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
TradeFM is a 524 million-parameter generative Transformer model introduced by Maxime Kawawa-Beaudan, Srijan Sood, Kassiani Papasotiriou, Daniel Borrajo, and Manuela Veloso to address market microstructure by learning from billions of trade events across more than 9,000 equities. The model employs scale-invariant features and a universal tokenization scheme to unify heterogeneous order flow data without asset-specific calibration. Integrated with a deterministic market simulator, TradeFM-generated rollouts replicate key stylized facts of financial returns and achieve 2-3 times lower distributional error than Compound Hawkes baselines while generalizing zero-shot to APAC markets with moderate perplexity degradation.
🔗 Source: Summary based on arxiv.org View Source | Found on Mar 02, 2026
The article presents a large-scale benchmark of modern deep learning architectures for financial time series prediction and position sizing, focusing on Sharpe ratio optimization. The study evaluates linear models, recurrent networks, transformer-based architectures, state space models, and recent sequence representation approaches using a daily futures dataset covering commodities, equity indices, bonds, and FX from 2010 to 2025. It assesses out-of-sample performance with measures including statistical significance, downside and tail risk metrics, breakeven transaction cost analysis, robustness to random seed selection, and computational efficiency. Hybrid models like VSN with LSTM achieve the highest overall Sharpe ratio.
🔗 Source: Summary based on arxiv.org View Source | Found on Mar 03, 2026
The article introduces FinTexTS, a large-scale text-paired stock price dataset constructed using a semantic-based and multi-level pairing framework. The authors, Jaehoon Lee and colleagues, extract company-specific context from SEC filings and use embedding-based matching to retrieve semantically relevant news articles. News articles are classified into four levels—macro-level, sector-level, related company-level, and target-company level—using large language models for multi-level pairing with the target company. Experiments on FinTexTS demonstrate improved stock price forecasting performance compared to keyword-matching approaches, with further enhancements observed when applying the method to proprietary curated news sources.
🔗 Source: Summary based on arxiv.org View Source | Found on Mar 04, 2026
Timer-S1 is a Mixture-of-Experts time series foundation model with 8.3 billion total parameters and 0.75 billion activated parameters per token, featuring a context length of 11,500. The model employs Serial Scaling across architecture, dataset, and training pipeline to address scalability challenges in pre-trained time series models. Timer-S1 integrates sparse TimeMoE blocks and generic TimeSTP blocks for Serial-Token Prediction, improving long-term forecasting while reducing inference costs and error accumulation. The curated TimeBench corpus contains one trillion time points with data augmentation to reduce bias. Timer-S1 achieves state-of-the-art MASE and CRPS scores on the GIFT-Eval leaderboard.
🔗 Source: Summary based on arxiv.org View Source | Found on Mar 06, 2026
The article by Kwok Ping Tsang and Zichao Yang analyzes transaction-level data from Polymarket’s 2024 U.S. presidential-election contracts to examine how prediction markets respond to major political shocks, specifically the first Biden-Trump debate, the Trump assassination attempt, and Biden’s withdrawal. The study finds significant bursts of trading activity during these events, especially among high-intensity incumbents, with pre-event net exposure predicting abnormal post-event trading and position changes. Price discovery dynamics vary: the debate shows transitory price pressure and partial reversal; the assassination attempt leads to more permanent repricing; Biden’s dropout results in heavy trading but muted net price changes and high two-sidedness.
🔗 Source: Summary based on arxiv.org View Source | Found on Mar 04, 2026
The article introduces RAG4CTS, a regime-aware, training-free Retrieval-Augmented Generation framework for Covariate Time-Series, addressing challenges in predictive maintenance of Pressure Regulating and Shut-Off Valves (PRSOV) such as data scarcity, short transient sequences, and covariate coupled dynamics. RAG4CTS constructs a hierarchical time-series native knowledge base for lossless storage and physics-informed retrieval of raw historical regimes. It employs a two-stage bi-weighted retrieval mechanism based on point-wise and multivariate similarities and an agent-driven strategy for dynamic context optimization. Deployed in Apache IoTDB at China Southern Airlines, it identified one PRSOV fault in two months with zero false alarms.
🔗 Source: Summary based on arxiv.org View Source | Found on Mar 06, 2026
The study by Donghao Huang and Zhaoxia Wang, accepted at PAKDD 2026, evaluated 504 configurations across seven large language model families on sentiment analysis tasks of varying complexity. Results showed reasoning effectiveness is task-dependent: binary classification performance dropped by up to -19.9 F1 percentage points, while 27-class emotion recognition improved by up to +16.0pp. Distilled reasoning variants underperformed base models by 3-18pp on simpler tasks, though few-shot prompting partially recovered performance. Few-shot learning outperformed zero-shot in most cases, with gains varying by architecture and task complexity. Reasoning incurred a computational overhead of 2.1x-54x and was justified only for complex emotion recognition tasks.
🔗 Source: Summary based on arxiv.org View Source | Found on Mar 02, 2026
The article introduces macro-conditioned scenario-context rollout (SCR), a method designed to generate plausible next-day multivariate return scenarios under market stress events, addressing the challenge of distribution shifts in portfolio rebalancing. The authors analyze the reward–transition mismatch that arises when incorporating scenario-based rewards into temporal-difference learning, which destabilizes reinforcement learning critic training. They propose constructing a counterfactual next state using rollout-implied continuations to augment the critic agent's bootstrap target, stabilizing learning and achieving a bias-variance tradeoff. In out-of-sample tests across 31 U.S. equity and ETF portfolio universes, SCR improves Sharpe ratio by up to 76% and reduces maximum drawdown by up to 53% compared with classic and RL-based baselines.
🔗 Source: Summary based on arxiv.org View Source | Found on Mar 02, 2026
Using a new measure called observed exposure, which combines theoretical LLM capability and real-world usage data, the article finds that AI’s actual coverage of tasks remains well below its theoretical potential; for example, Claude covers only 33% of Computer & Math tasks and 75% for Computer Programmers. Occupations with higher observed exposure are projected by the BLS to grow less through 2034, with every 10 percentage point increase in coverage linked to a 0.6 percentage point drop in growth projections. Highly exposed workers tend to be older, female, more educated, and higher-paid. No systematic increase in unemployment is found for these workers since late 2022, but there is suggestive evidence that hiring of younger workers (aged 22-25) into highly exposed occupations has slowed by about 14% compared to pre-ChatGPT rates.
🔗 Source: Summary based on anthropic.com View Source | Found on Mar 05, 2026
Phi-4-reasoning-vision-15B is a 15 billion parameter open-weight multimodal reasoning model released on March 4, 2026, by Microsoft and available via Microsoft Foundry, HuggingFace, and GitHub. The model uses a mid-fusion architecture with the SigLIP-2 Naflex vision encoder and Phi-4-Reasoning backbone, trained on 200 billion tokens of multimodal data—significantly less than models like Qwen VL or Gemma3. It excels at math and science reasoning, computer-use tasks, and general vision-language applications while maintaining low inference costs. The dataset emphasizes high-quality filtered open-source data supplemented by internal and synthetic sources. Evaluation used Eureka ML Insights and VLMEvalKit frameworks.
🔗 Source: Summary based on microsoft.com View Source | Found on Mar 04, 2026
Coherent, founded in 1971 and operating in more than 20 countries, is a global leader in photonics technology serving datacenter, communications, and industrial markets. The press release discusses forward-looking statements regarding the expansion of the supply relationship between Coherent and NVIDIA. NVIDIA is described as a world leader in AI and accelerated computing. Both companies caution that these statements are subject to risks and uncertainties, including potential changes to agreements, litigation, business relationship impacts, technological developments, market acceptance, and regulatory changes as detailed in their respective SEC filings.
🔗 Source: Summary based on nvidianews.nvidia.com View Source | Found on Mar 02, 2026
Correlations between the Mag 7 names have decreased from approximately 0.55–0.60 to around 0.20–0.30, indicating a shift from the “beta” phase to a “broadening” phase in the market. In this new phase, price appreciation will increasingly reflect measurable improvements in margins and return on invested capital by individual operators, with this process occurring at different times across industries. The article notes that passive management dominates during the beta phase, while active management is expected to outperform as investors rotate into small cap, mid cap, cyclical value, and international exposures during broadening.
🔗 Source: Summary based on im.natixis.com View Source | Found on Mar 03, 2026
OpenAI published a feature detailing how Balyasny utilizes AI in investment research, focusing on the Applied AI team composed of researchers, engineers, and domain experts who develop AI-native tools integrated into team workflows. These tools incorporate rigorous evaluation, traceable reasoning, and real-time feedback loops from inception. The article outlines four principles guiding this approach: model evaluation, feedback-driven design, firm-wide deployment, and the observed outcomes across investment teams. According to the feature, AI enables teams to apply first principles thinking more rapidly, process more data efficiently, and operate with greater structure.
🔗 Source: Summary based on bamfunds.com View Source | Found on Mar 06, 2026
Generative AI is reshaping the software industry by automating tasks such as note taking and meeting summarisation, creating new product classes, and driving increased compute demand. The Organisation for Economic Co-operation and Development found generative AI to be pervasive and innovative, enabling new applications. Despite a broad market sell-off in software stocks driven by fears of disruption from companies like Anthropic or OpenAI, the article argues that not all traditional software firms will be displaced; those with deep enterprise integration in customer management or financial systems are likely to become more necessary. Current valuations reflect severe disruption scenarios, presenting selective investment opportunities.
🔗 Source: Summary based on lombardodier.com View Source | Found on Mar 04, 2026
Lombard Odier’s Rethink Perspectives event in London, opened by Managing Partner Frédéric Rochat, addressed global economic uncertainty heightened by a new Middle East conflict. Chief Economist Samy Chaar stated Lombard Odier’s base case expects the conflict to last weeks, with oil price rises of about USD 10 per barrel not derailing expansion, as roughly 20% of global oil flows through the Strait of Hormuz. Labour markets show no recession signals across major regions. James Kynge highlighted China’s technological rise, leading in 66 of 74 critical technologies and producing six million STEM graduates annually, potentially accounting for nearly half of global industrial production by 2030.
🔗 Source: Summary based on lombardodier.com View Source | Found on Mar 07, 2026