Insights & Announcements
This study by Chang Liu analyzes over 35,000 documents from 1,125 hedge fund managers using topic modeling and sentiment analysis. Latent Dirichlet Allocation (LDA) with 20 topics yields the most interpretable and robust topic assignments, while Top2Vec achieves superior classification performance. Sentiment analysis reveals that DistilBERT outperforms FinBERT in generating sentiment scores for hedge fund documents. Sentiment scores produced by DistilBERT combined with Top2Vec show stronger correlations with subsequent fund performance than other model combinations. The research demonstrates that automated text analysis can provide investors with systematic investment signals and data-driven decision support tools. 🔗 Source: View Source | Found on Dec 10, 2025
This study analyzes market reactions and information spillovers from two major Japanese bank mergers: the 2005 creation of Mitsubishi UFJ Financial Group after the Financial Big Bang, and the 2018 merger involving Resona Holdings following the global financial crisis. Using event studies with market model, CAPM, and Fama-French three-factor model to estimate cumulative abnormal returns, along with VAR models for Granger causality and impulse response functions, and propensity score matching for causal effects, the research finds significant positive market reactions to both mergers and prolonged positive spillovers to other banks, suggesting synergistic effects within Japan’s banking sector.
🔗 Source: View Source | Found on Dec 10, 2025
The article introduces a new paradigm for constructing large causal models (LCMs) using large language models (LLMs), detailing ongoing experiments with the DEMOCRITUS system. DEMOCRITUS builds, organizes, and visualizes LCMs across diverse domains by extracting causal relations from targeted textual queries to LLMs. Unlike traditional methods relying on numerical data from experiments, this approach uses LLMs to propose topics, generate causal questions, and extract statements. The system addresses challenges in integrating fragmented and conflicting claims into relational causal triples within an LCM through novel categorical machine learning methods. Results span archaeology, biology, climate change, economics, medicine, and technology.
🔗 Source: View Source | Found on Dec 10, 2025
The article by Brian Ezinwoke and Oliver Rhodes investigates the use of Spiking Neural Networks (SNNs) for forecasting price movements in high-frequency financial data, addressing limitations of conventional models in capturing fine temporal structures. The study evaluates three SNN architectures—an unsupervised STDP-trained SNN, a novel SNN with explicit inhibitory competition, and a supervised backpropagation network—using high-frequency stock data converted into spike trains. Hyperparameter tuning was performed via Bayesian Optimization driven by Penalized Spike Accuracy (PSA), ensuring predicted spike rates matched empirical rates. The extended SNN model with PSA achieved the highest cumulative return of 76.8% in backtesting, outperforming the supervised alternative’s 42.54%.
🔗 Source: View Source | Found on Dec 08, 2025
The study analyzes hundreds of millions of anonymized user interactions with Comet, an AI-powered browser by Perplexity, and its integrated Comet Assistant. Adoption and usage are higher among earlier adopters, users in countries with greater GDP per capita and educational attainment, and those in digital or knowledge-intensive sectors such as technology, academia, finance, marketing, and entrepreneurship. Productivity & Workflow and Learning & Research comprise 57% of agentic queries; Courses and Shopping for Goods make up 22%. The top 10 out of 90 tasks account for 55% of queries. Personal use represents 55%, professional use 30%, and educational contexts 16%.
🔗 Source: View Source | Found on Dec 10, 2025
The article "Architectures for Building Agentic AI" by SÅ‚awomir Nowaczyk, submitted on 10 December 2025, defines agentic systems as goal-directed, tool-using decision makers operating in closed loops. It argues that reliability in agentic and generative AI is primarily an architectural property, emerging from principled componentisation—such as goal manager, planner, tool-router, executor, memory, verifiers, safety monitor, and telemetry—and disciplined interfaces like schema-constrained and validated tool calls. The chapter presents a taxonomy of agent types and provides design guidance on typed schemas, idempotency, permissioning, transactional semantics, memory hygiene and provenance, runtime governance including budgets and termination conditions, and simulate-before-actuate safeguards.
🔗 Source: View Source | Found on Dec 12, 2025
The article by Daniel Egger and Jacob Vestal, presents a new application of Hoeffding's Inequality for financial trading. Hoeffding's Inequality quantifies the maximum probability that n draws from a bounded random variable differ from its true expectation u by more than a specified tolerance t. The authors propose using deviations in trading strategy performance as an indicator of financial regime change; larger deviations suggest a lower probability that the current market regime persists. Changing Hoeffding probabilities can thus serve as an early warning signal for traders regarding potential shifts in financial regimes.
🔗 Source: View Source | Found on Dec 11, 2025
The article by Chuanhao Nie, Yunbo Liu, and Chao Wang reviews the application of artificial intelligence (AI) in anti-money laundering (AML), emphasizing its role in improving detection accuracy, lowering false-positive rates, and reducing manual investigation burdens. It discusses future research directions such as federated learning for privacy-preserving collaboration, fairness-aware and interpretable AI, reinforcement learning for adaptive defenses, and human-in-the-loop visualization systems. The paper proposes an AI-driven KYC solution using graph-based retrieval-augmented generation (RAG Graph) with generative models to enhance efficiency and transparency in KYC CDD/EDD workflows, supported by experimental results demonstrating high faithfulness and answer relevancy.
🔗 Source: View Source | Found on Dec 10, 2025
The article, "Can AI autonomously build, operate, and use the entire data stack?" by Arvind Agarwal, Lisa Amini, Sameep Mehta, Horst Samulowitz, and Kavitha Srinivas (submitted on 8 Dec 2025), discusses the challenges of enterprise data management across architecture, integration, quality, governance, and improvement. The authors argue for a paradigm shift from AI assisting specific roles to fully autonomous management of the entire data lifecycle by intelligent agents. They examine how each stage of the modern data stack could be autonomously managed and highlight open questions requiring further research to achieve self-sufficient systems usable by both humans and AI.
🔗 Source: View Source | Found on Dec 11, 2025
Accenture and Anthropic have announced an expanded partnership, forming the Accenture Anthropic Business Group to accelerate enterprise AI adoption. Approximately 30,000 Accenture professionals will be trained on Claude, creating one of the largest ecosystems of Claude practitioners globally. Accenture becomes a premier partner for coding with Claude Code, which now holds over half of the AI coding market. The partnership includes joint offerings for CIOs to measure value and drive large-scale AI adoption, with initial solutions targeting regulated industries such as financial services, life sciences, healthcare, and public sector. Anthropic’s enterprise market share has grown from 24% to 40%.
🔗 Source: View Source | Found on Dec 09, 2025
In 2026, AI is transitioning from a tool to a collaborative partner across industries, amplifying human expertise and transforming work processes. Aparna Chennapragada of Microsoft highlights AI agents as digital coworkers enabling small teams to launch global campaigns rapidly. Vasu Jakkal emphasizes the need for robust security measures for AI agents as they become integral to daily tasks. In healthcare, Dr. Dominic King notes Microsoft AI’s Diagnostic Orchestrator achieved 85.5% accuracy in solving complex cases in 2025, with generative AI products reaching millions. Mark Russinovich predicts denser, more efficient global AI infrastructure, while Mario Rodriguez reports GitHub’s monthly pull requests rose 23% to 43 million in 2025.
🔗 Source: View Source | Found on Dec 08, 2025
Anthropic has donated the Model Context Protocol (MCP) to the Agentic AI Foundation (AAIF), a directed fund under the Linux Foundation co-founded by Anthropic, Block, and OpenAI, with support from Google, Microsoft, AWS, Cloudflare, and Bloomberg. MCP has over 10,000 active public servers and is adopted by platforms such as ChatGPT, Cursor, Gemini, Microsoft Copilot, and Visual Studio Code. Enterprise-grade infrastructure supports MCP via AWS, Cloudflare, Google Cloud, and Microsoft Azure. MCP features include asynchronous operations and statelessness; its SDKs have 97M+ monthly downloads across Python and TypeScript.
🔗 Source: View Source | Found on Dec 09, 2025
The article presents the Tuning Config Recommender, developed after analyzing thousands of model tuning jobs and observing common user challenges such as CUDA out-of-memory errors, incorrect data configurations, and missed kernel optimizations. Integrated into the Foundation Model Stack (FMS) ecosystem, the tool uses rule-based flexibility and a knowledge-driven approach to generate optimal configurations with minimal user input. It processes inputs via an Intermediate Representation (IR), applies rules to produce JSON patches, and synthesizes runnable commands for fine-tuning stacks like fms-hf-tuning. The recommender aims to simplify configuration, improve efficiency, and reduce trial-and-error cycles for users.
🔗 Source: View Source | Found on Dec 08, 2025
Throughout 2025, MAI analyzed a sample of 37.5 million de-identified Copilot conversations to understand real-world usage patterns. The study found that health-related topics consistently dominated interactions on mobile devices, with users frequently seeking wellness tracking, health tips, and daily routine management. This trend remained steady across all days and months, highlighting the central role of health in digital habits. The analysis also noted seasonal spikes in specific queries, such as increased advice-seeking around Valentine’s Day in February. All data was processed with privacy safeguards by extracting only conversation summaries to determine topic and intent.
🔗 Source: View Source | Found on Dec 10, 2025
Codex, OpenAI's AI coding agent available in ChatGPT Plus, Pro, Business, Edu, and Enterprise plans, now integrates with the Hugging Face Skills repository to automate end-to-end machine learning experiments. Codex can validate datasets, select hardware based on model size (e.g., $1-2 for models under 1B parameters; $5-15 for 1-3B; $15-40 for 3-7B), generate training scripts, monitor progress via Trackio, and update training reports. It supports supervised fine-tuning, direct preference optimization, and reinforcement learning with verifiable rewards for models from 0.5B to 7B parameters. Trained models can be converted to GGUF for local deployment and pushed to the Hub.
🔗 Source: View Source | Found on Dec 11, 2025
At Convergence 2025, held December 9–12, Microsoft showcased advances in generative AI and agentic business applications that are transforming work by enabling autonomous enterprises powered by data, Copilot, and agents across Dynamics 365 and the Microsoft Cloud. The Dynamics 365 ERP Model Context Protocol (MCP) server now supports millions of ERP actions at scale with dynamic integration for live business signals. New agent templates—such as the Product Change Management Agent—automate workflows for manufacturers like Coca-Cola Beverages Africa, reducing approval times from weeks to days. Partner-built agents including Shop Floor by RSM, PayFlow Agent by HSO, and Quality Impact Recall Agent by Cegeka extend domain workflows through MCP integration.
🔗 Source: View Source | Found on Dec 09, 2025
NVIDIA achieved a record-breaking benchmark of 410 trillion traversed edges per second (TEPS) on the 31st Graph500 breadth-first search (BFS) list, using an accelerated computing cluster with 8,192 NVIDIA H100 GPUs at a CoreWeave data center in Dallas. The system processed a graph with 2.2 trillion vertices and 35 trillion edges, more than doubling the performance of comparable solutions while using just over 1,000 nodes versus about 9,000 for similar entries. This delivered three times better performance per dollar by leveraging technologies such as NVIDIA CUDA, Spectrum-X networking, NVSHMEM, and InfiniBand GPUDirect Async.
🔗 Source: View Source | Found on Dec 10, 2025
Marks assesses whether AI is in a bubble, distinguishing “mean-reversion” bubbles that destroy wealth from “inflection” bubbles that also accelerate progress. He argues AI has clear transformative potential but is surrounded by classic bubble features: euphoric narratives, speculative capital, circular deals, aggressive debt-financed infrastructure and great uncertainty over winners, profitability and asset longevity. Unlike past manias, today’s leaders generate real revenue and cash flow, yet valuations still embed large unknowns. He concludes nobody can reliably label today as a bubble; investors should avoid both all-in and all-out positions, instead taking moderate, selective exposure and treating AI debt with particular caution.
🔗 Source: View Source | Found on Dec 09, 2025
Man Group expects 2026 to be characterised by elevated uncertainty, driven by geopolitical risk, protectionism and policy divergence. Tariffs and other policy choices are likely to weigh on trade and labour markets, while AI both disrupts employment and exacerbates inequality. Together, these forces should slow global growth, with several major developed economies at risk of at least a mild recession, even as some emerging markets in Asia prove more resilient. Fragmented monetary and fiscal responses, rising debt-service burdens and more activist “visible hand” government intervention further raise macro volatility, reinforcing the case for diversification and agility.
🔗 Source: View Source | Found on Dec 11, 2025
Initially, tech giants funded data centres, chip development, and cloud infrastructure through strong free cash flow and high equity valuations, enabling sizeable capex with conservative capital structures. As AI-driven technology demand now far exceeds supply, global data centre capacity may need to increase sixfold by 2030, requiring approximately US$3 trillion in investment. Hyperscalers are expected to fund about half of this amount, with the remainder sourced externally. Consequently, there is a rapid shift from equity to credit markets for AI infrastructure financing, notably via public investment-grade corporate bonds and innovative hybrid deal structures.
🔗 Source: View Source | Found on Dec 08, 2025
AI has become a central risk-return driver in non-investment grade credit. Around $20 billion of recent AI-linked leveraged issuance, with up to $150 billion forecast, raises execution and leverage risks for narrow, project-financed structures. Most high yield and loan sectors are viewed as neutral or net beneficiaries, with only a minority structurally at risk, as shown in the industry heat maps on pages 2–3. Key upside areas include semiconductors, hardware, health care and utilities; more vulnerable are selected software, media and people-intensive services. Ultimately, scale, moats, proprietary data and credible AI execution will determine winners, reinforcing the importance of issuer-level security selection.
🔗 Source: View Source | Found on Dec 10, 2025