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 by Matt Wilson, submitted on 11 Dec 2025, establishes a precise correspondence between decision-making agents in partially observable Markov decision processes (POMDPs) and one-input process functions, which are the classical limit of higher-order quantum operations. The work identifies that an agent's policy and memory update combine into a process function w interacting with a POMDP environment via the link product. It presents dual interpretations: in physics, the process function acts as the environment for agent interventions; in AI, it encodes the agent while inserted functions represent environments. The perspective is extended to multi-agent systems through observation-independent decentralized POMDPs as domains for multi-input process functions.
π Source: arxiv.org View Source | Found on Dec 13, 2025
The article by Anna Perekhodko and Robert Εlepaczuk presents a hybrid modelling framework that combines a Stochastic Volatility (SV) model with a Long Short Term Memory (LSTM) neural network to forecast S&P 500 index volatility using daily data from January 1, 1998, to December 31, 2024. The SV model captures latent volatility dynamics and statistical precision, while the LSTM network detects complex nonlinear patterns. A rolling window approach is used for training and one-step-ahead forecasting. Statistical testing and investment simulations demonstrate that the hybrid SV-LSTM model outperforms standalone SV and LSTM models in volatility forecasting.
π Source: arxiv.org View Source | Found on Dec 17, 2025
This study by Pablo Hidalgo, Julio E. Sandubete, and Agustín García-García examines the predictive performance of Multilayer Perceptrons (MLP) and Long Short-Term Memory (LSTM) neural networks on economic time series using Intrinsic Mode Functions (IMFs). DeepSHAP is employed to estimate the marginal contribution of each IMF. Findings indicate that long-term trend IMFs are most influential for prediction, while high-frequency IMFs contribute less and may introduce noise, as shown by improved metrics when they are removed. The study also reveals that LSTM models distribute feature importance more evenly across IMFs compared to MLPs.
π Source: arxiv.org View Source | Found on Dec 17, 2025
This study applies Empirical Mode Decomposition (EMD) to the MSCI World index, extracting nine intrinsic mode functions (IMFs) using CEEMDAN. Each IMF is converted into a graph via four methods: natural visibility, horizontal visibility, recurrence, and transition graphs. Topological analysis reveals that high-frequency IMFs form dense small-world graphs, while low-frequency IMFs result in sparser networks with longer path lengths. Visibility-based methods are more sensitive to amplitude variability and yield higher clustering; recurrence graphs better preserve temporal dependencies. These findings inform the design of graph neural network architectures for financial time series modeling.
π Source: arxiv.org View Source | Found on Dec 17, 2025
The article "Local and Global Balance in Financial Correlation Networks: an Application to Investment Decisions" by Paolo Bartesaghi, Rosanna Grassi, and Pierpaolo Uberti introduces the concept of local balance as a measure of a single node's contribution to the overall balance in signed financial correlation networks. The authors investigate using deviations between local and global balance as a criterion for selecting outperforming assets, particularly during financial crises when most assets behave similarly and global balance peaks. Their research, supported by real financial data, confirms that focusing on assets with significant local-global balance deviations can improve investment decisions.
π Source: arxiv.org View Source | Found on Dec 13, 2025
The article presents a rigorous walk-forward validation framework for algorithmic trading, integrating interpretable hypothesis-driven signal generation, reinforcement learning, and strict out-of-sample testing. The methodology applies rolling window validation across 34 independent test periods and evaluates five market microstructure patterns on 100 US equities from 2015 to 2024. Results show modest annualized returns of 0.55% (Sharpe ratio 0.33), maximum drawdown of -2.76%, and market-neutral beta of 0.058, with performance dependent on market regime—positive in high-volatility periods (0.60% quarterly, 2020-2024) but negative in stable markets (-0.16%, 2015-2019). Aggregate results are statistically insignificant (p-value 0.34).
π Source: arxiv.org View Source | Found on Dec 17, 2025
The article "Model-First Reasoning LLM Agents: Reducing Hallucinations through Explicit Problem Modeling" by Annu Rana and Gaurav Kumar introduces Model-First Reasoning (MFR), a two-phase approach where large language models first construct an explicit model of the problem—defining entities, state variables, actions, and constraints—before generating a solution plan. Tested across domains such as medical scheduling, route planning, resource allocation, logic puzzles, and procedural synthesis, MFR reduces constraint violations and improves solution quality compared to Chain-of-Thought and ReAct methods. Ablation studies confirm the critical role of explicit modeling in these improvements. All prompts, evaluation procedures, and datasets are documented for reproducibility.
π Source: arxiv.org View Source | Found on Dec 18, 2025
The article, authored by Tetsuya Takaishi and submitted on 11 December 2025, investigates the use of single-qubit quantum circuit learning (QCL) for modeling volatility time series. Synthetic data generated with the Rational GARCH model, which captures volatility asymmetry, was used to evaluate QCL’s effectiveness. The study found that QCL-based predictions preserved the negative return-volatility correlation characteristic of asymmetric volatility dynamics. Additionally, analysis showed that the predicted series exhibited anti-persistent behavior and maintained multifractal structure, similar to the original synthetic data. The paper spans 9 pages with 10 figures and was accepted for the 14th International Conference on Mathematical Modeling in Physical Sciences.
π Source: arxiv.org View Source | Found on Dec 13, 2025
Oaktree argues recent stress in sub-IG credit is systematic, not systemic—a recurring product of late-cycle complacency, not broken “plumbing,” but a clear call for renewed credit discipline. Direct lenders should stay selective as abundant dry powder compresses middle-market spreads despite improving M&A pipelines. European direct lending is still in early growth, with bank retrenchment, large PE dry powder and higher public spending driving structural demand. The U.S. high yield market is higher quality than headline spreads suggest, with a record BB share and low defaults. Finally, EM equities have led global markets in 2025, supported by commodities, a weaker dollar and stronger balance sheets.
π Source: oaktreecapital.com View Source | Found on Dec 18, 2025
GMO warns that AI enthusiasm is driving a return to bubble-like equity pricing. Over 30% of U.S. market cap now trades above 10x sales, a concentration heavily driven by the “Magnificent 7,” but excess extends well beyond them, with roughly 8% of all U.S. stocks above 10x sales—levels close to 2000 and the 2020 growth bubble. GMO argues that even dominant firms can disappoint when expectations are extreme, while weaker names have been bid up on speculation, not fundamentals. They conclude today’s valuation extremes embed significant downside risk and call for renewed discipline in fundamentals and valuation.
π Source: gmo.com View Source | Found on Dec 18, 2025
Correlations among major hyperscalers have reached their lowest levels since the launch of ChatGPT, as investors shift focus within the AI trade and leadership moves to CapEx recipients. Government shutdowns have disrupted data availability, especially for labor markets, which are experiencing linear cooling and downside risks. Despite concerns about sticky inflation, wage growth is softening and shelter costs are disinflating, reducing persistent inflationary pressures. For 2026, tax refunds are expected to surge due to provisions in the One Big, Beautiful Bill; however, tepid consumer confidence may limit consumption gains. Optimism exists for a renewed CapEx cycle from restored tax policies but growth remains modest.
π Source: im.natixis.com View Source | Found on Dec 19, 2025
Neuberger Berman’s “Solving for 2026” highlights a moderate-growth but highly divergent outlook driven by policy cross-currents and AI. The U.S. and parts of Asia benefit from productivity gains and supportive policy, while Europe/UK face weaker growth and deeper cuts. AI is a multi-year driver, with returns shifting from enablers to high-quality adopters. Equities: favour US, China, Japan and EM Asia, emphasizing quality and clear monetisation. Fixed income: extend duration and use multi-sector, selective credit. Private markets: lean into secondaries/co-investments and AI-adjacent assets as liquidity providers. Key risks include hawkish policy error, AI capex disappointment, geopolitics, power constraints and credit shocks.
π Source: nb.com View Source | Found on Dec 16, 2025
Last week, stock markets declined despite the Federal Reserve cutting interest rates by 25 basis points, with two hawkish dissents (no change) and one dovish dissent (50 bps cut). The Fed also announced Reserve Management Purchases, committing to buy USD 40 billion in Treasury bills monthly from December 12 to increase banks’ reserves. The Swiss National Bank held rates at 0% and is expected to remain on hold through 2026; the Reserve Bank of Australia also held. The S&P 500 fell 0.6% as investors rotated out of AI-related tech stocks following disappointing results from Broadcom and Oracle Corp.
π Source: pictet.com View Source | Found on Dec 15, 2025
The paper contrasts human and robotic movement, arguing they are built on fundamentally different architectures and control logics. Human motion relies on compliant, self-repairing musculoskeletal tissue tightly coupled to predictive neural control, delivering stability, agility and extreme energy efficiency. Robots, in contrast, are typically rigid kinematic chains driven by discrete actuators and modular control, precise but brittle and power-hungry. Emerging trends—soft robotics, magnetic “muscles,” variable-stiffness joints and LLM-enabled controllers—are pushing towards partial convergence, with robots excelling in extreme or hazardous environments but still far behind in everyday agility and embodied “meaningful” action. Moravec’s paradox underlines why low-level sensorimotor skills remain the hardest frontier.
π Source: lansdownepartners.com View Source | Found on Dec 18, 2025
LSEG and Citi have entered a multi-year strategic partnership to deploy LSEG’s data, analytics, and workflow solutions at enterprise scale across Citi’s operations. The agreement will support Citi’s workflows in markets, investment banking, wealth, trading, risk, finance, and compliance by consolidating data access and standardizing governance. LSEG will provide AI-ready content including economic indicators, pricing data, company information, benchmarks, indices, fund data, commodities news, risk-intelligence and regulatory data. The partnership includes access to LSEG Workspace and APIs for real-time and historical content delivery while integrating World-Check risk-intelligence to enhance Citi’s compliance frameworks.
π Source: lseg.com View Source | Found on Dec 16, 2025
The NVIDIA RTX PRO 5000 72GB Blackwell GPU, now generally available as of December 18, 2025, offers robust agentic and generative AI capabilities with 2,142 TOPS of AI performance and 72GB of GDDR7 memory—a 50% increase over the previous 48GB model. This GPU delivers up to 3.5x image generation and 2x text generation performance compared to prior-generation hardware, slashes render times by up to 4.7x in engines like Arnold and Blender, and provides more than double graphics performance for engineering tasks. Early adopters include InfinitForm and Versatile Media, with availability from partners such as Ingram Micro and Leadtek.
π Source: blogs.nvidia.com View Source | Found on Dec 18, 2025
T5Gemma 2, published by Biao Zhang on December 18, 2025, is an advanced encoder-decoder model based on Gemma 3. It introduces tied word embeddings and merged decoder self- and cross-attention to reduce parameters, offering compact pre-trained models at sizes of 270M-270M (~370M total), 1B-1B (~1.7B), and 4B-4B (~7B) parameters. The model supports multimodal processing with a vision encoder for visual question answering, handles context windows up to 128K tokens using alternating local and global attention, and is trained on data covering over 140 languages.
π Source: blog.google View Source | Found on Dec 18, 2025