finance

This study systematically reviews RegTech applications in anti-money laundering and countering the financing of terrorism (AML/CFT) from 2020 to 2024, focusing on three core regulatory processes: customer due diligence, transaction monitoring, and regulatory reporting. Employing a literature analysis and technological integration framework, it examines how emerging technologies reshape AML/CFT compliance paradigms. The research proposes an innovative, synergistic architecture integrating artificial intelligence, blockchain, and big data analytics to enable real-time, intelligent, and cross-institutional/cross-border information sharing. Empirical evidence demonstrates that this integrated RegTech paradigm significantly enhances financial institutions’ capacity for autonomous financial crime risk identification, assessment, and mitigation. It simultaneously improves supervisory effectiveness and reduces compliance costs. The findings provide both theoretical grounding and actionable implementation pathways for building resilient, forward-looking AML/CFT frameworks aligned with evolving global regulatory expectations.

The financial domain lacks large-scale, open-source, semantically rich structured knowledge graphs (KGs), primarily due to the structural complexity of regulatory documents (e.g., annual reports) and stringent compliance requirements. Method: We introduce the first open-source financial KG derived from the latest SEC 10-K filings of S&P 100 companies. Our end-to-end pipeline features a table-aware chunking strategy and a schema-guided iterative extraction framework, augmented by a novel reflection-driven feedback mechanism and an LLM-as-a-Judge multi-dimensional evaluation system integrating intelligent parsing, rule-based validation, statistical verification, and large language model assessment. Contribution/Results: Experimental results demonstrate that our reflection-enabled agent achieves a compliance score of 64.8%, substantially outperforming single- and multi-step extraction baselines. It also attains state-of-the-art performance in precision, comprehensiveness, and relevance—marking significant advances in both accuracy and regulatory alignment for financial KG construction.

To address the limitations of large language models (LLMs) in regulatory compliance and numerical accuracy for Digital Regulatory Reporting (DRR), this paper introduces RKEFino1, a domain-specific financial language model. Methodologically: (1) it proposes the first numerical Named Entity Recognition (NER) task tailored to regulatory reporting, supporting dual-modality (sentence and table) financial entity identification; (2) it pioneers the systematic integration of multi-source structured regulatory knowledge—including XBRL, the Common Data Model (CDM), and the Meta-Object Facility (MOF)—into a lightweight financial foundation model, enhanced via domain-knowledge-informed fine-tuning, multi-task joint training, and regulatory ontology alignment. Experiments demonstrate that RKEFino1 significantly outperforms both general-purpose and state-of-the-art financial LLMs on key compliance tasks—including regulatory knowledge question answering, mathematical reasoning, and numerical NER—while exhibiting strong generalization capability. The model is publicly released on Hugging Face.

Financial foundation models (FFMs) face domain-specific challenges—including multimodal reasoning, regulatory compliance, and data privacy—that general-purpose foundation models cannot adequately address. To systematize FFM research, this work proposes the first tripartite taxonomy—FinLFM (language), FinTSFM (time-series), and FinVLFM (vision-language)—and comprehensively surveys advances in architecture design, training paradigms, datasets, and real-world deployment. We introduce a unified training framework integrating language modeling, time-series representation learning, and cross-modal alignment, leveraging heterogeneous financial data sources such as financial statements, market feeds, charts, and regulatory documents. Furthermore, we formally delineate the core capability boundaries distinguishing FFMs from general large language models. Key contributions include: (i) the first holistic survey of FFMs; (ii) Awesome-FinFMs—an open, dynamically updated resource repository; and (iii) a reproducible benchmark suite, a technical selection guide, and an open research roadmap.

This paper addresses the challenges of integrating and deeply understanding multimodal financial data—including textual, audio, visual, and video content, alongside market, macroeconomic, and alternative data. To this end, it proposes a novel paradigm: the Multimodal Financial Foundation Model (MFFM). Methodologically, it introduces the first systematic conceptual framework for MFFMs, transcending the linguistic limitations of existing Financial Large Language Models (FinLLMs) by unifying cross-modal alignment, heterogeneous representation fusion, and finance-domain adaptive pretraining and fine-tuning. Key contributions include: (1) establishing a comprehensive development roadmap for MFFMs; (2) open-sourcing the Awesome-MFFMs repository—a curated collection of models, datasets, and benchmarks; and (3) enabling complex financial applications such as intelligent investment research, risk management, and regulatory oversight, thereby advancing the field from unimodal analysis to cross-modal deep reasoning.

This work addresses the limited domain adaptability of large language models (LLMs) in financial natural language processing. We propose FinMA—a domain-specialized LLM built upon the PIXIU framework, integrating domain-adaptive pretraining with finance-oriented instruction tuning. To support this, we construct FIT, a high-quality financial instruction dataset, and conduct systematic evaluation using the FLARE benchmark. Experimental results demonstrate that FinMA significantly outperforms general-purpose baselines on financial sentiment analysis and classification tasks, validating the efficacy of domain-instruction fine-tuning. However, it exhibits notable performance bottlenecks on tasks demanding strong logical reasoning, fine-grained semantic understanding, or long-context processing—namely, numerical reasoning, financial named entity recognition, and long-document summarization. This study represents the first holistic effort to co-develop a financial instruction dataset, an adaptable training framework, and a dedicated evaluation benchmark. It provides a reproducible empirical foundation and methodological guidance for designing, evaluating, and analyzing the capability boundaries of financial LLMs.

This work addresses the limitations of existing large language models in the Indian financial ecosystem—particularly in domains such as digital payments, transaction disputes, and authorization management—and the absence of multilingual models supporting English, Hindi, and Hinglish. To bridge this gap, the authors present FiMI, the first large multilingual model tailored for Indian financial scenarios, built upon the Mistral Small 24B architecture. FiMI is developed through continued pretraining on 68 billion tokens of multilingual financial and synthetic data, followed by instruction tuning for multi-turn tool-augmented dialogues and domain-specific supervised fine-tuning. Experimental results demonstrate that FiMI Base outperforms baseline models by 20% on financial reasoning benchmarks, while FiMI Instruct achieves an 87% improvement in tool-calling tasks, all while maintaining competitive general-purpose capabilities.

This work addresses the challenge of evaluating large language model agents in finance, where existing methods fall short in assessing compliance and dynamic operational capabilities within realistic, executable tool environments. To bridge this gap, the authors introduce the first runnable benchmark grounded in real-world financial scenarios, integrating 760 executable financial APIs and 295 complex, tool-requiring queries. They propose a multidimensional evaluation framework that incorporates temporal relevance, intent categorization, and regulatory compliance. A key innovation is the Financial-Aware Tool Retrieval and Reasoning (FATR) mechanism, which enhances agent decision-making with domain-specific awareness. Additionally, the study releases the first open-source testing platform supporting auditable tool invocations, substantially improving the reliability of evaluations in terms of compliance, stability, and task completion quality.

Existing LLM-based trading agents predominantly rely on single-step prediction and lack explicit risk management mechanisms, leading to suboptimal performance under market volatility. To address this, we propose RISK-LLM—a novel framework featuring: (i) hierarchical market analysis to model multi-granularity dynamics; (ii) a dual-decision agent architecture that decouples signal generation from risk control; and (iii) multi-horizon reinforcement learning with risk-aware rewards, jointly optimizing returns and downside risk constraints (e.g., conditional value-at-risk). This work constitutes the first systematic integration of risk sensitivity into the LLM-driven trading paradigm. Extensive experiments across diverse markets—including A-shares and U.S. equities—demonstrate that RISK-LLM significantly improves the Sharpe ratio and enhances maximum drawdown control, outperforming state-of-the-art methods in both profitability and trading stability.

This study addresses the challenges of credit scoring in Kenya’s digital lending landscape, where data scarcity, institutional uncertainty, and pluralistic risk perceptions complicate algorithmic decision-making. Drawing on a nine-month ethnographic fieldwork in Nairobi, the research integrates alternative data construction, algorithmic modeling, and regulatory compliance strategies to examine how practitioners negotiate definitions of risk and model performance across technical and political dimensions. The project introduces “alignment” as a bidirectional translation process, elucidating the ongoing co-constitution of credit scoring models and their socio-institutional environments across three dimensions: cognition, modeling, and context. By extending alignment theory from human-computer interaction into the high-uncertainty domain of financial technology in the Global South, this work offers a novel perspective on algorithmic governance under conditions of epistemic and institutional ambiguity.

Current supervisory guidance SR 26-2 does not address generative AI and agent-based systems, creating a governance gap for financial institutions deploying these technologies in regulated processes—particularly concerning explainability, documentation, and compliance controls. This work proposes the first Generative AI Control Framework (GAICF) aligned with SR 26-2, extending model risk management principles to non-traditional AI systems through risk-tiered modeling and use-case mapping. Although designed for generative AI operating beyond formal model boundaries, the framework remains consistent with SR 26-2’s risk-based approach, offering financial institutions an actionable governance pathway to meet regulatory expectations when employing such systems in auxiliary tasks like policy analysis and regulatory interpretation.