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.
“Causal Inference in Financial Event Studies” by Paul Goldsmith-Pinkham and Tianshu Lyu shows that event studies using linear factor models produce inconsistent treatment effect estimates when factor models are misspecified. Bias is particularly severe during volatile periods, over long horizons, and when event timing correlates with market conditions. The authors derive identification conditions and asymptotic bias expressions, and propose synthetic control methods that construct replicating portfolios from control securities without imposing specific factor structures. Revisiting four empirical applications, they find some results reflect misspecification; traditional methods remain reliable for short-horizon studies with random event timing.
🔗 Source: View Source | Found on Nov 21, 2025
Eghbal Rahimikia, Hao Ni, and Weiguan Wang (arXiv:2511.18578) present the first comprehensive empirical study of TSFMs in global financial markets. Using a large-scale dataset of daily excess returns across diverse markets, it evaluates zero-shot inference, fine-tuning, and pre-training from scratch. Off-the-shelf pre-trained TSFMs perform poorly in zero-shot and fine-tuning, while models pre-trained from scratch on financial data deliver substantial forecasting and economic improvements; larger datasets, synthetic data augmentation, and hyperparameter tuning further enhance performance.
🔗 Source: View Source | Found on Nov 26, 2025
Hang Ding, Xue Wang, Tian Zhou, and Tao Yao introduce a single-stage, end-to-end diffusion framework for point forecasts. A unified Transformer serves as both denoiser and predictor, removing any need for external pre-trained or jointly trained regressors. SimDiff attains state-of-the-art point estimation by leveraging intrinsic output diversity with multiple inference ensembling to improve mean squared error. Innovations—normalization independence and a median-of-means estimator—enhance adaptability and stability, with extensive experiments showing significant outperformance.
🔗 Source: View Source | Found on Nov 26, 2025
Amir Kavoosi’s paper (arXiv:2511.16319) introduces the Quantitative Geometric Market Structuralist (QGMS) framework, a hybrid method integrating geometric pattern recognition with quantitative modeling to identify terminal zones of market movements. QGMS treats markets as geometric structures and uses blind-testing—concealing price, symbol, and temporal identifiers—to ensure verification without revealing the algorithmic core. Empirical tests across the 2008 Global Financial Collapse, the 2015 EUR CHF SNB event, the 2016 Brexit referendum, and the 2020 COVID-19 crash show consistent identification of structural endpoints before major reversals, establishing a foundation for evaluation and research into nonlinear structural forecasting models.
🔗 Source: View Source | Found on Nov 22, 2025
Miquel Noguer i Alonso introduces Financial Information Theory, applying entropy, Kullback-Leibler divergence, mutual information, and normalized mutual information (NMI) to financial time series. The paper derives these measures with proofs and proposes estimation algorithms. Using S&P 500 daily returns from 2000–2025, it validates regime detection, market efficiency testing, and portfolio construction. NMI serves as a diagnostic of the Efficient Market Hypothesis and highlights structural changes in 2008 and during COVID-19. Tools include entropy-adjusted Value at Risk, an information-theoretic diversification criterion, and an NMI-based market efficiency test, with superior regime detection versus autocorrelation or volatility approaches.
🔗 Source: View Source | Found on Nov 22, 2025
Silvia Onofri, Andrey Shternshis, and Stefano Marmi present “Emergence of Randomness in Temporally Aggregated Financial Tick Sequences,” introducing a model-free methodology to assess how tick-by-tick returns resemble random sequences. Beyond serial correlation or entropy, they apply many tests from the NIST Statistical Test Suite and TestU01 (Rabbit and Alphabit) to ultra-high-frequency trade data. They show that increasing transaction-time aggregation transforms highly correlated ticks into more random streams and reveal asset-specific, non-monotonic predictability patterns, enabling assessment and pseudo-random sequence generation from financial time series.
🔗 Source: View Source | Found on Nov 24, 2025
Fermat Leukam, Rock Stephane Koffi, and Prudence Djagba propose enhancing portfolio optimization using G-Learning with parametric tuning via the GIRL algorithm. The goal is to maximize portfolio value by a target date while minimizing periodic investor contributions. In a highly volatile market with a well-diversified, low-risk portfolio, reinforcement learning dynamically adjusts positions. Results show the Sharpe Ratio improved from 0.42 to 0.483. GIRL optimizes reward parameters (lambda = 0.0012 vs 0.002) with marginal performance impact, indicating G-Learning’s robustness. The work appears as arXiv:2511.18076; the arXiv-issued DOI is pending registration.
🔗 Source: View Source | Found on Nov 26, 2025
The letter argues AI-related equities exhibit classic bubble traits: extreme S&P 500 valuations, speculative VC funding, euphoric pricing of AI and quantum names, and mega-cap re-ratings untethered from realistic profit paths. Yet it targets the “agnostic investor”: someone unsure whether AI is mispriced but worried about downside. Versus 2007–08 or 2021, this resembles the 2000 internet bubble, where you could sidestep the mania without abandoning risk. The author recommends underweighting U.S. AI winners in favour of non-U.S. equities, deep value, small value and liquid alternatives, which offer attractive expected returns in both bubble and non-bubble scenarios.
🔗 Source: View Source | Found on Nov 25, 2025
Meta fell 11% on 30 October; the Nasdaq‑100 trades at 31x forward earnings and was up 20% YTD to 11 November versus 15.6% for the S&P 500. Bloomberg logged 140 ‘AI Bubble’ stories on day AMD announced OpenAI partnership and over 600 by 7 November. Vanguard projects AI boosting productivity 20% by 2035, lifting US GDP growth to 3% in the 2030s. Jamie Dimon expects an adjustment “in the next six months to two years.” The article advocates multi‑asset diversification via gold, money markets, short‑dated investment‑grade bonds, European financial debt, climate bonds, and long/short strategies.
🔗 Source: View Source | Found on Nov 25, 2025
Adam Berger and Brian Barbetta assess AI’s portfolio impacts. Barbetta says markets underestimate AI, with rapidly improving, compute-intensive models and possible insufficiency of planned data-center capacity, implying larger spending. He notes significantly better-than-market-average returns on invested capital and efficient redeployment. He also cites recent social media advances. Berger urges clear technology weighting, risk management of concentrated bets, checklists to course-correct, and attention to infrastructure tied to data centers. A poll of over 60 institutional investors found almost 40% underweight technology. Teams use AI to scan research and filings, cutting analyses from days to hours.
🔗 Source: View Source | Found on Nov 20, 2025
Claude Opus 4.5 is available on apps, the Claude API, and all three major cloud platforms, priced at $5/$25 per million tokens. In a 2‑hour performance engineering take‑home exam, it scored higher than any human candidate ever. A new API effort parameter lets developers trade speed for capability: at medium effort it matches Sonnet 4.5 on SWE‑bench Verified using 76% fewer output tokens; at highest effort it exceeds Sonnet 4.5 by 4.3 percentage points using 48% fewer tokens. Updates include Claude Code Plan Mode, desktop, Chrome (Max), Excel beta (Max, Team, Enterprise), chat summarization, and higher usage limits.
🔗 Source: View Source | Found on Nov 24, 2025
Anthropic’s Claude Opus 4.5, now in public preview in Microsoft Foundry, GitHub Copilot paid plans, and Microsoft Copilot Studio, with availability coming to Visual Studio Code via the Foundry extension. Anthropic reports 80.9% on SWE-bench; the model outperforms Sonnet 4.5 and Opus 4.1 and is priced at one third of previous Opus-class models. It adds Programmatic Tool Calling in Python, Tool Search, and Tool Use Examples, plus Foundry features Effort Parameter (Beta) and Compaction Control. Anthropic cites reduced misaligned responses, stronger prompt-injection robustness, and improved tool use across finance and cybersecurity.
🔗 Source: View Source | Found on Nov 26, 2025
Highlights enterprises building specialized AI agents on NVIDIA Nemotron open models. CrowdStrike’s Charlotte AI AgentWorks, powered by Nemotron and NVIDIA NIM, automates alert triage and remediation, raising triage accuracy from 80% to 98.5% and reducing manual effort tenfold. PayPal built a fine-tuning pipeline in two weeks, cut latency nearly 50%, and relies on open, modular models fine-tuned for payments and commerce to serve 430 million customers and 30 million merchants. Synopsys’ AI agents on NVIDIA accelerated infrastructure delivered a 72% productivity boost in early formal verification trials. Steps include NeMo-based lifecycle management and continuous fine-tuning.
🔗 Source: View Source | Found on Nov 24, 2025
Ironwood is custom silicon now available to Cloud customers. Acting as a hugely efficient parallel processor, it manages massive calculations and minimizes on‑chip data‑shuttle time, speeding complex AI so models run faster and smoother across their cloud. Purpose‑built for the age of inference, it is custom built for high‑volume, low‑latency AI inference and model serving, while also improving training. Ironwood delivers more than 4X better performance per chip for both training and inference versus the last generation, making it their most powerful and energy‑efficient custom silicon to date.
🔗 Source: View Source | Found on Nov 25, 2025
LLM Open Finance releases two 8B-parameter models trained on a balanced dataset: 54% financial, 20% translation, 16% general domain, 8% RAG, and 2% math/reasoning/coding. Finetuning focused on English, French, and German while preserving multilingual capabilities. The models show consistent gains over base models on financial tasks, including improved understanding of financial French acronyms and domain terminology, and evaluations report outperforming general-domain models on financial tasks and translation while maintaining strong general knowledge. They are RAG-ready. Commercial LLM Pro Finance models: Gemma Pro Finance 12B, Qwen Pro Finance R 32B; playground and paper (arXiv:2511.08621).
🔗 Source: View Source | Found on Nov 24, 2025