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 paper introduces ECLIPSE, an information-theoretic framework treating hallucination as a mismatch between semantic entropy and evidence capacity. It combines multi-sample clustering with a novel perplexity decomposition measuring evidence use, and proves, under mild conditions, a strictly convex entropy‑capacity objective with a unique stable optimum. On a controlled financial QA dataset (GPT‑3.5‑turbo, n=200 with synthetic hallucinations), ECLIPSE achieved ROC AUC 0.89 and average precision 0.90, outperforming a semantic entropy-only baseline (AUC 0.50). An ablation with Claude‑3‑Haiku, which lacks token-level log probabilities, dropped AUC to 0.59 and coefficient magnitudes by 95%; perplexity decomposition features had the largest learned coefficients.
🔗 Source: View Source | Found on Dec 05, 2025
Zhiming Lian’s paper evaluates Qwen3-8B for financial sentiment analysis and financial news classification. Using Noisy Embedding Instruction Finetuning, rLoRA (rank-stabilized low-rank adaptation), and FlashAttention to allow faster training with lower GPU memory, the study benchmarks against T5, BERT, RoBERTa, LLaMA1‑7B, LLaMA2‑7B, and Baichuan2‑7B. Qwen3‑8B consistently achieves higher classification accuracy and requires fewer training epochs. The synergy suggests it can potentially serve as a scalable, economical option for real-time financial NLP. Accepted to DECS 2025, pending publication in the ACM International Conference Proceeding Series.
🔗 Source: View Source | Found on Dec 03, 2025
Jie Zhang, Dongrui Liu, and Jing Shao proposes a unified framework to stress-test adaptive and procedural autonomous trading agents. TradeTrap targets market intelligence, strategy formulation, portfolio and ledger handling, and trade execution, evaluating robustness under controlled system-level perturbations. Using closed-loop historical backtesting on real US equity data with identical initial conditions, experiments show small perturbations can induce extreme concentration, runaway exposure, and large drawdowns. Code available; cite arXiv:2512.02261; DOI 10.48550/arXiv.2512.02261 via DataCite (pending registration).
🔗 Source: View Source | Found on Dec 04, 2025
Weikang Zhang and Alison Watts run 480 million Monte Carlo simulations across 378 non-stablecoin crypto assets, net of trading fees and the opportunity cost of 1-month Treasury bills. At 2-3 years, median excess return is -28.4%, the 1% conditional value at risk wipes out principal, and only the top quartile averages 1,326.7% excess returns. A Bayesian multi-horizon local projection shows realized risk-return predictors negligible, while the 24-week Fear and Greed EMA indicates a one-standard-deviation sentiment shock reduces top-quartile mean excess returns by 15-22 percentage points and median by 6-10 over 1-3 years.
🔗 Source: View Source | Found on Dec 04, 2025
Rory Milsom’s 56-page, 8-figure University of Edinburgh dissertation extends TheAgentCompany with a finance-focused environment to test whether general-purpose LLM agents can perform wealth‑management tasks accurately and economically. It introduces synthetic domain data, richer colleague simulations, and an automatic task-generation pipeline, and builds a benchmark of 12 task‑pairs across retrieval, analysis, and synthesis/communication with explicit acceptance criteria and deterministic graders. Each task has high- and low-autonomy variants using seeded finance-specific data. Findings show limitations stem from end-to-end workflow reliability more than mathematical reasoning, autonomy significantly affects outcomes, and flawed model evaluation has hindered benchmarking.
🔗 Source: View Source | Found on Dec 04, 2025
Juan C. King and Jose M. Amigo present a 24-page study (7 figures, 2 tables) integrating LSTM networks with decision-tree models (Random Forest, Gradient Boosting) for stock trading. LSTMs analyze asset price patterns, while tree models use company economic data. Numerical simulations on international companies with 10-weekday predictions show that combining fundamental and technical variables can outperform approaches that do not. Among decision trees, Random Forest performed best. The authors also report that selecting technical variables can further boost prediction performance. The work is referenced as Forecasting 2025, 7(3), 49 (DOI: 10.48550/arXiv.2512.02036).
🔗 Source: View Source | Found on Dec 04, 2025
Marek Adamczyk and MichaÅ‚ DÄ…browski study pairs trading by replicating an asset via risk factors from PCA, ETFs, and LSTM networks, then trading on mean-reverting residuals modeled with an Ornstein–Uhlenbeck process. They adapt Avellaneda and Lee (2008) to Poland, using WIG20, mWIG40, and sector indices, with local risk-free rates and transaction costs. Evaluating 2017–2019 and 2020, all methods profit in 2017–2019; PCA delivers ~20% cumulative return and an annualized Sharpe up to 2.63. In the COVID-19 recession, only the ETF approach remains profitable (~5% annual), while PCA underperforms and LSTM results are negative.
🔗 Source: View Source | Found on Dec 04, 2025
Hui Gong, Akash Sharma, and Francesca Medda present “The Quantum Network of Assets,” a density‑matrix framework (QNA) embedding cross‑asset dependencies using density operators, entropy, and mutual information without assuming physical quantum effects. They define ERI and QEWS. Using NASDAQ‑100 data from 2024–2025, quantum entropy shows smoother evolution and clearer regime distinctions than classical entropy; ERI rises during structural tightening despite low volatility. Around the 2025 US tariff announcement, QEWS increases pre‑event then collapses post‑announcement, indicating structural transitions can precede price movements. The 26‑page paper includes 3 figures and 4 tables (arXiv:2511.21515).
🔗 Source: View Source | Found on Nov 28, 2025
The piece argues that recent AI-driven market volatility masks a much more durable, multiyear capital cycle. Rather than a single boom/bust “dot-com” replay, AI’s buildout is layered across cash-generative incumbents, capital-intensive new models and a broader creditor base, shifting focus from ROE to ROIC. Phase one (infrastructure) is already supporting growth and productivity, with adoption (software/services) and outcomes (proven customer benefits) still ahead. For investors, the key is to choose where to sit in the capital stack, stay diversified across equity and credit, prioritize balance-sheet strength and capital discipline, and use valuation frameworks that fully reflect rising capital intensity.
🔗 Source: View Source | Found on Dec 01, 2025
Bridgewater’s Greg Jensen and Jas Sekhon say Google’s Gemini 3 is the best publicly available model, delivering the biggest capability jump since at least OpenAI’s o3 (possibly o1, late 2024) and outperforming models across text, vision, and video. They estimate Gemini 3 used at least 2-3 times, possibly an order of magnitude, more pre-training compute than GPT-4o and GPT-5, and it was trained entirely on Google TPUs. Google leads the AI race, pressuring Nvidia. DeepSeek’s V3.2 acknowledged a pre-training compute gap. They expect the biggest capex boom in 2026-2027 and an underappreciated global economic boost in the next two years.
🔗 Source: View Source | Found on Dec 03, 2025
Black Friday 2024 was the strongest day in a $314.9 billion Cyber Week, with TikTok Shop sales up 226% and Temu/Shein gaining. EU’s 2026 removal of the duty-free threshold for parcels under EUR 150 may raise costs. AI-driven traffic is expected to rise 520% year-on-year; over a third of consumers use AI; Target, Walmart, and Ralph Lauren launched assistants. Alternative data providers have grown to 2,000+, with Deloitte forecasting $137 billion by 2030. Europe’s Black Friday deliveries emitted 1.2 million tonnes of CO2, 94% above average week; about 30% of items are returned; 78% consider sustainability important.
🔗 Source: View Source | Found on Nov 27, 2025
AI adoption among U.S. businesses has doubled since 2024: about 10% used AI in the past two weeks and nearly 14% plan to within six months. Costs have plunged, with LLMs down 80%–99% and GPT‑3.5‑level inference over 280-fold cheaper between November 2022 and October 2024. AI-related activity grew over 50% year over year and contributed 30% of U.S. growth in H1 2025; capex is nearing 1% of GDP. Productivity gains could reach 0.3% near term and 0.6%–0.9% this decade, potentially 1.3% long-term. As of Q2 2025, the “AI 8” stocks accounted for 22% of S&P 500 earnings.
🔗 Source: View Source | Found on Nov 28, 2025
In August 2025, the BLS issued one of the largest NFP revisions, cutting May by 125,000 (from 144,000 to 19,000) and June by 133,000 (from 147,000 to 14,000). Two-year Treasury yields react most to the first-print surprise, with smaller responses to revisions. Since 2024, revision error has increased: the 12-month trailing MAD for the second-to-third vintage returned to GFC territory. Inflation-adjusted BLS funding has been squeezed, and an 8% cut is proposed for 2026 under “One Big Beautiful Bill.” Chair Powell signalled reliance on alternative gauges in September 2024 and July 2025. E.J. Antoni advocated suspending monthly reports.
🔗 Source: View Source | Found on Dec 02, 2025
Global equities remain supported by steady growth, easier policy and healthy earnings, but returns will vary. Expected outperformers vs MSCI World: European mid-caps and value, US growth, and emerging markets. EMs should benefit from a weaker dollar, lower real rates, higher commodity prices and AI; India: domestic investors own 18.5% of its market. Europe trades at a 25% discount to the US (vs 10% before Covid/Ukraine). US Core AI trades near 30x earnings (market 22x; >100x in dot-com); leaders’ earnings grow ~25 ppts faster, with ~20% growth expected. Favor quality, pharmaceuticals, technology, financials, industrials; the UK offers inexpensive stagflation protection.
🔗 Source: View Source | Found on Dec 03, 2025
Published on November 27, 2025, 15:20 by Amundi, the article says policy choices will drive markets. With US debt unprecedentedly high and inflation still above target, risks to the Fed’s independence persist, keeping US yields range-bound and favoring tactical duration and inflation protection. For 2026, European bonds remain a key call, with emphasis on peripheral bonds and UK Gilts. In credit, Amundi prefers euro investment grade with solid fundamentals and is cautious on US high yield due to exposure to regional banks and consumer dependence. Amundi expects the USD to continue weakening, though not linearly.
🔗 Source: View Source | Found on Dec 01, 2025
Mixture-of-experts (MoE) has become the architecture of choice for frontier models: the Artificial Analysis leaderboard’s top 10 open-source models—including DeepSeek-R1, Kimi K2 Thinking, gpt-oss-120B and Mistral Large 3—use MoE, with over 60% of open-source releases this year and a nearly 70x intelligence increase since early 2023. NVIDIA’s GB200 NVL72, a rack-scale system with 72 Blackwell GPUs, 1.4 exaflops, 30TB shared memory and 130 TB/s NVLink, scales expert parallelism and delivers a 10x performance per watt, enabling 10x token revenue. It provides 10x gains for DeepSeek-R1, Kimi K2 Thinking and Mistral Large 3, and is being deployed by major cloud providers.
🔗 Source: View Source | Found on Dec 03, 2025
Anthropic and Snowflake expanded their partnership with a multi-year $200 million agreement, bringing Claude models to the Snowflake platform and more than 12,600 customers across Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Azure. A joint go-to-market initiative will deploy AI agents to derive insights from structured and unstructured data with governance. Thousands of Snowflake customers process trillions of Claude tokens per month via Cortex AI; internal use includes Claude Code and a GTM AI Assistant. Claude achieves greater than 90% accuracy on complex text-to-SQL tasks (Snowflake benchmarks). Claude Sonnet 4.5 powers Snowflake Intelligence; Opus 4.5 was hosted day one.
🔗 Source: View Source | Found on Dec 03, 2025
On December 02, 2025 at 18:30 in Las Vegas, Amazon announced a comprehensive expansion of its Nova portfolio: four new models, a pioneering “open training” service to enable organizations to build custom model variants with Nova, and a service for creating highly reliable AI agents. Nova Forge allows companies to build optimized Nova variants by infusing proprietary data early in the training process through its unique “open training” approach.
🔗 Source: View Source | Found on Dec 02, 2025
On December 2, 2025, Mistral AI announced the open-source Mistral 3 family optimized for NVIDIA supercomputing and edge platforms. Mistral Large 3, a mixture‑of‑experts model, has 41B active parameters, 675B total parameters and a 256K context window. On NVIDIA GB200 NVL72, it uses NVLink, wide expert parallelism, NVFP4 and NVIDIA Dynamo, achieving gains over the H200. The suite adds nine compact Ministral 3 models for NVIDIA Spark, RTX PCs and Jetson, available via Llama.cpp and Ollama. Models are available today on platforms and cloud providers; NVIDIA optimized TensorRT‑LLM, SGLang and vLLM, and models are expected soon as NVIDIA NIM microservices.
🔗 Source: View Source | Found on Dec 02, 2025