The rise of artificial intelligence in algorithmic trading: a bibliometric study

Artificial intelligence (AI) is transforming how trading strategies are generated, evaluated, and executed in financial markets. The adoption of machine learning (ML), deep learning (DL), and reinforcement learning (RL) has shifted algorithmic trading from traditional rule-based systems towards adaptive, data-driven systems that learn nonlinear relationships, detect structural market patterns, and automate decision-making at scale.

Beyond its technical applications, the adoption of AI in financial markets is part of a broader process of digital transformation that is reshaping organisational decision-making, innovation processes, and competitive dynamics in the financial sector. AI-driven trading systems influence how investment firms allocate resources, manage risk, process information, and respond to changing market conditions. Algorithmic decision systems have consequently evolved from specialised analytical tools into strategic assets that affect organisational performance, market efficiency, and the structure of financial intermediation. Understanding the evolution of research in this area is therefore relevant not only from a technological perspective but also from management, innovation, and business-economics perspectives.

As data availability, computational capacity, and market microstructure complexity have expanded, AI models have moved from peripheral forecasting tools to central components of quantitative trading and market-microstructure research. Despite this expansion, the literature remains fragmented and conceptually under-integrated. Prior studies have examined specific algorithms, asset classes, or performance metrics, but no data-driven synthesis connects the evolution of methodological approaches, the structure of scientific collaboration, and the emerging research fronts within a unified analytical framework. In particular, little is known about how the intellectual foundations of AI-driven trading have developed over time, whether the field is converging towards a dominant paradigm or remains methodologically dispersed, and what these patterns imply for future theoretical development and managerial practice. This absence of an integrative perspective limits both academic understanding and the capacity of researchers, practitioners, and policymakers to assess the field's likely direction. To address this gap, the present study employs bibliometric techniques to map the evolution, impact, and thematic structure of AI research in algorithmic trading, offering a structured foundation for theory building, practical decision-making, and future research agendas.

By mapping more than three decades of scholarship, this study clarifies how the field has developed, where its core knowledge structures lie, and how emerging paradigms are shaping its trajectory.

The growth of AI-driven trading has been partly motivated by the limitations of traditional financial models in handling high-dimensional, nonlinear, and non-stationary market behaviour. Early academic contributions showed the potential of ML methods relative to classical econometric approaches, particularly their capacity to capture complex temporal structures and extract predictive signals from noisy data. Previous studies have reported that AI-based systems can outperform conventional linear benchmarks across diverse markets and trading horizons (Chen et al., 2016; e.g. Li et al., 2016; Sezer and Ozbayoglu, 2018; Shynkevich et al., 2017). These studies created a methodological foundation that subsequent research refined through increasingly sophisticated architectures.

As DL matured, researchers widely adopted Long Short-Term Memory (LSTM) networks, Convolutional Neural Networks (CNNs), and hybrid DL models to analyse financial time series, identify market regimes, and generate trading signals. Several studies reported advantages of these approaches in capturing long-range dependencies and uncovering latent market structures. LSTM-based prediction and policy-learning frameworks, hybrid convolutional-recurrent architectures, and attention-enhanced models, for example, reported strong forecasting and trading performance across equities, foreign-exchange, and cryptocurrency markets (Gómez-Martínez et al., 2022; Qing et al., 2022; e.g. Wu et al., 2019; Ye and Schuller, 2023). Collectively, this body of work positioned DL as a central methodological pillar in AI-based trading, often extending or outperforming traditional time-series approaches in predictive accuracy and adaptability.

RL expanded this paradigm further by conceptualising trading as a sequential decision-making problem in which an autonomous agent interacts with the market, receives rewards, and learns optimal policies. RL has been particularly influential in the development of end-to-end trading systems capable of self-optimisation. A growing body of research, from early RL portfolio frameworks for cryptocurrency markets to more recent studies incorporating macroeconomic indicators, multi-agent interactions, and derivatives, illustrates the versatility of RL-driven trading (Conegundes and Pereira, 2024; e.g. Jiang and Liang, 2017; Lin et al., 2022; Park et al., 2024; Sang et al., 2025). These contributions point to RL's capacity to integrate heterogeneous information sources, adapt to regime shifts, and optimise long-horizon risk-adjusted performance.

AI-driven trading research has also expanded beyond price prediction and strategy generation to encompass portfolio optimisation, trend-following, explainability, behavioural modelling, and volatility forecasting. Studies report that ML-enhanced allocation models can outperform traditional portfolio construction approaches, particularly under shifting market conditions or in emerging markets characterised by structural inefficiencies (e.g. Kawy et al., 2021; Ye and Schuller, 2023). AI-based trend-following systems have shown robustness during periods of extreme volatility, such as the COVID-19 crisis, outperforming rule-based strategies by adjusting dynamically to changing momentum structures (e.g. Gómez-Martínez et al., 2022). The integration of behavioural and sentiment-based data adds a further dimension. Studies incorporating investor sentiment, macroeconomic narratives, and volatility indicators into ML models have reported performance improvements and offer insights into market psychology. Examples include ML classifiers applied to sentiment-enhanced datasets for forecasting equity indices and models predicting volatility indices in markets with limited historical data (He et al., 2021; e.g. Wang and Xie, 2024). These contributions mark a convergence between behavioural finance and AI, expanding the informational scope of algorithmic trading.

Beyond their predictive capabilities, AI-based trading systems can also be interpreted through theoretical perspectives on market adaptation, behavioural finance, and strategic decision-making. From a market-adaptation perspective, AI models are mechanisms through which financial actors adjust to evolving information environments and competitive conditions. Behavioural finance suggests that market outcomes are influenced by cognitive biases, sentiment dynamics, and information-processing limitations, factors that AI systems increasingly seek to identify, exploit, or mitigate. At the organisational level, AI-driven trading platforms function as strategic decision-support systems, enabling firms to process large volumes of information, automate complex decisions, and respond quickly under uncertainty. The evolution of AI in algorithmic trading therefore reflects not only technological progress but also broader changes in how financial organisations generate, interpret, and act upon market knowledge.

Cryptocurrency markets, characterised by extreme volatility, decentralised structure, and rapid innovation, have served as a major proving ground for AI-driven trading research. DL and RL models have been widely deployed to capture rapid price swings, detect structural anomalies, and optimise cryptocurrency portfolios. The development of explainable-AI tools for cryptocurrency trading reflects a broader shift towards interpretability and transparency, addressing concerns about model opacity in markets with high volatility and substantial speculative activity (e.g. Fior et al., 2022; Jiang and Liang, 2017; Qing et al., 2022).

As the literature has matured, scholars have increasingly recognised the limitations of static learning systems in nonstationary and adversarial markets. This recognition has motivated meta-learning frameworks and cognitive-theoretic RL models designed to improve adaptability. Research in this developing area suggests that meta-RL and game-theoretic learning mechanisms can improve strategy robustness in markets characterised by structural shifts and interactions among competing agents (Lin et al., 2022; Park et al., 2024; e.g. Shen and Huang, 2025).

More recently, transformer architectures and large language models (LLMs) have opened a new stage in AI-driven financial decision-making. Earlier research focused primarily on predictive performance using ML, DL, and RL techniques. Current developments increasingly emphasise AI systems that integrate structured and unstructured information, support human decision-making, and operate within explainable and auditable environments. Recent research shows that LLM-based architectures are being developed as transparent, domain-adaptive components that support data management, semantic reasoning, regulatory compliance, and organisational governance rather than merely improving prediction accuracy (Karras et al., 2025). Similarly, agentic AI architectures integrating LLMs, retrieval-augmented generation (RAG), multimodal reasoning, and human-in-the-loop mechanisms reflect a transition towards intelligent financial systems that prioritise explainability, accountability, and adaptive decision support (Ahi et al., 2025). These developments suggest that AI in finance is moving beyond predictive modelling towards trustworthy, governance-oriented decision systems, a context relevant for interpreting the recent thematic trends identified in the present bibliometric analysis.

Despite the breadth of methodological innovation and domain expansion, the research remains fragmented. Studies are distributed across finance, engineering, computer science, and AI journals. Co-authorship networks exhibit weak global integration, and the field's intellectual structure is anchored by a relatively small number of highly cited works. These characteristics call for a structured, data-driven analysis of the field's development and of its thematic and collaborative architecture. Bibliometric methods are well suited to this task: they map citation structures, visualise thematic clusters, quantify collaboration networks, and track temporal research dynamics. Prior bibliometric work in finance and technology shows how science-mapping techniques capture the diffusion of innovations, the emergence of research fronts, and the evolution of interdisciplinary fields (Bhuiyan et al., 2025; Chen et al., 2025; e.g. Horobet et al., 2024). Applying these methods to AI in algorithmic trading therefore offers a framework for understanding how the field has evolved and where it is headed.

The contribution of this study lies in providing an integrative bibliometric assessment of research on artificial intelligence in algorithmic trading. Recent studies have examined related topics, including algorithmic trading, deep learning applications, and sentiment analysis in financial markets, but existing reviews tend to focus on specific technologies, narrower thematic domains, or shorter observation periods. This study instead examines the evolution of AI-driven algorithmic trading research over more than three decades (1993–2025) and combines publication trends, citation impact analysis, collaboration networks, and thematic mapping within a single analytical framework.

Unlike previous bibliometric and review studies, which have addressed algorithmic trading in general, specific AI techniques (e.g. deep learning), or particular financial applications such as sentiment analysis, this study offers the first bibliometric assessment devoted exclusively to artificial intelligence in algorithmic trading. Integrating publication trends, citation impact, collaboration networks, and thematic evolution over the period 1993–2025 yields a broader and more cohesive account of the intellectual development of this interdisciplinary research domain.

By integrating these complementary dimensions, the study describes how the field has evolved, identifies the publications and authors that have shaped its development, reveals patterns of scientific collaboration, and uncovers emerging research fronts. Rather than focusing exclusively on methodological advances, the analysis also examines the interdisciplinary convergence between artificial intelligence, quantitative finance, and managerial decision-making. In doing so, the study contributes to a better understanding of the intellectual evolution of AI-based trading research and provides a structured foundation for future theoretical, methodological, and practical developments.

Although several review studies have examined related areas, their scope differs substantially from that of the present work. Horobet et al. (2024) mapped the broader literature on algorithmic trading without focusing specifically on artificial intelligence techniques. Bhuiyan et al. (2025) conducted a systematic review of deep learning applications for algorithmic trading, emphasising predictive models and optimisation strategies rather than the overall intellectual structure of the field. Chen et al. (2025) analysed the bibliometric evolution of sentiment analysis in stock market research, concentrating on a specific AI application rather than the wider AI-driven trading ecosystem. In contrast, the present study provides the first bibliometric mapping focused specifically on artificial intelligence in algorithmic trading, integrating machine learning, deep learning, reinforcement learning, citation structures, collaboration networks, and thematic evolution over more than three decades (1993–2025). This broader perspective allows us to identify not only publication trends but also the intellectual foundations, interdisciplinary connections, and emerging research fronts of the domain.

Table 1 summarises the positioning of the present study relative to previous reviews.

Specifically, the study has four objectives: (1) to examine the temporal evolution of research output and identify key publication trends; (2) to assess the scientific influence of the journals, authors, and articles shaping the field; (3) to analyse collaboration patterns through co-authorship networks and evaluate the degree of interdisciplinary integration; and (4) to identify the main thematic structures and emerging research fronts through term co-occurrence analysis. By addressing these objectives, the study seeks to clarify the intellectual foundations of AI-driven algorithmic trading and to offer a structured basis for theoretical development, managerial insight, and future research agendas in finance, management, and business economics.

Given the rapid expansion, interdisciplinary nature, and methodological diversity of research on AI in algorithmic trading, a bibliometric approach is well suited to examining the field. Unlike traditional literature reviews, which typically synthesise findings from a limited number of studies, bibliometric methods enable the systematic analysis of large bodies of literature, revealing publication trends, influential contributions, collaboration patterns, and emerging thematic structures. The search strategy and data-collection procedures used to construct the bibliometric dataset are described in the Methodology section.

Bibliometric analysis has become a central method of research evaluation. Since Garfield's introduction of citation indexing in the 1960s (Garfield, 1964), bibliometric approaches have developed into tools for mapping the production, dissemination, and influence of knowledge. Supported by databases and software such as Web of Science, Scopus, VOSviewer, and bibliometrix, bibliometric analysis is particularly valuable in rapidly developing fields with high publication volumes (Aria and Cuccurullo, 2017). Applied to AI in algorithmic trading, it provides both a historical overview of the field and an assessment of its current research frontiers. When applied to algorithmic trading, bibliometric analysis can identify distinct streams of research. Recent studies point to clusters focused on ML and DL methods for stock price prediction and time series forecasting (Bhuiyan et al., 2025), as well as RL and optimisation techniques that support automated decision-making in trading strategies (Horobet et al., 2024). Other notable clusters include the integration of sentiment analysis and behavioural indicators into trading systems, where unstructured data such as news and social media play a growing role (Chen et al., 2025), and the examination of risk management and market microstructure, with particular attention to liquidity, volatility, and the stability of algorithmic systems (Demir et al., 2022).

To investigate the research landscape of AI in algorithmic trading, we conducted a bibliometric analysis based on publications indexed in the Web of Science (WoS) Core Collection (Clarivate Analytics, 2025), a widely used multidisciplinary database for scholarly output. The Web of Science Core Collection was selected for its rigorous journal selection criteria, standardised bibliographic records, and widespread use in bibliometric research, which support the reliability and reproducibility of science-mapping analyses. Broader databases such as Scopus or Dimensions include additional conference papers and regional publications; whether their inclusion would materially alter the publication patterns or intellectual structure reported here remains an empirical question.

The search strategy was designed to capture the intersection between AI and algorithmic trading. The search query applied was:

Here, TS denotes the Topic Search field, which includes title, abstract, and author keywords. The search was conducted within the Web of Science Core Collection.

The search strategy prioritised precision over breadth by focusing on terms directly associated with algorithmic and quantitative trading. Although this approach reduced the inclusion of peripheral studies, it may have excluded relevant contributions using alternative terminology, such as automated trading, trading systems, or AI-based financial decision-making.

The dataset comprised 417 documents published between 1993 and 2025 (Figure 1). The dataset consisted primarily of journal articles (70%), followed by conference proceedings (29%), and a small proportion of book chapters and reviews. For each record, we collected the available metadata, including title, authors, source, year of publication, abstracts, keywords, and citation counts. Unlike systematic literature reviews, which typically involve manual screening and exclusion stages, the present bibliometric study relied on a targeted search strategy. All 417 records retrieved from the Web of Science Core Collection were therefore retained for analysis, and no additional exclusion criteria were applied.

In accordance with established bibliometric protocols, the study used VOSviewer to conduct several network analyses. Descriptive statistics and figures were derived from the Web of Science export and VOSviewer outputs (Van Eck and Waltman, 2010). Co-authorship analysis was used to examine collaboration patterns, whereas term co-occurrence analysis based on titles and abstracts was used to identify the field's main thematic clusters. Citation analysis identified the most influential publications and contributors. Although export constraints prevented complete co-citation and keyword-network analyses, the available term co-occurrence networks provided useful evidence on the field's thematic organisation and research trends.

The evolution of publications in this field reveals a clear trajectory of growth. The earliest publication in the dataset appeared in 1993, yet research activity remained limited for more than two decades. A marked expansion began around 2018, coinciding with the growing adoption of DL and RL in financial applications. Publication output reached notable peaks in 2022 with 65 documents and in 2024 with 90, whereas the dataset contains 44 publications for 2025 as of 4 September, meaning that the final annual count is incomplete (Figure 2).

This trajectory reflects the emerging and rapidly expanding nature of the field.

Journal articles constitute approximately 70% of the dataset, reflecting the increasing consolidation of the field in peer-reviewed scholarly outlets. Conference proceedings account for approximately 29%, indicating the field's continuing links to computer science and engineering communities, in which preliminary findings are often disseminated through conferences. Reviews and book chapters constitute only a small proportion of the dataset. This acceleration in publication volume indicates that AI in algorithmic trading has moved from an experimental research niche towards an established interdisciplinary domain. The post-2018 surge coincides with wider access to computational resources (GPU/TPU), open-source frameworks (TensorFlow, PyTorch), and the mainstreaming of DL and RL. The temporal pattern also matches the diffusion curve typical of emerging technological paradigms: an initial exploratory phase (pre-2010), rapid expansion during methodological innovation (2018–2024), and a current phase of consolidation marked by increased journal specialisation. The predominance of journal articles over conference proceedings is a further indication that AI-based trading research is moving from technical feasibility studies towards theoretical and applied maturity.

The most relevant journals by publication count are presented in Table 2.

These journals illustrate the interdisciplinary nature of the field, spanning AI, computer science, and finance.

Table 3 shows the most productive authors by publication count.

Several other authors have 2 or 3 publications, reflecting a relatively dispersed author network without a single dominant research group.

Beyond descriptive indicators of productivity and impact, bibliometric methods allow for the exploration of the relational structure of a research field. Network analysis, applied to bibliographic data, captures how knowledge is created and disseminated through collaborations, citations, and thematic linkages. Mapping these relationships makes it possible to identify the intellectual structure of a domain, the main communities of researchers, and the emerging research fronts.

The co-authorship analysis indicates that collaboration within this field remains limited and dispersed. A few small groups of researchers co-publish, but most studies are produced either by individual authors or by teams that rarely collaborate beyond their immediate circles. The largest connected components contain only three to five authors, and many researchers appear as isolated nodes. This pattern reflects the emerging and interdisciplinary character of the field, in which expertise is spread across computer science, finance, and engineering departments. Among the most productive contributors are Gil Cohen and Yuling Huang, yet no single research hub or leading collaborative network has formed.

Figure 3 presents the co-authorship network for the 417 documents analysed. The map shows highly fragmented collaboration patterns: the majority of authors appear as isolated nodes or form very small clusters. Only a few weakly connected groups can be observed, and there is no large cohesive network of researchers working systematically together. This pattern is consistent with an emerging field in which contributions remain dispersed across institutions and disciplines, and most articles are authored by small independent teams.

The fragmentation observed in the co-authorship network suggests that knowledge diffusion in AI-based trading occurs through parallel innovation streams rather than centralised research hubs. Such structural dispersion is typical of rapidly evolving technological fields, where expertise develops simultaneously across institutions and geographies. This limits cross-fertilisation of ideas, but it also fosters methodological diversity, which may explain the rapid proliferation of heterogeneous models (Convolutional Neural Networks, Long Short-Term Memory, RL agents) in the literature. From an innovation-theory standpoint, this network topology indicates a pre-paradigmatic phase in which localised expertise precedes the formation of global epistemic communities. Strengthening cross-institutional collaboration could therefore accelerate the standardisation of methods and the translation of AI innovations into financial practice.

A co-occurrence analysis of terms drawn from titles and abstracts revealed three main thematic clusters within the field. The first, centred on methodological and experimental approaches, includes terms such as analysis, novel, propose, system, models, and accuracy, reflecting research devoted to the development and testing of new ML and AI techniques. The second cluster focuses on applications to financial trading, with frequent terms such as strategies, trading, stock, returns, financial, and results, capturing work on the design of trading strategies, the assessment of financial performance, and the practical integration of AI into trading systems. The third cluster, related to time-series and forecasting, is smaller but notable; it includes terms such as time-series, long-term, technique, and success, reflecting research that applies AI to time-series prediction and long-horizon financial forecasting.

Figure 4 presents the co-occurrence map of terms extracted from titles and abstracts, produced using VOSviewer. The visualisation identifies four principal clusters: methodological and experimental approaches (blue and green), applications to trading and financial markets (green), studies focused on accuracy, datasets, and forecasting methods (yellow), and conceptual or ethical discussions surrounding artificial intelligence (red). The map illustrates the field's dual focus on technical innovation and financial application, while also revealing emerging trends such as the growing interest in transformer-based models, represented by the purple cluster.

These clusters confirm the field's dual orientation towards technical and methodological innovation on the one hand and practical financial-market applications on the other.

The publications in the dataset had received 4,605 citations in total as of the data-extraction date, with an average of 11 citations per publication. The most frequently cited work is that of Sezer and Ozbayoglu (2018), titled Algorithmic financial trading with deep convolutional neural networks: Time series to image conversion approach, published in Applied Soft Computing, which has received 253 citations. This study introduced a method that converts financial time series into two-dimensional images derived from technical indicators, thereby enabling the use of convolutional neural networks originally developed for image recognition in financial data analysis. The proposed CNN model classifies market patterns into buy, sell, or hold signals and outperformed the reported traditional benchmarks, including Buy-and-Hold and other standard trading systems, in out-of-sample tests for both stocks and exchange-traded funds. By adapting computer-vision techniques to the financial domain, this study established a methodological benchmark for subsequent research on DL approaches to price prediction and algorithmic trading.

Several highly cited contributions, each with more than 100 citations, have shaped research at the intersection of AI and algorithmic trading. Chen et al. (2016) provided early evidence of the applicability of deep CNNs to noisy financial time-series data. Jiang and Liang (2017) introduced a model-free deep-RL framework for cryptocurrency portfolio management and reported strong back-tested performance relative to conventional strategies. Shynkevich et al. (2017) showed that forecasting performance depends materially on the relationship between the input-window length and the forecast horizon. Théate and Ernst (2021) subsequently introduced the Trading Deep Q-Network and a structured methodology for evaluating risk-adjusted trading performance. Together, these studies illustrate the growing influence of DL and RL on algorithmic-trading research and on standards of empirical performance evaluation.

The concentration of citations among a small number of studies suggests that early, high-impact contributions have disproportionately influenced subsequent research. This citation asymmetry reveals the field's dependence on methodological benchmarks, particularly those linking neural architectures to financial prediction. The position of Sezer and Ozbayoglu (2018) exemplifies the role of foundational papers in establishing conceptual legitimacy for AI-driven trading models. At the same time, because a complete co-citation analysis could not be conducted, the degree of theoretical integration cannot be assessed conclusively from co-citation evidence. This leaves room for future work to synthesise methodological advances into coherent theoretical frameworks that bridge financial modelling, ML, and behavioural finance.

The bibliometric analysis shows the rapid growth and diversification of research on AI in algorithmic trading over the past decade. The descriptive indicators confirm that the field has experienced a sharp increase in scientific output since 2018, coinciding with the widespread adoption of DL and RL techniques in finance. Journals such as Expert Systems with Applications and IEEE Access have emerged as key publication venues, while contributions remain relatively dispersed across other outlets in computer science and applied intelligence.

Citation metrics show that a small group of highly cited studies serves as the field's methodological foundation. The mean of approximately 11 citations per publication indicates growing scholarly visibility, although citation counts are sensitive to publication age. This concentration may encourage cumulative development around established architectures, but it may also narrow theoretical diversity if subsequent studies focus primarily on incremental extensions.

The co-authorship network shows that collaboration remains weak and fragmented. Most authors appear as isolated nodes or in very small clusters, with no dominant research hub or cohesive international network. This pattern is typical of emerging interdisciplinary domains, where scholars from finance, computer science, and engineering work in parallel but have not yet established strong collaborative ties. Strengthening co-authorship networks could improve knowledge integration and accelerate methodological innovation in the field.

The fragmented collaboration network suggests that AI-based trading research is still characterised by dispersed communities operating across finance, computer science, and engineering. This fragmentation encourages methodological diversity and experimentation, but it may also slow the development of common research standards, reduce knowledge transfer across disciplines, and delay the emergence of shared theoretical frameworks.

The co-occurrence map of terms reveals a more structured intellectual organisation within the field. Four main thematic clusters can be distinguished. The first, centred on methodological development, focuses on the creation of new models and the evaluation of their performance. The second, devoted to applications in trading strategies, examines stock prediction, financial returns, and automated trading systems. The third cluster, related to forecasting and time-series modelling, explores comparisons between traditional and AI-based predictive methods. The fourth, encompassing conceptual and ethical perspectives, addresses the broader implications and risks associated with the use of AI in financial markets. A smaller emerging cluster indicates growing interest in transformer-based architectures, a potential new front for research.

Viewed through the lens of technological diffusion and paradigm formation, the results indicate that AI in algorithmic trading is moving from a pre-paradigmatic stage towards early consolidation. The initial phase (pre-2018) was characterised by dispersed experimentation and heterogeneous methodologies, while the subsequent expansion phase shows increasing convergence towards DL and RL paradigms. Thematic clusters and citation hierarchies now function as epistemic anchors, structuring how new entrants to the field frame problems and methods. Yet the fragmented co-authorship network reveals that institutional diffusion lags behind intellectual diffusion: methodological ideas spread quickly, while cross-institutional collaboration has yet to catch up.

The bibliometric evidence also points to the interdisciplinary nature of this research ecosystem. AI-driven trading sits at the intersection of computational science, quantitative finance, and behavioural economics, domains that differ in their epistemic assumptions and validation standards. The observed thematic clusters reflect these tensions: one cluster prioritises algorithmic precision and model optimisation, while another focuses on interpretability, ethics, and market implications. Such epistemic plurality fragments collaboration networks in the short term, but it also fuels innovation by maintaining methodological diversity. In the long term, bibliometric trends suggest gradual epistemic integration, in which shared methodological norms (benchmarking, reproducibility, and hybrid models) may support closer links between AI research and financial-market theory.

Overall, these findings indicate that the field is simultaneously maturing and diversifying: intellectual structures are beginning to consolidate around a few methodological and applied clusters, while the social structure of research remains fragmented. The combination of strong thematic coherence and weak collaboration networks suggests that AI-enabled algorithmic trading is a developing research area with considerable potential for integration across disciplines and geographies.

The rapid growth of publications on AI in algorithmic trading can be largely explained by the convergence of technological, economic, and institutional drivers. Advances in computational power, the widespread availability of financial data, and the maturation of ML frameworks have substantially lowered entry barriers, enabling researchers from diverse disciplines to contribute to the field. At the same time, the economic relevance of algorithmic trading and the strategic value of AI-driven decision systems have intensified academic and professional interest, accelerating publication output since 2018. This expansion has, however, occurred in a fragmented manner. The bibliometric evidence shows that research has developed through connected streams across finance, computer science, and engineering, reflecting differences in epistemic traditions, validation standards, and publication outlets. Such fragmentation limits cumulative knowledge building and interdisciplinary collaboration, but it has also fostered methodological diversity and rapid experimentation. Looking ahead, the field is likely to enter a phase of consolidation characterised by stronger theoretical integration, increased cross-disciplinary collaboration, and greater emphasis on the explainability, governance, and robustness of AI-based trading systems. As regulatory scrutiny intensifies and market environments become more complex, future research is expected to move beyond performance optimisation towards integrative frameworks that align technological innovation with managerial decision-making and financial stability.

Taken together, the bibliometric patterns identified in this study suggest that AI in algorithmic trading is following a trajectory similar to that observed in other emerging technology-driven fields. The coexistence of strong thematic convergence and weak collaboration networks indicates that methodological consolidation is occurring more rapidly than institutional integration. Researchers are increasingly converging on a shared set of tools and analytical approaches, particularly deep learning and reinforcement learning, even though collaboration remains fragmented across disciplines and research communities. This pattern is characteristic of fields undergoing paradigm formation, in which intellectual convergence precedes the emergence of stable collaborative ecosystems. If this trend continues, future research is likely to become more interdisciplinary, integrating advances in artificial intelligence with broader concerns related to financial decision-making, governance, explainability, and market resilience.

From a theoretical perspective, this study contributes to the literature on management and business economics by clarifying the intellectual structure and evolutionary dynamics of AI in algorithmic trading. The mapping of citation patterns and thematic clusters shows that the field is converging towards a limited set of dominant methodological paradigms, particularly DL and RL, while remaining fragmented in terms of interdisciplinary integration and collaborative structures. This tension between methodological convergence and institutional dispersion points to a theoretical gap in how AI-driven financial innovation is conceptualised. The findings suggest opportunities for finance and management theory to better account for adaptive, data-driven decision systems, which complicate conventional assumptions regarding market efficiency, rationality, and model stability. By positioning AI-based trading as an emerging technological paradigm rather than a collection of isolated techniques, this study provides a conceptual foundation for future theory building that integrates financial economics, strategic management, and AI research.

The identification of emerging research fronts, such as transformer-based architectures, explainable AI, and meta-learning frameworks, extends current theoretical debates by signalling a shift from performance-oriented modelling towards concerns related to interpretability, robustness, and governance. These trends indicate that future theoretical development in algorithmic trading will need to move beyond predictive accuracy and incorporate organisational, ethical, and regulatory dimensions, reinforcing the relevance of management-oriented perspectives in AI-driven financial research.

For practitioners, including asset managers, quantitative trading firms, and fintech organisations, the findings offer insights into the strategic landscape of AI-driven algorithmic trading. The concentration of academic citations around a limited number of model architectures may indicate methodological homogeneity within the research literature; bibliometric evidence alone, however, cannot establish the degree of model homogeneity in industry practice. Practitioners should nevertheless consider diversity in model design, data sources, validation methods, and learning frameworks to improve resilience across changing market conditions.

The fragmented collaboration networks identified in this study also carry implications for organisational strategy. Although bibliometric collaboration patterns cannot be mapped directly onto organisational structures, the findings point to the potential value of interdisciplinary teams when developing AI-based trading systems. Encouraging collaboration between quantitative analysts, data scientists, and risk managers can improve model design, deployment, and oversight. In this regard, the growing prominence of explainable-AI tools represents a practical opportunity to connect technical performance with managerial accountability, supporting better communication between technical and non-technical stakeholders.

Finally, the emergence of ethical and governance-related research themes indicates the importance of responsible AI adoption in financial markets. Regulators and practitioners alike can use the structured overview provided by this study to identify issues likely to attract regulatory attention, including transparency, model risk, and market stability. By aligning trading innovation with explainability, robustness, and ethical considerations, organisations can improve performance while strengthening trust and legitimacy in increasingly automated financial environments.

The findings also have broader implications for financial institutions, regulators, and policymakers responsible for overseeing AI-enabled financial markets. Financial institutions should strengthen governance frameworks for AI-based trading systems by incorporating model validation, continuous monitoring, and explainability mechanisms to mitigate operational and model risks. For regulators and policymakers, the growing complexity of AI-driven trading points to the need for standards that balance innovation with transparency, accountability, and market stability. The bibliometric evidence presented in this study suggests that governance-related research is becoming an increasingly important component of the field, supporting the development of regulatory approaches capable of addressing the systemic implications of autonomous financial decision-making.

This study is subject to certain limitations. First, it relies exclusively on the Web of Science Core Collection, which, while authoritative, may omit relevant contributions indexed in Scopus, IEEE Xplore, or SSRN. Second, due to database export constraints, complete co-citation and author-keyword network analyses could not be conducted, limiting the scope of the intellectual structure captured. Finally, the field itself is still in a state of flux, meaning that bibliometric patterns will likely evolve rapidly in the coming years.

Future research should expand bibliometric mapping by incorporating multiple databases and by applying complementary techniques such as bibliographic coupling or text mining of full articles. As the field matures, it will also be important to study how interdisciplinary collaborations evolve, how new AI architectures (e.g. transformers, generative models) reshape trading research, and how ethical, regulatory, and societal concerns influence both academic and practical developments in AI-driven finance.

This study maps the development of research on AI in algorithmic trading between January 1993 and September 2025. Publication output increased sharply after 2018, and a limited number of journals and highly cited studies now serve as important publication outlets and intellectual anchors. Deep learning and reinforcement learning dominate the methodological landscape, while transformer-based, explainable, and governance-oriented approaches are emerging.

The network analyses reveal a contrast between the field's intellectual and social structures. Term co-occurrence patterns show increasingly coherent thematic groupings around methodological innovation, trading applications, forecasting, and ethical or governance concerns. By contrast, the co-authorship network remains fragmented, with limited cross-institutional and cross-disciplinary integration. The field therefore appears to be moving towards early intellectual consolidation without a correspondingly integrated collaborative structure.

These findings have implications for both research and practice. Scholars should strengthen the links between AI methods, financial theory, and rigorous empirical validation, with particular attention to reproducibility, transaction costs, model drift, explainability, and robustness across market regimes. Practitioners and policymakers should view AI-enabled trading not only as a source of predictive and operational benefits but also as a source of model, governance, and market-stability risks.

The study is limited by its reliance on the Web of Science Core Collection, the incomplete 2025 observation period, time-sensitive citation counts, and export constraints that restricted co-citation and keyword analyses. Future research should use multiple databases, apply bibliographic coupling and full-text analysis, and examine whether the field's growing methodological coherence is accompanied by stronger interdisciplinary collaboration. Such work will help clarify how emerging AI architectures can be integrated into transparent, robust, and responsible financial decision systems.

Generative artificial intelligence tools were used solely for language support, including grammatical and stylistic revision, readability improvements, translation of selected passages, and assistance with manuscript organisation. They were not used to generate research data, conduct bibliometric analyses, select references, interpret findings, develop theoretical arguments, or formulate conclusions. All methodological decisions, data-collection procedures, analyses, interpretations, and conclusions were undertaken and validated by the authors, who assume full responsibility for the accuracy, originality, and integrity of the manuscript.