem\u003eHow global markets are adapting to AI and digitalization\u003c/em\u003e. ION.\u003c/li\u003e\n\u003cli\u003eJick, T. D. (1979). Mixing qualitative and quantitative methods: Triangulation in action. \u003cem\u003eAdministrative Science Quarterly, 24\u003c/em\u003e(4), 602\u0026ndash;611. https://doi.org/10.2307/2392366\u003c/li\u003e\n\u003cli\u003eKelliher, C. (2022). \u003cem\u003eQuantitative finance with Python: A practical guide to investment management, trading, and financial engineering\u003c/em\u003e. Chapman \u0026amp; Hall/CRC.\u003c/li\u003e\n\u003cli\u003eKe, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., \u0026hellip; Liu, T.-Y. (2017). LightGBM: A highly efficient gradient boosting decision tree. In \u003cem\u003eAdvances in Neural Information Processing Systems\u003c/em\u003e (Vol. 30, pp. 3146\u0026ndash;3154). Curran Associates, Inc.\u003c/li\u003e\n\u003cli\u003eLoughran, T., \u0026amp; McDonald, B. (2016). Textual analysis in accounting and finance: A survey. \u003cem\u003eJournal of Accounting Research, 54\u003c/em\u003e(4), 1187\u0026ndash;1230. https://doi.org/10.1111/1475-679X.12123\u003c/li\u003e\n\u003cli\u003eMcKinney, W. (2018). \u003cem\u003ePython for data analysis: Data wrangling with pandas, NumPy, and IPython\u003c/em\u003e (2nd ed.). O\u0026rsquo;Reilly Media.\u003c/li\u003e\n\u003cli\u003eMeyer, J. W., \u0026amp; Rowan, B. (1977). Institutionalized organizations: Formal structure as myth and ceremony. \u003cem\u003eAmerican Journal of Sociology, 83\u003c/em\u003e(2), 340\u0026ndash;363. https://doi.org/10.1086/226550\u003c/li\u003e\n\u003cli\u003eNazareth, N., \u0026amp; Reddy, Y. V. R. (2023). Financial applications of machine learning: A literature review. \u003cem\u003eExpert Systems with Applications, 219\u003c/em\u003e, 119640. https://doi.org/10.1016/j.eswa.2023.119640\u003c/li\u003e\n\u003cli\u003eNVIDIA. (2024). \u003cem\u003eState of AI in financial services: 2024 trends\u003c/em\u003e. NVIDIA.\u003c/li\u003e\n\u003cli\u003eNVIDIA. (2025). \u003cem\u003eState of AI in financial services: 2025 trends\u003c/em\u003e. NVIDIA.\u003c/li\u003e\n\u003cli\u003eNYSE Arca. (2023). \u003cem\u003eSafeguarding the market: NYSE Arca\u0026rsquo;s surveillance system\u003c/em\u003e. NYSE Arca.\u003c/li\u003e\n\u003cli\u003eOrlikowski, W. J., \u0026amp; Iacono, C. S. (2001). Research commentary: Desperately seeking the \u0026ldquo;IT\u0026rdquo; in IT research\u0026mdash;A call to theorizing the IT artifact. \u003cem\u003eInformation Systems Research, 12\u003c/em\u003e(2), 121\u0026ndash;134. https://doi.org/10.1287/isre.12.2.121\u003c/li\u003e\n\u003cli\u003ePaszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., \u0026hellip; Chintala, S. (2019). PyTorch: An imperative style, high-performance deep learning library. In \u003cem\u003eAdvances in Neural Information Processing Systems\u003c/em\u003e (Vol. 32, pp. 8024\u0026ndash;8035). Curran Associates, Inc.\u003c/li\u003e\n\u003cli\u003ePatton, M. Q. (2015). \u003cem\u003eQualitative research \u0026amp; evaluation methods\u003c/em\u003e (4th ed.). Sage.\u003c/li\u003e\n\u003cli\u003ePedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., \u0026hellip; Duchesnay, \u0026Eacute;. (2011). Scikit-learn: Machine learning in Python. \u003cem\u003eJournal of Machine Learning Research, 12\u003c/em\u003e, 2825\u0026ndash;2830.\u003c/li\u003e\n\u003cli\u003eSpence, M. (1973). Job market signaling. \u003cem\u003eQuarterly Journal of Economics, 87\u003c/em\u003e(3), 355\u0026ndash;374. https://doi.org/10.2307/1882010\u003c/li\u003e\n\u003cli\u003eThornton, P. H., \u0026amp; Ocasio, W. (2008). Institutional logics. In R. Greenwood, C. Oliver, K. Sahlin, \u0026amp; R. Suddaby (Eds.), \u003cem\u003eThe SAGE handbook of organizational institutionalism\u003c/em\u003e (pp. 99\u0026ndash;129). Sage. https://doi.org/10.4135/9781849200387.n4\u003c/li\u003e\n\u003cli\u003eTonello, M. (2025). AI risk disclosures in the S\u0026amp;P 500: Reputation, cybersecurity, and regulation. \u003cem\u003eHarvard Law School Forum on Corporate Governance\u003c/em\u003e.\u003c/li\u003e\n\u003cli\u003eU.S. Department of the Treasury. (2024). \u003cem\u003eUses, opportunities, and risks of artificial intelligence in financial services\u003c/em\u003e. U.S. Department of the Treasury.\u003c/li\u003e\n\u003cli\u003eUberti-Bona Marin, L. G., Rijsbosch, B., Spanakis, G., \u0026amp; Kollnig, K. (2025). Are companies taking AI risks seriously? A systematic analysis of companies\u0026rsquo; AI risk disclosures in SEC 10‑K forms. In \u003cem\u003eProceedings of the SoGood Workshop at ECML PKDD 2025\u003c/em\u003e. (Preprint available at arXiv:2508.19313).\u003c/li\u003e\n\u003cli\u003eU.S. Government Accountability Office. (2025). \u003cem\u003eArtificial intelligence: Use and oversight in financial services\u003c/em\u003e (GAO‑25‑107197). U.S. Government Accountability Office.\u003c/li\u003e\n\u003cli\u003eU.S. Securities and Exchange Commission. (2023). \u003cem\u003eConflicts of interest associated with the use of predictive data analytics by broker-dealers and investment advisers\u003c/em\u003e (Release No. 34‑97990; IA‑6353; File No. S7‑12‑23). U.S. Securities and Exchange Commission.\u003c/li\u003e\n\u003cli\u003eVanderPlas, J. (2022). \u003cem\u003ePython data science handbook: Essential tools for working with data\u003c/em\u003e (2nd ed.). O\u0026rsquo;Reilly Media.\u003c/li\u003e\n\u003cli\u003eVaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., \u0026hellip; Polosukhin, I. (2017). Attention is all you need. In \u003cem\u003eAdvances in Neural Information Processing Systems\u003c/em\u003e (Vol. 30). Curran Associates, Inc. (Original work available at https://arxiv.org/abs/1706.03762)\u003c/li\u003e\n\u003cli\u003eWeil, Gotshal \u0026amp; Manges LLP. (2023). \u003cem\u003eSEC disclosures of artificial intelligence technologies: Survey results and examples from 2023 company filings\u003c/em\u003e. Weil, Gotshal \u0026amp; Manges LLP.\u003c/li\u003e\n\u003cli\u003eYin, R. K. (2018). \u003cem\u003eCase study research and applications: Design and methods\u003c/em\u003e (6th ed.). Sage.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"NYSE, financial institutions, Python, disclosure, libraries, Outcome-Claims Index, triangulation, replication","lastPublishedDoi":"10.21203/rs.3.rs-8884680/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8884680/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eArtificial intelligence (AI) is diffusing rapidly across financial services, but public disclosure of AI and software capability remains sparse and heterogeneous, even as Python has become the dominant language for analytics and model deployment. This research constructs a systematic map of \u003cb\u003ePython-framed AI disclosure\u003c/b\u003e for a strict frame of 180 New York Stock Exchange (NYSE) financial institutions. A triangulated corpus of regulatory filings, employer portals, corporate communications, and sectoral press is assembled through January 2026, and each firm is coded into one of three disclosure states: explicit Python, indirect AI, or none. Named Python libraries are detected via sentence-level dictionary matching and contextual filters, then mapped to seven analytical dimensions (natural language processing, machine learning, deep learning, reinforcement learning, probabilistic modeling, optimization, and visualization). An Outcome Claims Index (OCI) flags quantified performance assertions (e.g., risk reduction, accuracy gains) and supports both frequentist and Bayesian inference when such claims are rarely observed.\u003c/p\u003e \u003cp\u003eThe results show that 76.7% of firms disclose explicit Python usage, 2.2% disclose only indirect AI references, and 21.1% disclose neither, with a visibility-weighted explicit share of 0.629. Disclosure patterns vary strongly by subsector, and a common backbone of pandas, NumPy, and scikit-learn coexists with toolkits tailored to text-intensive, tabular risk, payments, and market microstructure tasks. OCI values are effectively zero across subsectors, indicating that quantified outcome claims are rarely placed in the public record. The study delivers a subsector-resolved empirical catalog of Python adoption in finance and a replication-ready pipeline for measuring tool-level AI disclosure.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e","manuscriptTitle":"Public Signals of Python‑Enabled AI in Finance: Disclosure Patterns and Outcome Claims in NYSE Institutions","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-18 08:09:57","doi":"10.21203/rs.3.rs-8884680/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"37ad9ce2-e10f-4159-8179-129b5a5b7fae","owner":[],"postedDate":"February 18th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":62950911,"name":"Finance"},{"id":62950912,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2026-02-18T08:09:57+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-18 08:09:57","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8884680","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8884680","identity":"rs-8884680","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}