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.
The article "Forward-Oriented Causal Observables for Non-Stationary Financial Markets" by Lucas A. Souza, submitted on 31 December 2025, investigates short-horizon forecasting in financial time series under strict causal constraints. The study constructs an interpretable causal signal from micro-features such as momentum, volume pressure, trend acceleration, and volatility-normalized price location using causal centering, linear aggregation, a one-dimensional Kalman filter for stabilization, and an adaptive forward-like operator. Application to high-frequency EURUSDT (1-minute) data demonstrates that these causally constructed observables can yield substantial economic relevance in specific regimes but may degrade following regime shifts.
🔗 Source: arxiv.org View Source | Found on Jan 01, 2026
The article "Empowering Small Language Models with Factual Hallucination-Aware Reasoning for Financial Classification" by Han Yuan, Yilin Wu, Li Zhang, and Zheng Ma introduces a three-step AAAI pipeline—Association Identification, Automated Detection, and Adaptive Inference—to address factual hallucinations in small language models (SLMs) used for financial classification. Experiments on three representative SLMs demonstrate that factual hallucinations are positively correlated with misclassifications; encoder-based verifiers can effectively detect these hallucinations; and incorporating feedback on factual errors enables adaptive inference that improves classification performance. The authors aim to enhance the trustworthiness and effectiveness of SLMs in finance.
🔗 Source: arxiv.org View Source | Found on Jan 06, 2026
Uni-FinLLM is a unified multimodal large language model developed by Gongao Zhang, Haijiang Zeng, and Lu Jiang to jointly process financial text, numerical time series, fundamentals, and visual data for micro-level stock prediction and macro-level systemic risk assessment. The model employs a shared Transformer backbone with modular task heads and utilizes cross-modal attention and multi-task optimization. Uni-FinLLM achieves 67.4% stock directional accuracy (up from 61.7%), 84.1% credit-risk accuracy (up from 79.6%), and 82.3% macro early-warning accuracy, outperforming existing baselines in finance-related prediction tasks.
🔗 Source: arxiv.org View Source | Found on Jan 07, 2026
This paper by Mariluz Mate examines the challenges of defining and interpreting causal effects in environments where firms interact through spatial or network connections. It argues that under interdependence, causal effects are not uniquely defined, even with correctly specified or learned interaction structures and ideal identifying conditions. The author develops a framework for firm-level economies with unobserved but learnable interaction structures and formalizes three counterfactual regimes: partial equilibrium, local interaction, and network consistent equilibrium. The study shows that standard spatial autoregressive estimates correspond to different causal effects depending on the chosen counterfactual regime and derives identification conditions for each.
🔗 Source: arxiv.org View Source | Found on Jan 05, 2026
The article introduces an operator-theoretic framework for causal analysis in multivariate time series, defining directional influence via sensitivity of second-order dependence operators to order-preserving temporal deformations. Under linear Gaussian assumptions, the method aligns with linear Granger causality but extends to capture collective and nonlinear directional dependence beyond pairwise predictability. The framework establishes existence, uniform consistency, and valid inference for non-smooth supremum--infimum statistics using shift-based randomization. Simulations show correct size and strong power against distributed and bulk-dominated alternatives. An empirical application to high-dimensional daily financial return series demonstrates episodic, stress-dependent causal organization with sparse, horizon-heterogeneous transmission channels.
🔗 Source: arxiv.org View Source | Found on Jan 06, 2026
Meta announced agreements with Oklo, TerraPower, and Vistra to support clean, reliable nuclear energy for its operations. The partnership with TerraPower includes funding for two new Natrium units generating up to 690 MW by 2032 and rights to energy from six additional units totaling 2.1 GW by 2035, for a potential total of eight units with 2.8 GW capacity and 1.2 GW storage. The Oklo agreement supports development of a nuclear campus in Pike County, Ohio, potentially adding up to 1.2 GW by 2030. With Vistra, Meta will purchase over 2.1 GW from three plants and support uprates totaling 433 MW in the early 2030s.
🔗 Source: about.fb.com View Source | Found on Jan 09, 2026
NVIDIA has released new open models, data, and tools to advance AI across industries, including the Nemotron family for agentic AI, Cosmos platform for physical AI, Alpamayo for autonomous vehicles, Isaac GR00T for robotics, and Clara for biomedical applications. The company provides open-source training frameworks and vast datasets: 10 trillion language tokens, 500,000 robotics trajectories, 455,000 protein structures, and 100 terabytes of vehicle sensor data. Leading firms such as Bosch, ServiceNow, CrowdStrike, Palantir, Salesforce and Uber are adopting these technologies. Models and resources are available on GitHub and Hugging Face for flexible developer access.
🔗 Source: blogs.nvidia.com View Source | Found on Jan 05, 2026
In the second half of 2025, global adoption of generative AI tools rose by 1.2 percentage points, with about one in six people worldwide now using such tools. The Global North saw faster growth than the Global South, resulting in 24.7% usage among its working age population versus 14.1% in the Global South. The UAE led globally with 64.0% adoption, ahead of Singapore at 60.9%. The U.S., despite leading in infrastructure and model development, ranked 24th with a 28.3% usage rate. DeepSeek’s open-source platform gained traction across China, Russia, Iran, Cuba, Belarus, and Africa through partnerships like Huawei’s.
🔗 Source: blogs.microsoft.com View Source | Found on Jan 08, 2026
NVIDIA founder and CEO Jensen Huang opened CES 2026 in Las Vegas by unveiling Rubin, NVIDIA’s first extreme-codesigned, six-chip AI platform now in full production, which reduces token generation costs to one-tenth of the previous platform. Huang introduced Alpamayo, an open reasoning model family for autonomous vehicles, with the first passenger car featuring Alpamayo—built on NVIDIA DRIVE—launching soon in the Mercedes-Benz CLA. NVIDIA also announced DLSS 4.5 with Dynamic Multi Frame Generation and support for over 250 games and apps, new GeForce NOW Apps for Linux PC and Amazon Fire TV, and G-SYNC Pulsar monitors available this week.
🔗 Source: blogs.nvidia.com View Source | Found on Jan 06, 2026
On January 5, 2026, Microsoft announced the acquisition of Osmos, an agentic AI data engineering platform aimed at simplifying complex and time-consuming data workflows. Osmos utilizes agentic AI to convert raw data into analytics and AI-ready assets within OneLake, the unified data lake central to Microsoft Fabric. The Osmos team will join Microsoft’s Fabric engineering organization to further develop simpler and more intuitive AI-ready data experiences. This acquisition supports Microsoft Fabric’s objective of unifying all data and analytics into a single secure platform for organizations.
🔗 Source: blogs.microsoft.com View Source | Found on Jan 05, 2026
Over the next decade, AI is expected to boost US productivity growth from 1.60% to 1.85% annually, with direct impacts averaging 0.15 percentage points and indirect effects around 0.1 percentage points per year. Despite negative policy trends shaving -0.15% off potential growth and less favorable demographics due to an aging population and falling immigration, overall US potential growth is projected to average 2.1%, reaching up to 2.4% by 2035 as AI adoption rises, especially in the service sector. The US focuses on closed AI models for proprietary ecosystems, while China pursues open-weight strategies for broader accessibility and industrial integration.
🔗 Source: lombardodier.com View Source | Found on Jan 07, 2026
Apollo-managed funds and affiliates led a $3.5 billion capital solution for Valor Compute Infrastructure L.P. (VCI), managed by Valor Equity Partners, to support VCI’s $5.4 billion acquisition and lease of data center compute infrastructure, including NVIDIA GB200 GPUs, to a subsidiary of xAI Corp for model training and Grok development. The financing uses a triple net lease structure. NVIDIA invested in VCI as an anchor Limited Partner alongside other institutional investors. As of September 30, 2025, Apollo had approximately $908 billion in assets under management; Valor had about $55 billion as of December 31, 2025.
🔗 Source: apollo.com View Source | Found on Jan 07, 2026