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
OpenAI and Amazon announced a multi-year strategic partnership on February 27, 2026, to accelerate AI innovation. Amazon will invest $50 billion in OpenAI, starting with $15 billion and an additional $35 billion pending certain conditions. AWS will be the exclusive third-party cloud distribution provider for OpenAI Frontier, enabling organizations to build and manage teams of AI agents. OpenAI will consume approximately 2 gigawatts of Trainium capacity through AWS infrastructure as part of an expanded $138 billion agreement over eight years. The partnership includes co-developing a Stateful Runtime Environment powered by OpenAI models, launching soon on Amazon Bedrock.
π Source: Summary based on press.aboutamazon.com View Source | Found on Feb 27, 2026
Since 2019, Microsoft and OpenAI have maintained a strong partnership focused on responsible AI advancement, with Microsoft holding an exclusive license and access to OpenAI’s intellectual property across models and products. The terms of their relationship, including commercial and revenue share arrangements, remain unchanged as of the February 27, 2026 announcement. Azure continues as the exclusive cloud provider for stateless OpenAI APIs, including those resulting from collaborations with third parties such as Amazon. OpenAI’s first-party products like Frontier will remain hosted on Azure. The contractual definition and processes regarding AGI are also unchanged under this partnership.
π Source: Summary based on blogs.microsoft.com View Source | Found on Feb 27, 2026
Meta announced a multi-year agreement with AMD to power its AI infrastructure using up to 6GW of AMD Instinct GPUs, supporting modern AI models. Shipments for the first GPU deployments will begin in the second half of 2026, utilizing the Helios rack-scale architecture developed in collaboration with AMD. The partnership includes alignment across silicon, systems, and software enabling vertical integration within Meta’s infrastructure stack. This collaboration is part of Meta’s Compute initiative to scale infrastructure for personal superintelligence and diversify its technology stack alongside its Meta Training and Inference Accelerator (MTIA) silicon program.
π Source: Summary based on about.fb.com View Source | Found on Feb 24, 2026
On February 26, 2026, Huawei Cloud held an online launch event themed “AI Coding Practitioners, Full Code Power,” announcing the public beta release of CodeArts, an AI-powered coding solution for developers and enterprises. CodeArts integrates code large models, IDEs (including VS Code and JetBrains), self-developed modes, and supports HarmonyOS’s ArkTS language. It features code generation, knowledge Q&A, unit test case generation, expert skills indexing, compliance-driven development, and connects to GLM-5.0 and DeepSeek-V3.2 models as well as proprietary ones. The platform claims a 30% token saving per task through efficient context understanding and offers free public beta access with comprehensive tutorials on its website.
π Source: Summary based on huaweicloud.com View Source | Found on Feb 26, 2026
Secretary of War Pete Hegseth announced on X that the Department of War will designate Anthropic as a supply chain risk, following stalled negotiations over two exceptions requested by Anthropic: mass domestic surveillance of Americans and fully autonomous weapons. Anthropic has supported American warfighters since June 2024 and maintains its stance against these uses, citing reliability concerns and fundamental rights violations. The designation, unprecedented for an American company, would only affect Department of War contract work under 10 USC 3252; individual customers and commercial contracts remain unaffected. Anthropic intends to challenge the designation in court and continues to support its users.
π Source: Summary based on anthropic.com View Source | Found on Feb 28, 2026
Anthropic announced the acquisition of Vercept, whose co-founders are Kiana Ehsani, Luca Weihs, and Ross Girshick, to enhance its AI’s ability to perform complex tasks in live applications. Vercept will discontinue its external product and integrate with Anthropic. This follows the launch of Claude Sonnet 4.6, which improved computer use skills on OSWorld from under 15% in late 2024 to 72.5%. Sonnet 4.6 now approaches human-level performance on tasks such as navigating complex spreadsheets and completing web forms across browser tabs. Vercept is Anthropic’s latest acquisition after Bun.
π Source: Summary based on anthropic.com View Source | Found on Feb 25, 2026
The article, authored by Alyssa Hughes and published on February 26, 2026, presents CORPGEN, an agent framework developed to simulate corporate environments with autonomous digital employees in Multi-Horizon Task Environments (MHTEs), where each task requires 10–30 dependent steps within five-hour sessions. Testing three agent systems under increasing loads from 12 to 46 tasks showed baseline completion rates dropping from 16.7% to 8.7%, while CORPGEN maintained or improved performance, completing 15.2% of tasks at the highest load compared to baselines’ 4.3%. The research was a collaboration between Microsoft’s CTO Office and MAIDAP, with contributions from several Microsoft teams and individuals.
π Source: Summary based on microsoft.com View Source | Found on Feb 27, 2026
NVIDIA reported record fourth-quarter revenue of $68.1 billion, up 20% from the previous quarter and 73% year-over-year, with fiscal 2026 revenue reaching $215.9 billion, a 65% increase. Fourth-quarter GAAP and non-GAAP gross margins were 75.0% and 75.2%, while fiscal margins were 71.1% and 71.3%. Quarterly GAAP and non-GAAP earnings per diluted share were $1.76 and $1.62; for the year, $4.90 and $4.77 respectively. NVIDIA returned $41.1 billion to shareholders in fiscal 2026, with $58.5 billion remaining under repurchase authorization, and announced a quarterly dividend of $0.01 per share payable April 1, 2026.
π Source: Summary based on nvidianews.nvidia.com View Source | Found on Feb 25, 2026
In 2025, the S&P 500 Index gained 17.9%, driven largely by AI-related earnings and expectations, with the Communication Services sector returning 33.6% and Information Technology 24.0%. The ‘Magnificent Seven’ stocks—Alphabet, Apple, Amazon, Meta Platforms, Microsoft, Nvidia, and Tesla—led massive AI investments; Alphabet alone spent an estimated USD 175 billion. Despite nearly 75% of surveyed businesses using AI in some way, 86% reported virtually no productivity improvement. Consumer staples also performed strongly with a sector gain of 15.8%, and Walmart’s market capitalization reached USD 1 trillion in early February.
π Source: Summary based on gam.com View Source | Found on Feb 27, 2026
Generate has identified a monoclonal antibody in phase I trial targeting the SARS-CoV-2 spike protein and has dosed the first patient in a study for mild-to-moderate asthma using another monoclonal antibody. The company’s AI platform can generate de novo antibodies, demonstrated across nine distinct targets, and last year published full details of its generative AI model Chroma in Nature, releasing its code as open-source. Following this, Generate received 2,000 applications for five machine learning internships. The company now routinely analyzes datasets of a million variants per target in one quarter the historical time required.
π Source: Summary based on pictet.com View Source | Found on Feb 27, 2026
The article by Kunihiro Miyazaki, Takanobu Kawahara, Stephen Roberts, and Stefan Zohren proposes a multi-agent large language model (LLM) trading framework that decomposes investment analysis into fine-grained tasks rather than relying on coarse-grained instructions. The system was evaluated using Japanese stock data—including prices, financial statements, news, and macro information—under a leakage-controlled backtesting setting. Experimental results demonstrate that fine-grained task decomposition significantly improves risk-adjusted returns compared to conventional designs. Further analysis indicates that alignment between analytical outputs and downstream decision preferences is crucial for performance. Standard portfolio optimization exploiting low correlation and variance achieves superior results.
π Source: Summary based on arxiv.org View Source | Found on Feb 27, 2026
The paper titled "Could Large Language Models work as Post-hoc Explainability Tools in Credit Risk Models?" by Wenxi Geng, Dingyuan Liu, Liya Li, and Yiqing Wang investigates the use of large language models (LLMs) as post-hoc explainability tools for credit risk predictions. Using a LendingClub personal lending dataset, three commercial LLMs—GPT-4-turbo, Claude Sonnet 4, and Gemini-2.0-Flash—were evaluated. The results show strong evidence for LLMs functioning effectively as translators of model outputs but low alignment when providing autonomous explanations. Few-shot prompting improved feature overlap for logistic regression but not consistently for XGBoost.
π Source: Summary based on arxiv.org View Source | Found on Feb 24, 2026
The article introduces AlphaForgeBench, a framework developed by Wentao Zhang and colleagues to address limitations in evaluating Large Language Model (LLM)-based trading agents. The authors identify that current benchmarks suffer from severe behavioral instability, including high run-to-run variance and irrational action flipping, due to stateless autoregressive architectures and sensitivity in action mappings. AlphaForgeBench reframes LLMs as quantitative researchers who generate executable alpha factors and strategies rather than direct trading actions. This approach decouples reasoning from execution, enabling deterministic and reproducible evaluations aligned with real-world quantitative research workflows, as demonstrated across multiple state-of-the-art LLMs.
π Source: Summary based on arxiv.org View Source | Found on Feb 24, 2026
This study by Krishna Neupane, Prem Sapkota, and Ujjwal Prajapati establishes the causal effects of market sentiment on bank profitability using a causal forest machine learning methodology that controls for confounding variables and analyzes heterogeneity and non-linearities. The research utilizes FinancialBERT to generate sentiment scores from quarterly reports as causal interventions affecting profitability metrics such as returns and volatilities. Drawing on data from NEPSE, NRB, and individual financial institutions, SHAP analysis identifies influential profit predictors. The findings reveal statistically significant causal associations between balance sheet strength, expense management variables, and profitability through Average Treatment Effect analyses.
π Source: Summary based on arxiv.org View Source | Found on Feb 23, 2026
The article by Amine Kobeissi and Philippe Langlais investigates retrieval failures in retrieval-augmented generation (RAG) for long-document financial question answering, focusing on cases where the correct document is retrieved but the specific page or chunk containing the answer is missed. The authors evaluate multiple retrieval strategies—dense, sparse, hybrid, and hierarchical with reranking and query reformulation—on a 150-question subset of FinanceBench. They introduce an oracle-based analysis to establish empirical upper bounds and present a domain fine-tuned page scorer using a bi-encoder for page-level relevance in financial filings, resulting in significant improvements in page recall and chunk retrieval.
π Source: Summary based on arxiv.org View Source | Found on Feb 23, 2026
The article presents a hybrid framework for aspect-based sentiment analysis that leverages large language models (LLMs) for aspect identification and traditional machine-learning methods for scalable sentiment classification. The authors, Vishal Patil, Shree Vaishnavi Bacha, Revanth Yamani, Yidan Sun, and Mayank Kejriwal, used ChatGPT to analyze sampled restaurant reviews and developed sentiment classifiers with human-labeled data. This approach was applied to 4.7 million reviews spanning 17 years from a major online platform. Regression analysis showed that machine-labeled aspects significantly explained variance in overall restaurant ratings across various dining experiences, cuisines, and regions.
π Source: Summary based on arxiv.org View Source | Found on Feb 25, 2026
The paper by Szymon Lis, Robert Εlepaczuk, and PaweΕ Sakowski examines whether short-term market overreactions in Apple Inc. (AAPL) stocks can be predicted and monetized as momentum signals using high-frequency emotional information from Twitter and machine learning methods. The study constructs an intraday dataset combining volatility-normalized returns with transformer-based emotion features, defines overreactions as extreme return realizations relative to volatility and transaction costs, and models them as a three-class prediction problem. Machine learning classifiers significantly outperform benchmark rules at ultra-short horizons, while classical behavioral momentum effects dominate at 10-minute frequencies. SHAP analysis highlights volatility and negative emotions as key drivers of predicted overreactions.
π Source: Summary based on arxiv.org View Source | Found on Feb 24, 2026
The study by Baris Arat, Hasan Fehmi Ates, and Emre Sefer investigates forecasting weekly sub-city real estate price indices in Dubai using over 350,000 transactions from the Dubai Land Department (2015-2025). The authors construct indices for 19 regions and evaluate forecasts from 2 to 34 weeks ahead by integrating transaction history, Sentinel-1 SAR backscatter, news sentiment (lexical tone and semantic embeddings), and macroeconomic context. Results show that at horizons up to 10 weeks, price history suffices; beyond 14 weeks, sentiment and SAR are essential. At long horizons (26-34 weeks), the multimodal model reduces mean absolute error from 4.48 to 2.93—a statistically significant 35% reduction.
π Source: Summary based on arxiv.org View Source | Found on Feb 24, 2026