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 (Arxiv papers, major AI/Tech/Data companies, investment firms). 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
The paper evaluates the impact of large language model (LLM)-based news sentiment analysis on stock price movement prediction, comparing DeBERTa, RoBERTa, and FinBERT models. DeBERTa achieved the highest accuracy at 75%, while an ensemble of all three models increased accuracy to approximately 80%. The study also found that incorporating sentiment news features provided slight benefits to certain stock market prediction models, specifically LSTM-, PatchTST-, and tPatchGNN-based classifiers as well as PatchTST- and TimesNet-based regression models.
🔗 Source: Summary based on arxiv.org View Source | Found on Feb 03, 2026
The article introduces MAS4TS, a tool-driven multi-agent system designed for general time series tasks, developed by Weilin Ruan and Yuxuan Liang. MAS4TS employs an Analyzer-Reasoner-Executor paradigm that incorporates agent communication, visual reasoning over time series plots using a Vision-Language Model to extract temporal structures, and latent reconstruction of predictive trajectories. The system features three specialized agents coordinating via shared memory and gated communication, with a router selecting task-specific tool chains. Experiments on multiple benchmarks show that MAS4TS achieves state-of-the-art performance across diverse time series tasks with strong generalization and efficient inference.
🔗 Source: Summary based on arxiv.org View Source | Found on Feb 04, 2026
PredictionMarketBench, introduced by Avi Arora and Ritesh Malpani in a paper submitted on January 28, 2026, is a SWE-bench-style benchmark designed for backtesting algorithmic and LLM-based trading agents on prediction markets using deterministic, event-driven replay of historical limit-order-book and trade data. The framework standardizes episode construction from raw exchange streams, provides an execution-realistic simulator with maker/taker semantics and fee modeling, and offers a tool-based agent interface supporting both classical strategies and tool-calling LLM agents. Four Kalshi-based episodes covering cryptocurrency, weather, and sports are released. Baseline results indicate naive agents underperform due to transaction costs and settlement losses.
🔗 Source: Summary based on arxiv.org View Source | Found on Feb 03, 2026
The article introduces an interactive framework where Large Language Models (LLMs) generate and explain Cypher graph queries, allowing users to iteratively refine them using natural language. This approach is applied to real-world Knowledge Graphs (KGs) to enhance accessibility to complex datasets while maintaining factual accuracy and semantic rigor. The core quantitative evaluation uses a 90-query benchmark on a synthetic movie KG to assess query explanation quality and fault detection across multiple LLMs, supplemented by two smaller experiments on the Hyena KG and the MaRDI (Mathematical Research Data Initiative) KG.
🔗 Source: Summary based on arxiv.org View Source | Found on Feb 06, 2026
The article, "Generative AI for Stock Selection" by Keywan Christian Rasekhschaffe, examines the automation of feature discovery in U.S. equities using generative AI. Large language models with retrieval-augmented generation and structured prompting synthesize features from analyst, options, and price-volume data for input into a tabular machine-learning model forecasting short-horizon returns. Across multiple datasets, AI-generated features show Sharpe improvements ranging from 14% to 91%, depending on dataset and configuration. Retrieval quality significantly impacts outcomes, and AI-generated signals are weakly correlated with traditional features, supporting their combination while reducing manual engineering effort.
🔗 Source: Summary based on arxiv.org View Source | Found on Feb 03, 2026
The article introduces a novel LLM-guided no-regret portfolio allocation framework that combines online learning dynamics, market sentiment indicators, and large language model (LLM)-based hedging to create high-Sharpe ratio portfolios for risk-averse investors and institutional fund managers. The method utilizes a follow-the-leader approach enhanced with sentiment-based trade filtering and LLM-driven downside protection. Empirical results show that this approach outperforms a SPY buy-and-hold baseline by 69% in annualized returns and 119% in Sharpe ratio. The research was authored by Muhammad Abro and Hassan Jaleel and submitted on January 16, 2026.
🔗 Source: Summary based on arxiv.org View Source | Found on Jan 27, 2026
The article, accepted by ICASSP 2026, introduces a Reinforcement Learning framework with Fine-grained Knowledge Verification (RLFKV) to reduce hallucinations in financial Retrieval-Augmented Generation (RAG) systems. The method decomposes financial responses into atomic knowledge units and evaluates each for correctness to compute a fine-grained faithful reward, enhancing alignment with retrieved documents. An informativeness reward is also used to ensure the policy model retains at least as many knowledge units as the base model. Experiments on the Financial Data Description (FDD) task and the new FDD-ANT dataset show consistent improvements using this approach.
🔗 Source: Summary based on arxiv.org View Source | Found on Feb 06, 2026
Claude Opus 4.6 introduces significant improvements over its predecessor, including enhanced coding skills, better code review and debugging, and a 1M token context window in beta. It achieves state-of-the-art performance on benchmarks such as Terminal-Bench 2.0, Humanity’s Last Exam, GDPval-AA (outperforming GPT-5.2 by 144 Elo points and Opus 4.5 by 190 points), and BrowseComp. The model supports up to 128k output tokens and offers adaptive thinking with four effort levels. Safety evaluations show low rates of misaligned behavior and over-refusals, with new cybersecurity probes implemented for enhanced protection. Pricing remains at $5/$25 per million tokens.
🔗 Source: Summary based on anthropic.com View Source | Found on Feb 05, 2026
By mid-2025, Qwen became the model with the most derivatives on Hugging Face, with over 113,000 models using it as a base and over 200,000 repositories tagging Qwen, surpassing Meta's Llama and DeepSeek. Alibaba led in organization-wide derivatives, nearly matching Google and Meta combined, while aligning model development with cloud and hardware infrastructure. Tencent accelerated open releases from May 2025 under Tencent HY. ByteDance contributed open-source artifacts like UI-TARS-1.5 and Seed-Coder; its Doubao app surpassed 100 million DAU in December 2025. Baidu shifted to open source with Ernie 4.5 and announced Kunlunxin’s IPO on January 1, 2026.
🔗 Source: Summary based on huggingface.co View Source | Found on Feb 03, 2026
Quantum computers, which can solve problems beyond the reach of classical supercomputers, are poised to revolutionize fields such as drug discovery, materials science, and energy. However, their capabilities also pose a significant security risk: large-scale quantum computers could break current public-key cryptosystems that protect sensitive information like bank transfers and classified data. Although such quantum computers do not yet exist, malicious actors may already be conducting “store now, decrypt later” attacks by collecting encrypted data in anticipation of future decryption capabilities. The article emphasizes the urgent need to prepare for these impending security challenges.
🔗 Source: Summary based on blog.google View Source | Found on Feb 06, 2026
Google has expanded its AI infrastructure by partnering with NVIDIA to offer the new Vera Rubin GPU platform and developing its own TPUs for a decade. In December 2025, Google announced plans to acquire Intersect, a provider of data center and energy infrastructure solutions. Gemini serving unit costs were reduced by 78% in 2025 through optimizations. Gemini 3 Pro, which processes three times as many daily tokens as 2.5 Pro, powers Google Antigravity with over 1.5 million weekly active users since launch two months ago. First-party models now process over 10 billion tokens per minute via API use, up from 7 billion last quarter.
🔗 Source: Summary based on blog.google View Source | Found on Feb 04, 2026
Satya Nadella announced that Hayete Gallot is rejoining Microsoft as Executive Vice President, Security, reporting directly to him, after serving as President, Customer Experience for Google Cloud and previously spending over 15 years at Microsoft in senior roles. The security team will now report to Hayete, with Ales Holecek appointed Chief Architect for Security under her leadership. Charlie Bell will transition from leading the Security, Compliance, Identity, and Management organization to a new role focused on engineering quality, also reporting to Nadella and partnering with Scott Guthrie and Mala Anand on the Quality Excellence Initiative.
🔗 Source: Summary based on blogs.microsoft.com View Source | Found on Feb 04, 2026
Fundamental, an AI company founded in October 2024 by DeepMind alumni, emerged from stealth on February 5, 2026 with $255M in funding—$30M Seed and $225M Series A led by Oak HC/FT and joined by Valor Equity Partners, Battery Ventures, Salesforce Ventures, Hetz Ventures, and notable angel investors. The company launched NEXUS, a Large Tabular Model (LTM) for enterprise data prediction. Fundamental partnered with AWS to offer NEXUS directly to AWS customers and secured seven-figure contracts with Fortune 100 enterprises for predictive use cases such as demand forecasting, price prediction, and customer churn.
🔗 Source: Summary based on press.aboutamazon.com View Source | Found on Feb 05, 2026
At CES 2026, NVIDIA introduced a suite of open physical AI models and frameworks to accelerate development of humanoids, autonomous vehicles, and other physical AI systems. These tools include NVIDIA Cosmos world models, Isaac technologies such as the new Isaac Lab-Arena framework, Alpamayo portfolio for autonomous vehicles, and OSMO for training orchestration. Caterpillar’s Cat AI Assistant uses NVIDIA Nemotron models on Jetson Thor modules for natural language interaction in heavy vehicles. LEM Surgical’s FDA-cleared Dynamis Robotic Surgical System employs Jetson AGX Thor and Holoscan for spinal procedures. NEURA Robotics trains its 4NE1 humanoid using Isaac Sim and collaborates with SAP via Omniverse Blueprint integration.
🔗 Source: Summary based on blogs.nvidia.com View Source | Found on Jan 29, 2026
The research paper presents the first large-scale analysis of disempowerment potential in AI conversations, examining 1.5 million Claude.ai interactions from December 2025. Severe disempowerment—where AI extensively shapes users’ beliefs, values, or actions—was rare: reality distortion occurred in about 1 in 1,300 conversations, value judgment distortion in 1 in 2,100, and action distortion in 1 in 6,000. Amplifying factors such as user vulnerability (1 in 300), attachment (1 in 1,200), reliance (1 in 2,500), and authority projection (1 in 3,900) increased risk. Rates of moderate or severe disempowerment potential rose between late 2024 and late 2025.
🔗 Source: Summary based on anthropic.com View Source | Found on Jan 28, 2026
The article, published on February 5, 2026, discusses research that challenges the assumptions of many reinforcement learning (RL) models in psychology and neuroscience. The researchers used a hybrid modelling approach integrating artificial neural networks into interpretable cognitive architectures to estimate general forms for each algorithmic component and evaluate their necessity and sufficiency. Analyzing a large dataset of human reward-learning behaviour, they found that successful models require independent and flexible memory variables capable of tracking rich representations of past experiences. Their results question RL models based solely on incremental updating of scalar reward predictions.
🔗 Source: Summary based on deepmind.google View Source | Found on Feb 06, 2026
NVIDIA founder and CEO Jensen Huang and Dassault Systèmes CEO Pascal Daloz announced at 3DEXPERIENCE World in Houston a major partnership to build a shared industrial AI architecture, merging NVIDIA’s accelerated computing and AI libraries with Dassault Systèmes’ Virtual Twin platforms. This collaboration, the largest between the companies in over 25 years, aims to create physics-based “world models” for simulating products, factories, and biological systems before physical construction. The partnership integrates technologies such as NVIDIA BioNeMo, BIOVIA world models, SIMULIA AI-based Virtual Twin Physics Behavior, CUDA-X libraries, Omniverse physical AI libraries with DELMIA Virtual Twin, and the 3DEXPERIENCE agentic platform.
🔗 Source: Summary based on blogs.nvidia.com View Source | Found on Feb 03, 2026
On January 29, 2026, during the UK Prime Minister’s visit to China, LSEG and ICBC signed a Memorandum of Understanding in Beijing to strengthen long-term strategic cooperation across global financial markets, data and analytics, cross-border RMB, sustainable finance, and financial innovation. The agreement establishes a framework for expanded collaboration in capital markets, trading, clearing, data services, RMB business, and emerging technologies. Fiona Bassett represented LSEG while Zhang Weiwu signed on behalf of ICBC. Both institutions emphasized their commitment to advancing financial innovation and enhancing UK-China cooperation within the global financial market ecosystem.
🔗 Source: Summary based on lseg.com View Source | Found on Jan 30, 2026
The article explains that major technological advances, such as steam engines, electricity, internal combustion engines, and computers, only became transformative when their applications expanded beyond initial uses. Steam engines began as tools for pumping water in coal mines in the early 18th century but revolutionized productivity when used for railroads, which altered travel times and trade routes. Electricity initially replaced steam engines and gas lamps but reorganized factories with decentralized motors and reshaped cities through streetcars and subways. Internal combustion engines powered cars and airplanes, changing living patterns and enabling new industries. Computers drove significant productivity gains when applied to networked systems and digital marketplaces.
🔗 Source: Summary based on pictet.com View Source | Found on Feb 05, 2026
In 2025, AI-related capital expenditure accounted for an estimated 92% of US GDP growth in the first half of the year, with nonfarm business productivity rising 4.9% in Q3 and unit labor costs declining for two consecutive quarters. Despite these gains, long-term trends show a decline in research productivity: Apple’s research productivity fell from 1.92% to 0.04% gross profit growth per thousand R&Ders over two decades, while median research productivity growth across US firms was -10%. International studies confirm similar declines. Firms must increase R&D investment to maintain innovation output, highlighting challenges despite rapid AI adoption.
🔗 Source: Summary based on man.com View Source | Found on Feb 02, 2026
Karen Karniol-Tambour, Greg Jensen, and Bob Prince discuss the new investment regime characterized by artificial intelligence and modern mercantilism. Karen outlines the current investment environment, highlights potential risks ahead, and identifies key questions that will shape future developments. Greg explains their approach to understanding AI and emphasizes that the next phase of the AI boom presents increased dangers. Bob provides strategies for investors to construct resilient portfolios in this context. These insights were shared during an excerpt from their Q4 CIO call, focusing on navigating today’s evolving financial landscape.
🔗 Source: Summary based on bridgewater.com View Source | Found on Feb 03, 2026
GMO’s Jeremy Grantham and Edward Chancellor argue the US equity market is in a two-sigma “bubble,” turbocharged by AI enthusiasm after ChatGPT arrested the 2022 selloff. They frame AI as a classic, highly visible general-purpose innovation that invites “new era” narratives, easy-money extrapolation, and a temporary suspension of valuation discipline. Chancellor maps today’s setup onto historical technology manias: exuberant storytelling, leverage and retail speculation, media “puffery,” frothy private funding, immature tech and uncertain monetization, and—most importantly—massive capital overcommitment (hyperscaler capex) that can compress future returns. They expect an eventual shakeout with falling profits and multiples.
🔗 Source: Summary based on gmo.com View Source | Found on Jan 29, 2026