Insights & Announcements
A weekly Newsletter on technology applications in investment management with an AI / LLM and automation angle. We combine 100% human curation/selection with LLM standardisation, summarisation, and more deterministic search/collection, classification and workflow - powered by Kubro(TM). Curated news, announcements, and posts, primarily directly from sources (Arxiv papers, major AI/Tech/Data companies, investment firms). See disclaimers at the bottom. Please DM with feedback and requests.
The paper, submitted on 9 Jul 2026 by Rian Dolphin, Joe Dursun, Jarrett Blankenship, Katie Adams, and Quinton Pike, presents a two-stage system for grounded event extraction from SEC 8-K filings. It uses a three-tier taxonomy of 119 event types, constrains tags to valid taxonomy entries, anchors each tag to a verbatim quote with fuzzy n-gram validation, and then re-grades the quote against the category definition. Applied to 292,984 filings from 2022 to 2026, the system produced 601,088 grounded event tags. Over 5,125 stratified tags, precision rose from 12% to 96% as quality scores increased, while unsupported tags fell from 8% to near zero.
🔗 Source: Summary based on View Source from arxiv.org | Found on Jul 10, 2026
The paper analyzes how four forces reshape the AI industry from 2026 to 2030: a DRAM/HBM price surge, frontier-capable open-weight models such as GLM-5.2, rapid inference-efficiency gains, and Meta and xAI entering compute resale on fleets bought before memory repricing. It finds the entrant-incumbent cost gap persists, at 3.2x in 2026, 1.9x in 2027, and 3-4x by 2029-2030. Training splits into a $18-38B frontier tier and a mass tier falling toward $5M. Solvency requires roughly 2x annual token-demand growth for four years with sticky premium pricing.
🔗 Source: Summary based on View Source from arxiv.org | Found on Jul 09, 2026
The paper, submitted on 5 Jul 2026 by Yiqing Wang, Yixin Kang, Luyun Lin, and Siqi Mao, says SR 26-2 modernizes U.S. model risk management by replacing SR 11-7 with a more risk-based and materiality-sensitive supervisory framework. It notes that generative and agentic AI are excluded, creating a governance challenge for banking organizations and other financial institutions. The authors propose the Generative AI Control Framework (GAICF), an SR 26-2-compatible governance framework for generative AI-enabled financial workflows that translates model risk management principles into a layered control structure for applications embedded in regulated banking processes.
🔗 Source: Summary based on View Source from arxiv.org | Found on Jul 07, 2026
Ati q Ur Rehman’s paper, “Unveiling Public Opinion: A Study of Sentiment Analysis Using LSTM and Traditional Models,” examines sentiment analysis on Twitter using logistic regression, random forest, naïve bayes, gradient boosting, and LSTM networks. The study uses a Kaggle Twitter dataset preprocessed with tokenization, lemmatization, and stopword elimination. It reports that the LSTM model achieved 90.98% training accuracy, 80.00% testing accuracy, and a 0.92 micro-average ROC-AUC score, and found that LSTM outperformed conventional machine learning techniques in capturing contextual and sequential textual aspects.
🔗 Source: Summary based on View Source from arxiv.org | Found on Jul 10, 2026
The paper compares nine zero-shot pretrained time series foundation models with eight econometric specifications for forecasting realized volatility using the VOLARE dataset. It covers 50 assets across equities, foreign exchange, and futures, and evaluates three forecast horizons with formal pairwise and multi-model tests. The models do not show a uniform advantage: pooled losses favor foundation models, but the gain is driven by a few outlier assets. Only Tiny Time Mixers (TTM) beats the Log-HAR benchmark at every horizon, and only narrowly. A simple equal-weight average of TTM and Log-HAR matches the best single model and enters the Model Confidence Set for 98% to 100% of assets.
🔗 Source: Summary based on View Source from arxiv.org | Found on Jul 07, 2026
The paper studies market stability among self-interested agents using a multi-agent marketplace simulation with 18 DeepSeek-V3 LLM agents, each with complementary production specialties, trading within a constrained social network. It compares eight mechanism conditions under progressive troll injection over 200 rounds and identifies Mediation as the top-performing mechanism. In a second phase, adversarial red-teaming with iteratively prompt-optimised LLM-driven trolls finds that the best attack, v6, reduces honest-agent utility by 13.3% but does not collapse the market. The authors conclude that Mediation can sustain positive honest-agent utility under optimised attack.
🔗 Source: Summary based on View Source from arxiv.org | Found on Jul 10, 2026
The paper “Volatility in Prediction Markets: A Structural Approach,” submitted on 9 Jul 2026 by Weiye Xi, Ciamac C. Moallemi, Mallesh Pai, and Shouqiao Want, develops and estimates a volatility model for binary prediction markets. The model combines a Wright-Fisher deadline-resolution component and a Glosten-Milgrom order-flow component. Using a large panel of Kalshi contracts, the authors report that these structural variables have substantial forecasting power, outperforming ARCH/GARCH benchmarks. Adding residual GARCH dynamics gives the best overall forecasts. Volatility is highest near fifty-fifty prices and near resolution, with economics contracts smoother and sports contracts more jump-like.
🔗 Source: Summary based on View Source from arxiv.org | Found on Jul 10, 2026
In a July 9, 2026 article, Corey Hoffstein discussed with the author on the Flirting with Models podcast how edge in quantitative investing has shifted from raw data to what is done with data. At Two Sigma, researchers, feature engineers, and other engineers turn creatively identified facts into predictive features, using hypothesis-driven testing and moving on when data does not support a prior. The article says LLMs expand research by generating new textual data, speeding exploration of ideas that once took months, including CEO microexpression analysis. Two Sigma aims to use AI to amplify originality and orthogonality while avoiding homogeneous results.
🔗 Source: Summary based on View Source from twosigma.com | Found on Jul 09, 2026
The article says major technology waves contain smaller “handoffs” or “shakeouts,” marked by new IPOs, more private companies, sharper drawdowns, and greater volatility, often replacing the leading firms. It asks whether software is in such a phase now. Based on quarterly net new ARR, AI labs are starting to add more revenue than both hyperscalers and the entire public software universe. The article notes that when software shifted from the enterprise era to SaaS, only a handful of the largest names carried over, and the current AI leaders, including OpenAI and Anthropic, did not exist during the last shakeout.
🔗 Source: Summary based on View Source from coatue.com | Found on Jul 07, 2026
Goldman Sachs Asset Management said corporate AI adoption is accelerating, but scaling is constrained by available computing power. Brook Dane and Sung Cho said enterprise AI use is broadening beyond coding, with one company reporting that the top 5% of users consume three times the tokens of the median company. They described a compute-constrained environment affecting ASICs, memory chips, fiber optics, and data-center links, while noting inference demand is rising. They said AI has not hurt online advertising; instead, AI queries and platform use are creating more valuable ad inventory. They also said valuations remain attractive despite a June selloff in AI-related tech stocks.
🔗 Source: Summary based on View Source from goldmansachs.com | Found on Jul 09, 2026
Central bankers use the real neutral rate, or r*, to judge whether policy is restrictive or stimulative. It is defined as the real short-term interest rate at full employment and stable inflation, and it changes over time with demographics, productivity, and savings-investment trends. Because r* is unobservable, it must be estimated with considerable uncertainty. A market-based proxy is the 5-year, 5-year-forward rate (5y5y), which reflects the expected future average overnight rate plus a term premium. Event studies of 5y5y real and nominal rates can benchmark market reactions to new AI model releases.
🔗 Source: Summary based on View Source from pimco.com | Found on Jul 10, 2026
AE Studio and Anthropic introduced GRAM, or Gradient-Routed Auxiliary Modules, a method for controlling dual-use knowledge in large language models. GRAM adds removable modules to each Transformer layer and routes learning from dual-use data, such as virology, cybersecurity, nuclear physics, and a niche programming language, into category-specific compartments while freezing general weights. In tests on synthetic stories, realistic web-text mixtures, and models from 50 million to 5 billion parameters, deleting a module removed the related capability with little or no loss in general performance. The approach matched data filtering, and removed knowledge was harder to recover than after post-training unlearning.
🔗 Source: Summary based on View Source from anthropic.com | Found on Jul 09, 2026
On July 7, 2026, Philipp Schmid announced new capabilities for Managed Agents in the Gemini API: background execution, remote MCP server integration, custom function calling, and refreshed credentials across interactions. The updates are described as addressing developer feedback and product needs to help build reliable, production-ready agents. In the Gemini Interactions API, users call a single endpoint while Gemini handles reasoning, code execution, package installation, file management, and web information inside an isolated cloud sandbox.
🔗 Source: Summary based on View Source from blog.google | Found on Jul 08, 2026
IBM announced major updates to IBM Bob, its agentic software development platform, including new multi-agent capabilities, built-in AI cost and use analytics through Bobalytics, and pre-built workflows for modernizing enterprise systems. IBM said 85% of DevSecOps professionals surveyed agree AI has shifted the bottleneck from writing code to reviewing and validating it. Jack Henry said IBM Bob helped accelerate RPG development workflows and improve code quality, while Blue Pearl said a legacy modernization program originally projected to take nine months with 14 engineers was completed in three days. IBM Bob is also adding premium packages for IBM Z, IBM i, and Java modernization.
🔗 Source: Summary based on View Source from newsroom.ibm.com | Found on Jul 10, 2026
LangChain tuned its Deep Agents harness for NVIDIA Nemotron 3 Ultra and reported the highest accuracy among open models, more completed tasks, higher throughput, and 10x lower inference cost per run than leading closed models. Against LangChain’s Deep Agents benchmark, Nemotron 3 Ultra also reached business task parity with the highest-scoring closed models. No retraining was required; the gains came from tuning system prompts, tool descriptions, and middleware. The tuned profile is available through LangChain, and NVIDIA NemoClaw for LangChain Deep Agents packages the open stack for enterprise use.
🔗 Source: Summary based on View Source from blogs.nvidia.com | Found on Jul 08, 2026
Anthropic’s Long-Term Benefit Trust (LTBT) appointed Dr. Ben Bernanke, a Distinguished Fellow at the Brookings Institution and former Chair of the Federal Reserve, as its newest member. Bernanke led the Fed from 2006 to 2014, including during the 2008 global financial crisis and the recovery that followed. Before government service, he spent more than two decades as an academic economist, much of it at Princeton, and won the 2022 Nobel Prize in Economic Sciences. The LTBT is an independent body that supports Anthropic’s mission of responsibly developing advanced AI for the long-term benefit of humanity and can appoint members to Anthropic’s board.
🔗 Source: Summary based on View Source from anthropic.com | Found on Jul 09, 2026
Microsoft says identity security is becoming a key pressure point as AI accelerates cyberattacks across the attack chain. At RSA earlier this year, it unveiled a unified identity risk score in Microsoft Entra that correlates signals across accounts, sessions, workloads, and applications to support real-time access decisions. Microsoft also announced an updated Entra ID Protection experience, a coming identity-focused RBAC role in public preview, and just-in-time access via privileged identity management. The Conditional Access Optimization Agent and Security Alert Triage Agent are being extended with more detections, recommendations, automatic attack disruption, and predictive shielding.
🔗 Source: Summary based on View Source from techcommunity.microsoft.com | Found on Jul 09, 2026
The article describes an iterative pipeline for generating financial news headlines with NVIDIA NeMo Data Designer, NeMo Curator, and Nemotron models. A 50,000-headline baseline retained only 17,348 unique headlines, with 65% removed as near-duplicates. The full run produced 502,536 unique headlines across 13 categories in 82 iterations, using about 6 days of compute on a single 8-way NVIDIA B200 node. The process used global semantic deduplication, category-weight correction, and farthest-from-centroid few-shot selection. Batch size was reduced from 50,000 to 35,000 after the deduplication drop rate exceeded about 80%.
🔗 Source: Summary based on View Source from developer.nvidia.com | Found on Jul 10, 2026
The article says countries are investing in domestic AI infrastructure, local data and homegrown expertise to build AI capabilities tailored to local citizens, services and regulations. It identifies five ingredients of a national AI strategy: AI imperative, AI-ready workforce, AI models and data, AI factories and public-private partnerships. Since 2019, NVIDIA’s AI Nations initiative has supported AI ecosystems and workforce development worldwide. Examples include ThinkDeep helping France’s Ministry of Economy and Finance cut document search times from two days to two minutes and save 2 million euros for 10,000 employees; India’s Sarvam serving 22 official languages; and Widelabs improving legal services in Brazil.
🔗 Source: Summary based on View Source from blogs.nvidia.com | Found on Jul 07, 2026
At ICML 2026, NVIDIA had 74 accepted papers, with about 2,000 papers citing NVIDIA GPUs and 145 citing NVIDIA Nemotron. Research highlighted open frontier models and open AI infrastructure across vision, video generation, reinforcement learning, agent training, inference, robotics, autonomous vehicles, and biomedical research. DreamDojo used NVIDIA Cosmos to learn from human video and support robot policy evaluation. BioNeMo supported life sciences work such as FLIP2 and KERMT. Sakana AI built Fugu and Fugu-Ultra on Nemotron 3 Ultra, and KiloCode reported up to 90% token cost reductions using Nemotron.
🔗 Source: Summary based on View Source from blogs.nvidia.com | Found on Jul 06, 2026
The article asks who should benefit from advanced technological innovation and argues for an expansive answer: technologies, including advanced AI, should be designed, developed, and distributed to benefit everyone. It grounds this claim in five moral arguments: human rights, beneficence, contingencies of birth, the global tree of knowledge, and global economic justice. Together, these arguments support the case for globally beneficial technologies.
🔗 Source: Summary based on View Source from deepmind.google | Found on Jul 07, 2026
The AI Economy Institute (AIEI) announced its third cohort of researchers on July 7, 2026, to study how AI is being adopted across economies, industries, and communities. The third global research call focuses on frontier firms and how they are reshaping work and the broader economic landscape. Researchers will examine firm-level and economy-wide impacts, including job design, skill demands, productivity, and regional economic development. AIEI said it has received more than 800 responses to its research calls since launch. The institute said the cohort aims to provide empirical evidence to help policymakers, firms, and institutions make decisions in a rapidly evolving AI economy.
🔗 Source: Summary based on View Source from blogs.microsoft.com | Found on Jul 08, 2026