Year-end tax strategies for property investors: accelerated depreciation claims, loss harvesting through asset disposals, and cross-border restructuring before new rules take effect.

Annual survey analysing run-off insurance portfolios globally. Covers capital management strategies, regulatory frameworks, and market consolidation trends affecting closed insurance books.

PwC analysis of global financial services M&A trends for 2025, covering digital transformation catalysts, regulatory environment impacts, and strategic consolidation patterns across banking, insurance and asset management sectors.

PwC analyses global real estate M&A trends for 2025, covering deal activity patterns, sector performance, and key factors driving property investment decisions across commercial and residential markets.

Annual PwC survey tracks crypto hedge fund industry evolution, covering assets under management, investment strategies, performance metrics, and regulatory developments across global markets.

Financial services M&A fell 15% by volume and 23% by value in H1 2024. Higher interest rates and regulatory scrutiny create challenging environment. Banks focus on domestic consolidation whilst insurers pursue cross-border deals.

Real estate M&A activity dropped 15% globally in H1 2024 compared to H1 2023. Elevated interest rates creating valuation mismatches between buyers and sellers. Opportunistic capital targeting distressed assets.

Global insurance run-off market exceeds $400bn in liabilities. Two-thirds of insurers actively manage legacy books (up from 45% in 2022). Key drivers include regulatory capital relief and operational cost reduction of 25-40%.

Financial services M&A activity remains subdued in 2024 due to elevated interest rates, increased regulatory oversight, and persistent buyer-seller valuation mismatches. Cross-border transactions declined 40% from 2021 levels.

PwC compares real estate transfer tax across 17 jurisdictions. Rates range from 0% (Jersey) to 15% (UK SDLT). Key rates: Belgium 10-12.5%, Germany 3.5-6.5%, France 5.8%, Netherlands 10.4%. Covers exemptions and cross-border structuring.

PwC analysis of international real estate investment structures, covering tax implications, legal frameworks, and regulatory requirements across global markets for institutional investors.

KPMG's risk and regulation resource centre covers regulatory developments across BFSI sectors including banking supervision, insurance regulation, and asset management compliance requirements.

Banks face new transformation challenges as traditional cost reduction methods plateau. Technology integration and operational redesign become key priorities for sustainable efficiency gains.

The FCA and PRA have introduced measures to boost the mutuals sector, such as a new Mutual Societies Development Unit and reduced application times for new societies. These initiatives aim to double the sector's size, affecting credit unions, building societies, and mutual insurers. This is crucial for enhancing financial inclusion and community support.

Non-bank financial institutions have grown significantly, driven by investment funds and hedge funds. Bond markets expanded due to surging government debt. NBFIs create uncertainty for monetary policy transmission.

The study finds that cybercrime legislation significantly enhances banking stability. This effect is more pronounced in countries with stringent penalties and robust enforcement mechanisms. The laws improve funding liquidity and mitigate operational risk, contributing to a secure banking environment. BFSI professionals should advocate for strong cybercrime laws to bolster stability.

Rory Milsom's study benchmarks large language model (LLM) agents for wealth-management tasks, revealing that workflow reliability and autonomy significantly impact performance. The research introduces a benchmark of 12 task-pairs for wealth management assistants, focusing on retrieval, analysis, and communication. It concludes that incorrect model evaluation has hindered effective benchmarking.

This paper explores the monopoly pricing of weather index insurance through a game-theoretic model involving insurers and farmers. It finds that using convolutional neural networks (CNNs) for modeling index payoffs reduces basis risk and increases insurer profits. The study also shows that expanding pricing flexibility leads to higher equilibrium profits.

ARCADIA, a new AI framework, improves causal discovery in corporate bankruptcy analysis by integrating large-language-model reasoning with statistical diagnostics. It outperforms traditional models such as NOTEARS and GOLEM by producing more reliable causal graphs. This advancement is crucial for sectors dealing with high-stakes financial decisions, offering a fully explainable and intervention-ready pipeline.

The study examines how global banks' net worth shocks influence emerging market economies (EMEs). Positive shocks lead to currency appreciation, reduced borrowing costs, and increased capital flows, boosting investment and credit. These effects are robust and vary by firm leverage and currency exposure, highlighting the importance for EME financial stability.

Teemu Pennanen and Waleed Taoum develop a statistical model for the SOFR term structure, suitable for risk management and derivatives pricing in incomplete markets. The model incorporates macroeconomic factors influencing central bank policy rates, allowing for calibration to historical data and market quotes. This approach aids in large-scale simulations for risk management and portfolio optimization.

The DeFi TrustBoost Framework integrates blockchain and Explainable AI to improve trust in decentralized finance, particularly for small business loans from low-wealth households. It emphasizes confidentiality, compliance with data protection laws, and resistance to adversarial attacks. The framework also supports tamper-proof auditing and data storage strategies for financial collaboration.

Real estate tokenisation emerges in underserved markets with lower prices and limited credit access. Trading volumes surge 35% over two days following natural disasters, but requires platform buyback features with higher insolvency risks.

Dyna.Ai promotes its agentic AI as a solution for financial institutions to achieve 'autonomous efficiency,' where AI handles routine tasks, allowing human focus on strategic growth. Only 24% of financial institutions see consistent AI returns, highlighting the need for strategic deployment. The agentic AI market is projected to grow significantly, impacting sectors like lending and fraud detection.

After pandemic lows, credit card and auto loan delinquencies have risen to pre-pandemic levels, with auto loan delinquencies notably increasing among lower-income households. The Federal Reserve's analysis indicates that fading fiscal support and increased borrowing costs are key factors. BFSI professionals should monitor these trends for risk management and lending strategies.

Matt DeLauro from SEON emphasizes the need for integrated fraud and AML operations to address digital financial crime. SEON's platform offers a unified approach, combining immediate protection rules with adaptive algorithms. This is crucial for APAC's unique challenges, such as synthetic identity fraud, and aligns with regulatory demands for transparency and speed.

The Federal Reserve study shows that mutual fund to ETF conversions, totaling $80 billion by 2024, enhance market liquidity and reduce stock volatility. A significant conversion event on June 11, 2021, involving $30 billion in equities, demonstrated these effects. This trend is driven by the tax efficiencies of ETFs.

DeepKnown-Guard presents a model-based safety framework for Large Language Models (LLMs), achieving a 99.3% risk recall rate and a perfect 100% safety score on proprietary high-risk tests. This framework integrates fine-tuned safety classification and Retrieval-Augmented Generation to ensure secure and traceable AI outputs, crucial for BFSI sectors relying on AI.

CompressionAttack identifies prompt compression in LLMs as a new security vulnerability. It uses HardCom and SoftCom strategies to manipulate LLM behavior, achieving up to 87% attack success rate. This poses significant risks for AI systems in BFSI sectors, necessitating stronger security measures.

Ant Group's Eric Jing highlighted AI's potential to serve as virtual CFO and COO for SMEs, addressing their lack of digital skills and manpower. Antom Copilot, part of Ant International, automates payment processes, reducing integration time by over 90% and improving chargeback rates. This initiative supports SMEs in global competition.

The study by Andrea Maurino highlights how data quality issues like missing values and noisy attributes affect machine learning models used in credit risk assessment. Using controlled data corruption, the research evaluates the robustness of models such as Random Forest and SVM. The findings underscore the importance of data quality in maintaining model accuracy, offering a framework for practitioners to enhance data pipeline robustness.

The study introduces a self-improving prompt framework that uses synthetic data to enhance the performance of large language models in financial document analysis. This method, tested on the DocMath-Eval benchmark, shows improved accuracy and robustness over traditional methods. It eliminates the need for external labels, making it cost-effective and adaptable.

The paper presents VerIH, a dataset designed to train language models to effectively prioritize instructions, improving their reliability and control. This approach enhances models' ability to handle conflicting instructions and increases robustness against security threats like jailbreak and prompt injection attacks. The findings are significant for developing more controllable language models.

The study reveals that Large Language Models (LLMs) face challenges in role-playing villainous characters due to their safety alignment, which conflicts with portraying morally ambiguous personas. This limitation is critical for BFSI sectors using AI for creative tasks, as it affects the authenticity of simulations involving negative scenarios.

The MERaLiON-SER model enhances speech emotion recognition by integrating weighted categorical cross-entropy and Concordance Correlation Coefficient losses. It excels in multilingual settings, surpassing open-source encoders and large Audio-LLMs. This model is pivotal for developing emotion-aware audio systems, offering improved paralinguistic understanding and cross-lingual generalisation.

FinRpt is a new benchmark designed to automate Equity Research Report (ERR) generation using a multi-agent framework and large language models (LLMs). It introduces a dataset construction pipeline integrating seven financial data types and an evaluation system with 11 metrics. The framework, FinRpt-Gen, shows strong performance in ERR generation, potentially transforming equity research.

Stella C. Dong's paper introduces a framework for evaluating large language models (LLMs) in reinsurance, focusing on governance, data lineage, and assurance. The framework aligns with Solvency II and other regulations, improving transparency and reducing informational frictions in risk transfer. This is crucial for reinsurance firms seeking to integrate AI responsibly.

SPEAR-MM, a framework for financial LLM adaptation, retains 91.2% of general reasoning capabilities compared to 69.7% with standard methods, while maintaining 94% of domain-specific gains. This approach significantly reduces computational costs by 90%, making it suitable for resource-constrained financial institutions. It offers interpretable trade-off control for model adaptation.

SWE-Compass provides a unified framework for evaluating LLMs in software engineering, addressing limitations of existing benchmarks. It covers 8 task types, 8 programming scenarios, and 10 languages, using 2000 instances from GitHub pull requests. This benchmark helps diagnose and advance LLMs' coding capabilities, aligning with real-world developer practices.

Cooperative Market Making (CMM) is a novel framework that distills large language model features into smaller models for market making in financial trading. By decoupling features across layer, task, and data dimensions, CMM enhances performance over traditional methods. BFSI professionals in financial services can leverage CMM to optimize trading strategies.

This systematic review examines the application of Explainable AI (XAI) in finance, focusing on post-hoc interpretability techniques such as SHAP and feature importance analysis. It identifies a need for multidisciplinary approaches that integrate financial expertise with advanced explainability methods. The findings are crucial for improving transparency in financial AI applications.

This study offers a cost-benefit analysis for organizations deciding between on-premise deployment of large language models (LLMs) and commercial LLM services. It evaluates hardware requirements, operational expenses, and performance benchmarks of open-source models like Qwen and Llama. The findings help determine the breakeven point for local deployment versus cloud services.

The AIA Forecaster, utilizing a large language model, achieves human superforecaster-level performance on the ForecastBench benchmark. It combines agentic search, a supervisor agent, and statistical calibration to improve forecasting accuracy. While it underperforms against market consensus on a new benchmark, it enhances predictions when combined with market data.

Research shows that LLMs can develop behavioral self-awareness, allowing them to predict their behaviors without supervision. This is achieved through a single rank-1 LoRA adapter and is domain-specific. The findings highlight potential safety issues in AI evaluation, relevant for AI governance and risk management in financial services.

The paper finds that smaller language models, such as Granite-3-8B and Qwen2.5-7B, achieve 100% output consistency in financial tasks, while larger models like GPT-OSS-120B show only 12.5% consistency. This challenges the belief that larger models are better for production. The study introduces a deterministic test harness and a three-tier model classification system for risk management.

The paper presents a five-layer architecture to improve the reliability of large language models (LLMs) in high-stakes decisions. It identifies cognitive biases as a challenge and proposes a structured calibration process to maintain effective human-AI partnerships. This approach is crucial for sectors relying on AI for strategic decisions, ensuring defensible valuations and sustainable investments.

OpsAgent is a novel multi-agent system designed to improve incident management in cloud systems by converting observability data into structured descriptions. It features a dual self-evolution mechanism for continuous improvement and has demonstrated superior performance on the OPENRCA benchmark. This system is particularly relevant for cloud service providers seeking efficient and sustainable incident management solutions.

'Learn-to-Ask' is a framework that enables large language models (LLMs) to act proactively by learning from offline expert data. This approach bypasses the need for complex user simulators and has been successfully deployed in a large-scale AI service, outperforming human experts in medical datasets. This method is particularly relevant for high-stakes domains.

'String Seed of Thought' (SSoT) is a novel prompting method that improves LLMs' performance in generating responses that align with target distributions. This method is particularly useful for applications requiring non-deterministic behaviors, such as human-behavior simulation and content diversification. SSoT enhances response diversity, crucial for test-time scaling.

NMIXX, a domain-adapted neural embedding model, enhances cross-lingual financial text analysis, particularly in low-resource languages like Korean. It outperforms existing models on Korean financial benchmarks by significant margins, demonstrating the importance of domain-specific adaptation. BFSI professionals can leverage these tools for improved multilingual financial analysis.

This study introduces a calibration method for process reward models (PRMs) using quantile regression to improve accuracy in large language models (LLMs). The calibrated PRMs enable an instance-adaptive scaling (IAS) framework that adjusts compute budgets dynamically, reducing inference costs without compromising accuracy. This approach is particularly relevant for mathematical reasoning tasks.

The study finds that enhancing the quality of chosen responses significantly boosts DPO performance, while the quality of rejected responses is less impactful. This insight is critical for constructing effective preference datasets for large language model alignment. BFSI professionals should consider these findings when developing AI models for customer interaction.

XBreaking reveals how targeted noise injection can break the security alignment of Large Language Models (LLMs). This method identifies exploitable patterns in LLMs, posing risks for sectors relying on AI for secure operations. BFSI professionals should assess AI model vulnerabilities to safeguard against potential breaches.

KoBankIR, a new domain-specific information retrieval benchmark for banking, is introduced using a novel query generation pipeline. This pipeline employs large language models (LLMs) and enhanced answerability assessment to align closely with human judgments. The study reveals that current retrieval models face challenges with KoBankIR's complex queries, indicating a need for improved techniques in financial domains.

The Nemotron Nano V2 VL by NVIDIA offers significant improvements in document understanding and video comprehension. It builds on the Nemotron Nano V2 with a hybrid Mamba-Transformer LLM and innovative token reduction techniques. The model is available in BF16, FP8, and FP4 formats, with datasets and training code shared for broader use.

AIRepr introduces a framework to evaluate and improve the reproducibility of workflows generated by large language models (LLMs) in data science. By benchmarking reproducibility-enhancing prompts against standard ones, the study found improved accuracy and reproducibility in analyses. This framework supports scalable, automated assessment, crucial for reliable human-AI collaboration.

TimeCopilot, introduced by Azul Garza and Renée Rosillo, is an open-source framework that combines Time Series Foundation Models with Large Language Models to automate the forecasting process. It supports feature analysis, model selection, and forecast generation, providing natural language explanations. The framework achieved state-of-the-art results on the GIFT-Eval benchmark, offering a cost-effective solution for reproducible and explainable forecasting.

Agent Spec introduces a standardized language for AI agents, promoting interoperability across frameworks by defining common components and evaluation methods. This standardization is crucial for consistent performance assessment and system integration, particularly in AI-driven financial services where agent reliability and efficiency are paramount.

This study introduces a semi-automated approach to generate Software Architecture Descriptions (SADs) from source code by combining reverse engineering with large language models. The method recovers static and behavioral architectural views, offering a scalable alternative to manual documentation. This enhances system understanding and maintainability, particularly relevant for software engineering.

Conformal Information Pursuit (C-IP) is proposed to enhance the efficiency of Large Language Models (LLMs) in interactive question-answering tasks. By using conformal prediction sets, C-IP provides a robust method for estimating uncertainty, leading to better predictive performance and shorter query chains compared to traditional Information Pursuit (IP) methods. This approach is particularly beneficial in interactive settings, such as medical consultations, where it offers improved interpretability and competitive performance.

TAMAS provides a benchmark for assessing adversarial risks in multi-agent LLM systems, revealing their vulnerability to attacks. The study evaluates ten LLMs across various scenarios, introducing the Effective Robustness Score (ERS) to measure safety versus task effectiveness. This research underscores the need for stronger defenses in multi-agent systems, crucial for BFSI sectors relying on AI collaboration.

AnytimeReasoner enhances the reasoning capabilities of large language models by optimizing performance under variable token budgets. It uses Budget Relative Policy Optimization (BRPO) to improve token efficiency and robustness in training. This approach outperforms existing methods in mathematical reasoning tasks, offering significant improvements in training and deployment efficiency.

Researchers have developed automated methods to extract court-established facts from Slovak criminal court opinions. Using advanced regular expressions and large language models, they achieved extraction accuracy rates of 97% and 98.75%, respectively, and 99.5% when combined. This advancement could significantly enhance data availability for legal and financial analysis.

The AI agent Solly, developed using self-play and deep reinforcement learning, has reached elite human performance in Liar's Poker, a game requiring multi-player engagement and strategic bidding. Solly's win rate exceeded 50%, and it outperformed large language models in both heads-up and multi-player formats, demonstrating advanced strategic capabilities.

This survey explores cognitive edge computing as a method to deploy large AI models on resource-constrained devices. It discusses model optimization techniques like quantization and distillation, system architecture strategies such as on-device inference, and adaptive intelligence methods. The paper outlines a standardized evaluation protocol and identifies challenges like energy reporting and safety evaluation.

The paper investigates whether LLMs exhibit human-like cognitive behaviors using frameworks like the Thematic Apperception Test and Moral Foundations Theory. Findings show LLMs create coherent narratives, are influenced by framing, and make moral judgments. These insights are crucial for AI transparency and ethical deployment, impacting AI safety and development.

BudgetMem introduces a memory-efficient architecture for processing long contexts in language models, achieving a 72.4% reduction in memory usage with minimal F1 score degradation (1.0%). This approach is crucial for deploying advanced language models on resource-constrained hardware, particularly benefiting sectors like financial services that require processing extensive documents.

The iTool approach enhances large language models (LLMs) by using reinforced fine-tuning with dynamic deficiency calibration, achieving a 13.11% performance improvement over base models. It addresses the limitations of synthetic data in complex scenarios, outperforming larger models by 6.5%. This method is crucial for advancing AI capabilities in complex tool-use tasks.

LiveStar, a new live streaming assistant, enhances online video understanding by achieving a 19.5% improvement in semantic correctness and a 12.0% increase in frames per second (FPS). It uses adaptive streaming decoding and memory-aware acceleration, making it relevant for sectors using real-time video data, such as financial services monitoring.

Adaptive Branching Monte Carlo Tree Search (AB-MCTS) is proposed to improve large language model (LLM) inference by using external feedback to decide whether to expand or revisit candidate responses. This method outperforms traditional sampling and standard MCTS, highlighting the importance of response diversity and multi-turn refinement. This advancement is crucial for complex coding and engineering tasks.

The study demonstrates that large language models (LLMs), when trained with uncertainty-aware objectives, can effectively optimize experimental processes. This approach nearly doubles the discovery rate of high-yielding conditions in pharmaceutical synthesis from 24% to 43%. It highlights the potential of LLMs to replace domain-specific feature engineering with natural language interfaces, enhancing reliability through uncertainty quantification.

MAC-Flow is a new framework for multi-agent coordination that balances performance and computational efficiency. It achieves 14.5 times faster inference compared to diffusion-based methods, with performance comparable to Gaussian policy-based methods. This advancement is crucial for sectors relying on real-time multi-agent systems, such as robotics and AI-driven financial services.

Enterprise Deep Research (EDR) introduces a multi-agent system for enterprise analytics, featuring adaptive query decomposition and specialized search agents. It excels in automated report generation and real-time streaming, validated on internal datasets and benchmarks like DeepResearch Bench. EDR's performance surpasses existing systems without human steering, offering significant potential for BFSI sectors.

The TeaRAG framework enhances Retrieval-Augmented Generation by compressing retrieval content and reasoning steps, achieving a 4% increase in Exact Match and reducing token output by 59% on specific datasets. This efficiency is crucial for applications requiring high accuracy with limited computational resources.

The paper presents DeepForgeSeal, a novel framework for deepfake detection using latent space-driven semi-fragile watermarking. It employs Multi-Agent Adversarial Reinforcement Learning to balance robustness and sensitivity in watermarking. The method shows significant improvements over current approaches, with over 4.5% better performance on CelebA and 5.3% on CelebA-HQ benchmarks.

The authors propose a framework that constructs a verifiable Long Chain-of-Thought (LCoT) knowledge base, leading to the creation of SciencePedia. This approach enhances scientific reasoning by providing explicit derivational chains, improving transparency and accuracy. The framework uses a Socratic agent to generate questions and multiple models to verify answers, resulting in a high-fidelity knowledge base.

The Pluralistic Behavior Suite (PBSUITE) evaluates large language models (LLMs) on their ability to adhere to custom behavioral policies in multi-turn interactions. While LLMs show less than 4% failure in single-turn settings, their compliance drops to 84% in adversarial multi-turn scenarios. This highlights the need for improved alignment techniques in real-world applications.

This research investigates the role of imagination in cognitive agents, comparing human and LLM imagination networks. It finds that human networks show strong correlations between centrality measures, unlike LLMs, which lack clustering. This suggests significant differences in internal world models, impacting AI development towards human-like imagination.

TOBUGraph, a knowledge graph-based retrieval framework, improves LLM performance by capturing deeper semantic relationships beyond RAG's text-to-text similarity. It enhances retrieval accuracy and reduces hallucinations, as demonstrated in a real-world application, TOBU. This advancement is significant for improving user experience in data retrieval tasks.

BBVA has integrated AI across its operations, reaching 11,000 employees and saving approximately 3 hours per employee each week. The bank reports efficiency improvements of up to 80% in certain workflows. This initiative is supported by a structured learning environment and governance, reducing shadow AI risks.

This study challenges the assumption that first layers in language models are best for influence estimation. It finds middle attention layers offer better insights. The research introduces new aggregation methods, such as ranking and vote-based approaches, which enhance performance. A new metric, Noise Detection Rate, is proposed for evaluating influence score efficacy without retraining models.

The Chain-of-Query (CoQ) framework addresses challenges in SQL-aided table understanding by employing a multi-agent approach. It uses natural-language-style table schema representations and a clause-by-clause SQL generation strategy to improve query quality. Experiments show CoQ significantly enhances accuracy and reduces invalid SQL rates compared to existing methods.

The paper introduces AVIATOR, an AI-agentic workflow for injecting realistic software vulnerabilities. It achieves 91%-95% success rates, outperforming current methods in accuracy and scope. This advancement is crucial for improving automated vulnerability detection and repair systems, which are vital for cybersecurity in BFSI sectors.

HugAgent is a new benchmark designed to evaluate large language models (LLMs) on their ability to simulate individualized human reasoning. It shifts focus from population-level consensus to individual reasoning styles. The benchmark uses a dual-track design for collecting human reasoning data and stress testing models. This is crucial for AI development in personalized applications.

The Trustworthiness Calibration Framework (TCF) offers a methodology to assess phishing email detectors using large language models (LLMs) like GPT-4, LLaMA-3-8B, and DeBERTa-v3-base. Experiments show GPT-4 has the strongest trust profile. This framework aids in evaluating model dependability beyond accuracy, crucial for real-world deployment.

The paper introduces Activation-Informed Merging (AIM), a technique that improves the merging of large language models by incorporating activation-space information. AIM enhances model performance by up to 40% across benchmarks, offering a flexible solution applicable to existing methods. This advancement is significant for computational efficiency in AI model development.

ProRefine, an inference-time optimization method, refines AI prompts using textual feedback, improving accuracy by 3-37% on mathematical reasoning tasks. This method allows smaller AI models to match the performance of larger ones, making AI systems more cost-effective and accessible. It is particularly relevant for AI-driven financial services requiring complex reasoning.

The McKinsey Global Survey 2025 reveals that while 88% of organizations use AI in at least one business function, only 39% report enterprise-level EBIT impact. High performers, representing 6% of respondents, achieve significant benefits by focusing on growth and innovation, redesigning workflows, and scaling AI use across multiple functions.

Boston Consulting Group reports that retail banks can enhance profits by $370 billion by 2030 through AI. This potential stems from AI's ability to improve efficiency, customer experience, and risk management. Retail banks should invest in AI technologies to capitalize on these opportunities and remain competitive.

The research by Karadal and Kekulluoglu explores how ChatGPT's responses vary based on inferred political orientations. Using personas with distinct political views, the study finds that ChatGPT's outputs align with these inferred views, indicating potential bias. This has implications for AI ethics and user interaction, particularly in politically sensitive contexts.

POLIS-Bench offers a new evaluation suite for large language models (LLMs) in bilingual governmental policy scenarios. It includes an up-to-date bilingual corpus, scenario-grounded tasks, and a dual-metric evaluation framework. The study reveals a performance hierarchy among models, with reasoning models excelling in stability and accuracy, especially in compliance tasks.

The study reviews 445 LLM benchmarks, identifying issues in measuring phenomena like safety and robustness. It emphasizes the need for strong construct validity to ensure reliable evaluations. The authors provide eight recommendations to enhance benchmark development, crucial for AI and NLP researchers.

The Financial Stability Board (FSB) has announced its 2025 list of Global Systemically Important Banks (G-SIBs). This list is vital for regulatory compliance, as it determines additional capital requirements and oversight for these banks. BFSI professionals, particularly in corporate banking, should review the list to understand potential impacts on risk management and regulatory strategies.

The Financial Stability Board's 2025 list of Global Systemically Important Banks (G-SIBs) identifies banks whose failure could significantly impact the global financial system. This list is crucial for regulators and financial institutions to ensure adequate capital buffers and risk management practices.

The ECB's Financial Stability Review highlights systemic risks from euro area banks' connections with non-bank financial intermediaries (NBFIs). Banks' short-term liabilities from NBFIs are prone to flight risk, while credit provision to leveraged NBFIs could lead to asset price shocks. This is critical for global systemically important banks (G-SIBs) that transform these liabilities into credit. Professionals should monitor these linkages to mitigate potential financial instability.

The 2025 EU-wide stress test integrates contagion risk, showing a 29 basis point average CET1 ratio depletion, with first-round effects increasing by 12%. Investment funds, especially equity funds, face the greatest losses. This system-wide approach underscores the need to consider spillover effects across financial sectors, particularly affecting banks with less sophisticated hedging strategies.

WinnowRAG, a new framework for retrieval-augmented generation, improves accuracy by filtering out irrelevant documents. It uses query-aware clustering and a critic LLM to evaluate and retain useful documents. This model-agnostic approach is adaptable to various tasks without requiring fine-tuning, showing effectiveness over current baselines.

Researchers developed an agent-based simulation using Large Language Models to study misinformation vulnerabilities. They found that mental schemas significantly influence how misinformation is interpreted, more so than professional backgrounds. This validates the use of LLMs as proxies in studying information responses, offering insights into trust and susceptibility in social systems.

The EncouRAGe framework facilitates the evaluation of Retrieval-Augmented Generation (RAG) systems, emphasizing scientific reproducibility and local deployment. It reveals that RAG systems underperform compared to Oracle Context, while Hybrid BM25 achieves the best results across datasets. BFSI professionals can leverage these insights to enhance AI-driven customer service tools.

JPMorgan Chase, under Derek Waldron's leadership, is transforming into an AI-first bank with its LLM Suite platform. This initiative, supported by an $18 billion annual tech budget, empowers employees with AI tools, fostering innovation and productivity. The platform's adoption has been significant, with nearly half of the bank's employees using AI tools daily, leading to a cultural shift within the organization.

The paper introduces ATLAS, a framework using Item Response Theory for evaluating large language models (LLMs). It reduces the number of evaluation items by 90% without losing precision, as demonstrated on the HellaSwag benchmark. This approach reveals discrepancies in model rankings compared to static benchmarks, impacting model evaluation practices.

The Office of the Comptroller of the Currency (OCC) and the Federal Deposit Insurance Corporation (FDIC) have retracted their guidance on leveraged lending. This decision was made due to concerns that the guidance was overly restrictive and could hinder lending activities. This change may impact corporate banking sectors involved in leveraged lending.

The publication discusses how real-time payment systems are transforming the payments landscape and influencing the future of the financial services industry. It likely covers drivers such as instant settlement, customer expectations, competitive dynamics, and emerging infrastructures, and how banks and payment providers must adapt their strategies, technology, and operations to stay relevant.

The study finds that banks' regulatory risk tolerance (RRT) increased after the Great Financial Crisis (GFC), driven by reduced management buffer targets. Banks with volatile buffer shocks set higher targets and reversion speeds, indicating strategic choices. High-RRT banks tend to cut lending when management buffers deplete, affecting the real economy.

Tokenisation, driven by distributed ledger technology, is growing but remains small-scale. It offers potential benefits such as efficiency and transparency but also poses risks like liquidity mismatches and operational fragilities. The Financial Stability Board highlights vulnerabilities such as leverage and interconnectedness, suggesting regulatory improvements to mitigate these risks.

The 2024 BIS survey indicates that 91% of surveyed central banks are exploring central bank digital currencies (CBDCs), with a stronger focus on wholesale CBDCs. The decline of cash and rise of tokenisation are key drivers. Over one-third of jurisdictions have accelerated CBDC work due to stablecoin developments.

Agentic AI represents a shift in enterprise automation, allowing AI agents to perform complex tasks beyond information retrieval. Azure AI Foundry offers a comprehensive platform for developing these agents, supporting tool use, reflection, planning, and collaboration. BFSI sectors can leverage these capabilities to enhance process efficiency and reduce operational costs.

RegTech solutions automate compliance and risk management for financial institutions, addressing complex regulatory requirements. In 2024, banks faced $19.3 billion in penalties, underscoring the importance of RegTech. Key growth drivers include complex regulations, digitalization, and cloud adoption. Financial institutions should focus on data quality and consider specialized RegTech solutions.

Bank resolution aims to maintain financial stability by preserving a failing bank's critical functions. Public support should be credibly funded and used only when systemic risk is present. Jurisdictions need policies to recoup public support and governance to ensure social legitimacy. Flexibility in providing support is crucial to manage systemic risks effectively.

The $2.4 trillion global payments industry is growing at 5% annually, but faces rising fraud risks, including deepfakes and synthetic ID fraud. Online payment fraud could surpass $362 billion by 2028. Providers must adopt advanced technologies to detect fraud while maintaining customer satisfaction. This balance is crucial to prevent revenue loss and maintain trust.

The new Monetary Policy Conditions Index (MCI) combines short-term interest rates and central bank balance sheet sizes into a single measure. This index captures the effects of balance sheet policies even when the effective lower bound is not binding. The MCI's impact on output, inflation, and financial conditions is analyzed using a Bayesian vector autoregression framework, offering insights into the effectiveness and unintended consequences of unconventional monetary policies.

AI is reshaping the insurance industry by enhancing sales, underwriting, claims management, and customer service. Insurers that fully integrate AI achieve significantly higher TSR, with leaders outperforming laggards by 6.1 times. To succeed, insurers must adopt a comprehensive AI strategy, rewire operations, and overcome barriers like talent shortages and legacy systems.

The study explores the impact of retail CBDC on the banking system, highlighting potential risks of disintermediation. Survey data from German households suggests a shift from bank deposits to CBDC, raising financial stability concerns. Implementing a holding limit on CBDC could enhance stability by reducing the risk of bank runs.

The study examines central bank and media sentiments on digital currencies across 15 major economies from 2016 to 2022. It finds that central bank sentiment significantly influences media sentiment and impacts financial markets. Positive central bank sentiment correlates with negative cryptocurrency returns and affects banking and payment-related stocks.

The BIS paper suggests a methodology to incorporate physical climate risks into banks' credit risk models, using an extension of the Vasicek model. This model treats physical risk as a stochastic factor affecting asset values, aiding in economic capital allocation and loan loss provisioning. It preserves portfolio invariance, making it suitable for regulatory purposes.

To succeed, new CROs should quickly establish a clear risk strategy aligned with the organization's goals and foster a strong risk culture. Building relationships with the board and supervisors is crucial. These steps help CROs optimize their impact and navigate early challenges effectively, ensuring long-term success.

Natural Language Reinforcement Learning (NLRL) uses Language Value Functions (LVF) to replace scalar values with interpretable linguistic evaluations. Tests show 85% accuracy vs 61% for baseline LLMs in breakthrough game, plus stable 0.9 win rates in tic-tac-toe vs random opponents.

Financial firms must check customer identity using clean ID copies without watermarks or blackouts. Remote identification requires photo/video selfies. Documents retained 5 years after customer relationship ends under anti-money laundering laws.

A 2024 study by De Nederlandsche Bank found that 15 of 36 insurers use AI, primarily for fraud detection and risk assessment. The AI Act, effective August 2024, provides a regulatory framework. Insurers see AI as a tool to enhance efficiency and customer experience, but face challenges in data quality and internal expertise.

Hebbia's Matrix, leveraging OpenAI's models, automates 90% of finance and legal workflows, achieving 92% accuracy in complex document processing. This platform uses multiple AI agents to handle tasks traditionally requiring extensive human effort, offering significant efficiency gains for financial and legal professionals.

The study identifies key challenges in multi-agent LLM systems, such as coordination failures and communication inefficiencies. These issues hinder the effective deployment of LLMs in complex environments. The findings are crucial for BFSI sectors relying on AI for decision-making, as they highlight the need for improved system designs to enhance reliability and performance.

Nubank, a major digital financial services platform, leverages OpenAI's technology to improve customer service and internal operations. Key initiatives include an AI-powered assistant resolving 55% of Tier 1 inquiries and a custom enterprise search tool used by over 5,000 employees. These solutions enhance customer satisfaction and operational efficiency.

SoFi, a digital banking platform, intends to raise $1.5 billion to support its business expansion. This move highlights SoFi's growth strategy in the competitive digital banking sector. The capital will likely be used to enhance its product offerings and market reach, impacting the retail banking landscape.

TMMFs circulate on public blockchains like stablecoins but offer money market rates and securities regulation. They use wallet allow-listing for compliance while serving DeFi protocols as collateral and savings instruments.

BIS study finds ChatGPT can perform high-level intraday liquidity management in wholesale payment systems. AI agent maintained precautionary buffers, prioritized urgent payments, and optimized settlement speed vs liquidity trade-offs across simulated scenarios.

Agentic AI represents the next frontier in artificial intelligence, capable of autonomously solving complex, multi-step problems through sophisticated reasoning and iterative planning. This technology is transforming industries by personalizing customer service, streamlining software development, and facilitating patient interactions. BFSI professionals should explore agentic AI to enhance customer service and automate routine tasks.

OpenAI has established a $4 billion revolving credit facility with major banks including JPMorgan Chase and Citi, adding to $6.6 billion in new funding. This provides OpenAI with over $10 billion in liquidity, enabling investment in AI research and infrastructure expansion. The facility underscores strong partnerships with leading financial institutions.

In 2024, data science will be driven by advancements in AI, machine learning, and big data, with significant implications for careers in these fields. Professionals will need to focus on AI ethics, real-time analytics, and interdisciplinary skills. Enhanced data privacy and democratization of data science tools will also be crucial.

AI is increasingly adopted, with 75% of organizations using it, though only 8% apply it extensively. This indicates early integration stages, affecting productivity and monetary policy. The ECB's Information Systems department is its largest, highlighting AI's role in economic modeling and policy implementation. AI's potential parallels historical tech shifts, like ICT.

Financial services are undergoing significant changes driven by digital transformation, data analytics, and regulatory compliance. Fintechs are challenging traditional banks, while open banking and AI are creating new opportunities. Compliance with regulations such as PSD2 and ESG is essential. Cybersecurity remains a top priority due to rising cybercrime.

AI's resurgence, driven by large language models like ChatGPT, now assists in 18% of computer science papers, up from 2% pre-2022. Quantum computing, despite its challenges, is poised to revolutionize fields like data science and AI. Zero Trust security models are becoming industry standards, with companies like Cisco investing heavily.

Digital transformation involves integrating digital technology across business areas, fundamentally changing operations and customer value delivery. Key components include digital strategy, data analytics, and customer-centricity. For BFSI, it enhances customer experience and operational efficiency, crucial for maintaining competitiveness. Professionals should focus on strategy and technology investments to drive transformation.

Deloitte's report, 'Agentic Enterprise 2028', highlights the transformative potential of Agentic AI in achieving cost savings, job creation, and accelerated growth by 2028. The BFSI sector, particularly retail banking and asset management, can benefit from AI-driven efficiencies and innovation. BFSI professionals should explore AI integration strategies to enhance operational efficiency and customer engagement.

The Financial Conduct Authority (FCA) has outlined plans to bolster the mutuals sector, which includes credit unions and building societies. This initiative aims to enhance market integrity and promote competition, aligning with the FCA's broader strategy for 2025-2030. The move is expected to benefit consumers by providing more diverse financial options.

The Financial Conduct Authority (FCA) is seeking feedback on proposals to reduce late fees for regulatory returns. This move is intended to alleviate financial pressures on firms, encouraging timely compliance and supporting market integrity. The consultation reflects the FCA's ongoing commitment to promoting competition and protecting consumers.

Scammers are increasingly creating fake bank websites and apps to steal personal information. The FDIC advises consumers to verify the legitimacy of banks using the FDIC BankFind tool. This is crucial for protecting personal data and financial security. BFSI professionals should educate customers on identifying phishing, smishing, and vishing scams.

The ISDA, SIFMA, and IIB have provided feedback to the SEC on proposed regulations for security-based swaps. These organizations aim to influence regulatory frameworks to ensure they align with industry practices and promote market stability. This is crucial for derivatives markets, impacting risk management and compliance strategies.

By 2030, AI will revolutionize the insurance industry, enabling real-time policy issuance and claims processing. Insurers must adapt to AI-driven changes in distribution, underwriting, and pricing. The integration of AI will lead to new product categories and personalized pricing, enhancing customer experience and operational efficiency.

FinCEN's ruling states that banks cannot be designated as providers of prepaid access under MSB regulations. This affects companies distributing bank-issued prepaid cards, as they may not need to register as MSBs if the bank has primary oversight. This clarification impacts compliance requirements for financial institutions involved in prepaid card programs.