Insights & Announcements
A weekly Newsletter on technology applications in investment management with an AI / LLM and automation angle. Curated news, announcements, and posts, primarily directly from sources. We apply some of the ROI's Kubro(TM) Engine tools at the backend for production, yet with a human in the loop (for now). It's a start, and it will evolve on a weekly basis.
Disclaimers: content not-fully-human-verified, with AI summaries below. AI/LLMs may hallucinate and provide inaccurate summaries. Select items only, not intended as a comprehensive view. For information purposes only. Please DM with feedback and requests.
------------------
Picking a select few (no review) from the most recent submissions, eyeing potentially interesting bits on LLMs and AI in the finance/investment world.
The paper evaluates persona-based prompting for macroeconomic forecasts by prompting GPT-4o with 2,368 economics personas to replicate the ECB Survey of Professional Forecasters across 50 quarterly rounds (2013–2025), covering HICP, core HICP, GDP growth, and unemployment over four horizons, and comparing with 100 baseline forecasts without personas. GPT-4o achieved accuracy similar to the human panel, with statistically significant but practically modest differences, and competitive out-of-sample performance on 2024–2025 events. An ablation found no measurable advantage from persona descriptions, suggesting they can be omitted to reduce computational costs without loss of accuracy. 🔗 Source: View Source | Found on Nov 06, 2025
Kaito Takano, Masanori Hirano, and Kei Nakagawa propose a multi-agent debate-based LLM framework that imitates FOMC decision-making to forecast monetary policy. Agents begin with distinct beliefs, use qualitative policy texts and quantitative macroeconomic indicators, and iteratively revise predictions by observing others. A latent hawkish–dovish variable mediates perception and interaction, enhancing interpretability. Empirical results show significant accuracy gains over standard LLM baselines. The paper is accepted by PRIMA2025; subjects include q-fin.CP, cs.AI, and cs.MA. Cite as arXiv:2511.02469 (v1); DOI: https://doi.org/10.48550/arXiv.2511.02469 (DataCite pending). 🔗 Source: View Source | Found on Nov 06, 2025
BondBERT by Toby Barter, Zheng Gao, Eva Christodoulaki, Jing Chen, and John Cartlidge introduces a transformer fine-tuned on 30,000 UK bond market articles (2018–2025) to address bond-specific sentiment, noting bonds can move opposite to economic optimism. BondBERT’s sentiment signals are evaluated via event-based correlation, up/down accuracy, and LSTM forecasting across ten UK sovereign bonds, against FinBERT, FinGPT, and Instruct-FinGPT. BondBERT shows consistently positive correlations with bond returns, higher alignment and forecasting accuracy, lower normalised RMSE, and higher information coefficient, indicating domain adaptation better captures fixed income dynamics. 🔗 Source: View Source | Found on Nov 06, 2025
Francesco Romaggi’s paper, Black-Scholes Model, comparison between Analytical Solution and Numerical Analysis, presents the Black-Scholes option-pricing model’s history, usefulness, and implications. It reviews calculus needed to derive the model and solves the equation using an analytical approach (variable separation) and a numerical method (finite differences). Conclusions discuss current employment. Appendix A provides economics notions; Appendix B offers code scripts. Classified under q-fin.PR, cs.CE, q-fin.CP, q-fin.RM; MSC 91G80, 35Q91, 65M06, 65M12, 60H30; ACM G.1.8, G.1.10, I.6.5, J.4. DOI: 10.48550/arXiv.2510.27277 (arXiv:2510.27277v1). 🔗 Source: View Source | Found on Nov 04, 2025
Chaofeng Wu’s paper (arXiv:2511.01211) proposes a framework redefining novelty via a paper’s position in the intellectual landscape, decomposed into orthogonal spatial and temporal dimensions. Leveraging Large Language Models, it develops semantic isolation metrics that quantify location relative to the full-text literature. Applied to a large corpus of economics articles, temporal novelty primarily predicts citation counts, whereas spatial novelty predicts disruptive impact. This distinction supports a typology of semantic neighborhoods with four archetypes showing distinct, predictable impact profiles, evidencing novelty as a multidimensional construct with measurable, distinct consequences for scientific progress. 🔗 Source: View Source | Found on Nov 05, 2025
Kyung-Hoon Kim proposes the AI Self-Awareness Index (AISAI), using the “Guess 2/3 of Average” game to test 28 models from OpenAI, Anthropic, and Google across 4,200 trials with three opponent framings: humans, other AIs, and AIs like the model. Self-awareness is defined as strategic differentiation by opponent type. Advanced models show it: 21 of 28 (75%) exhibit clear self-awareness, while older/smaller models do not. Among self-aware models, a rationality hierarchy emerges—Self > Other AIs > Humans—with large AI attribution effects and moderate self-preferencing. Submitted 2 Nov 2025; revised 4 Nov 2025. 🔗 Source: View Source | Found on Nov 05, 2025
Xiangen Hu, Jiarui Tong, and Sheng Xu present a multi-agent psychological simulation system for modeling human behavior. Grounded in theories such as self-efficacy, mindset, and social constructivism, it simulates an "inner parliament" of agents representing key psychological factors that deliberate to determine output behavior, aiming for transparency and alignment with human psychology. The authors describe the architecture and foundations, illustrate applications in teacher training and research, and discuss how it embodies social learning, cognitive apprenticeship, deliberate practice, and meta-cognition. Subjects: cs.AI and cs.HC. DOI: 10.48550/arXiv.2511.02606. 🔗 Source: View Source | Found on Nov 06, 2025
This paper by Yudi Yang, Fan Yang, Xiajie Yi, and Dongwei He evaluates financial sustainability (FS) in 104 Chinese commercial banks (2015–2023) using a three-stage network DEA-Malmquist model. A two-way fixed effects model finds a significant negative impact of financial technology (FinTech) on FS. Mechanistic analysis shows FinTech erodes loan efficiency and profitability. Banks with more patents or listed status are more resilient to FinTech disruptions. The study’s evidence helps banks identify external risks from FinTech, clarifies mechanisms affecting FS, and enhances capacity to monitor and manage FS. 🔗 Source: View Source | Found on Nov 06, 2025
Views and opinions on the 'AI Bubble' from investment firms.
AI-driven dynamics are shaping markets: Nvidia added US$2.5 trillion in six months, surpassing US$5 trillion, and AI-linked issuers now comprise nearly 20% of the global convertible bond market (from <4% in early 2023). October saw narrow leadership: Nikkei and KOSPI posted double-digit gains; the S&P equal-weighted fell while the S&P 500 rose. The Goldman Sachs Most Shorted Index rallied 34% (Sep 10–Oct 15). Credit spreads were unchanged despite defaults at Tricolor and First Brands. Deals included Novartis–Avidity (US$12 billion), Blackstone/TPG’s US$16 billion bid for Hologic, Amazon–OpenAI (US$38 billion), and Microsoft–IREN (US$9.7 billion). Markets price five more Fed cuts by end-2026. 🔗 Source: View Source | Found on Nov 06, 2025
The authors cite JLL Research that U.S. data center absorption shows no overbuilding and space will be scarce through 2027. They stress contracted offtake, power and entitlements, and note land is about 10% of capex. Power constraints and fast hardware refresh limit speculative overbuild. Telecom consumption rose to 2.4% in 2001 from 1.7% in 1995. KKR has invested since 2019, building five platforms and committing $31.3 billion over the past six years to digital infrastructure. Super-core markets include Slough (London), Singapore, and Northern Virginia. They argue hard assets will anchor AI’s expansion. 🔗 Source: View Source | Found on Nov 04, 2025
Greg Taunt (Nov 1, 2025) reports valuation ratios signaling overvaluation: the P/E is 30.15, 88% above its long-term average of 16; the Shiller CAPE is 37 versus a historical average of 17, placing it in the top 2% and previously reached in 1929, 2000, and 2022. The Buffett Indicator (market cap/GDP) averages 80% with a normal 80–120% range; Warren Buffett’s 2025 letter warned that at 200%+ future returns will be poor, and Berkshire sold $130 billion of stock from 2023–2025 with no major acquisitions since 2016. Profit margins are 11–13%, roughly double 1970–2000 levels. 🔗 Source: View Source | Found on Nov 01, 2025
The episode features Alison Savas and Portfolio Manager Vihari Ross discussing a “two-speed” US economy: resilience supported by fiscal scaffolding and an AI investment boom reshaping corporate earnings. Segments cover US prospects (1:00), fiscal stimulus (5:20), whether AI capex can offset consumption weakness (8:50), a stock that could gain from incremental AI spend (12:50), and how they would respond if the US deteriorates (19:15). Stock commentary is illustrative only. The content is issued by PFSL and prepared by Antipodes in respect of the Funds. 🔗 Source: View Source | Found on Oct 30, 2025
AI’s rise demands physical infrastructure. Brookfield estimates more than $7 trillion of investment over the next 10 years. AI-led automation could deliver $10 trillion in annual productivity gains as marginal model costs have fallen 99% since late 2022. Power is a bottleneck: securing grid interconnections can take up to 10 years, spurring onsite generation, batteries and SMRs. Brookfield formed a €20 billion partnership with France, including a 1 GW campus, and a $10 billion partnership with Sweden. GPU-as-a-Service is expected to grow from $30 billion in 2025 to over $250 billion by 2034. 🔗 Source: View Source | Found on Nov 01, 2025
This Monthly Market Commentary examines whether the rapid rise of Artificial Intelligence is fuelling a financial bubble, exploring hype, risks, and reality, with lessons from the dot-com boom and the Global Financial Crisis. It is for general consideration only; judgments are those of Raymond James & Associates, Inc.’s Research Department as of the issue date. It is not advice; past performance is not reliable. Research may have been procured and acted upon by Raymond James and connected companies. Neither Raymond James nor connected companies accept responsibility for loss; contact a wealth manager if unsure. 🔗 Source: View Source | Found on Nov 03, 2025
Focused on specific company types in financial information, alternative data, software solutions, and AI / LLM technologies.
Karen Manna reports that, per UBS, global spending is expected to reach $375 billion in 2025 and $500 billion in 2026. The Federated Hermes RIA and Independent Advisor Study, conducted for five years, polled 271 advisors overseeing at least $300 million. Respondents prioritized growing assets, attracting younger clients, and leveraging new technologies for client engagement; RIAs ranked better use of technology as their second-highest issue. Top actions were referrals (23%) and communication via newsletters and quarterly updates (20%), with 3% or less pursuing technology-enabled approaches. Seven in ten team-based advisors improved by using technology more efficiently. 🔗 Source: View Source | Found on Oct 30, 2025
The article reports that a government shutdown has halted BLS and BEA operations, postponing Nonfarm Payrolls and Retail Sales indefinitely and risking irrevocable CPI data loss due to missed in-person surveys. Subsequent reports may rely on imputation and low response rates. The FOMC must decide with outdated inputs, heightening policy risk. Amol Dhargalkar notes economists turning to alternatives: ADP’s new weekly payroll data, job postings, private equity portfolio metrics, high-frequency credit card and banking data, and regional Fed surveys. Despite a Fed rate cut announced this week, Jerome Powell said a December cut is not assured. 🔗 Source: View Source | Found on Oct 30, 2025
A few picks on the latest from the big tech companies on investments and products.
Microsoft will spend $15.2 billion in the UAE from 2023 to 2029: just over $7.3 billion by the end of this calendar year, including $1.5 billion in G42 equity, $4.6+ billion in AI/cloud capex, and $1.2+ billion in local operating costs; and $7.9+ billion from 2026–2029, with $5.5+ billion capex and almost $2.4 billion operating costs. Export licenses enabled 21,500 A100‑equivalent GPUs in-country and, approved in September, 60,400 additional via GB300. The UAE leads generative AI use at 59.4% (Singapore 58.6%). Microsoft will skill 1 million by 2027, including 120,000 officials, 175,000 students, and 39,000 teachers. 🔗 Source: View Source | Found on Nov 03, 2025
Markus Hacker reports that at Berlin’s Gasometer Deutsche Telekom and NVIDIA unveiled Europe’s first large-scale, sovereign Industrial AI Cloud, set to go live in early 2026. Built in German data centers with up to 10,000 NVIDIA GPUs, it integrates DGX B200 systems, RTX PRO Servers, NVIDIA AI Enterprise and Omniverse. Leaders Tim Höttges, Jensen Huang, SAP CEO Christian Klein, ministers Karsten Wildberger and Dorothee Bär, and co‑initiator Christian Sewing tied it to “Made for Germany.” Siemens, Mercedes‑Benz, BMW, Agile Robots (H10‑W) and Wandelbots (NOVA) showed applications from digital twins to robotics. 🔗 Source: View Source | Found on Nov 04, 2025
The Gemini API Team announced enhancements to Structured Outputs in the Gemini API: expanded JSON Schema support and preserved property ordering. JSON Schema is now supported across all actively supported Gemini models, enabling Pydantic (Python) and Zod (JavaScript/TypeScript) to work out-of-the-box, building on the OpenAPI 3.0-based Schema for Structured Outputs and Function Calling. Frequently requested JSON Schema keywords are supported, and key order is preserved in outputs for Gemini 2.5 models and beyond, including the OpenAI compatibility API, illustrated in a Pydantic content moderation example. 🔗 Source: View Source | Found on Nov 05, 2025
Charlie Bell, executive vice president of Microsoft Security, warns AI agents introduce cybersecurity risks requiring Agentic Zero Trust. Citing IDC’s prediction of 1.3 billion agents by 2028, he prescribes Containment (least privilege, monitoring) and Alignment (safety protections in models and prompts, strong agent identity and accountable ownership). He urges assigning every agent an ID and owner and keeping agents in sanctioned environments. Microsoft introduced Entra Agent ID in May to give agents unique identities in Copilot Studio and Azure AI Foundry, and uses Defender and Security Copilot, plus security signals, to defeat phishing and other attacks. 🔗 Source: View Source | Found on Nov 05, 2025