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.

SELECTIONS FROM ARXIV

🔹 Manifold-Constrained Hyper-Connections Introduced as New Framework for Data Structure Optimization

The article, "mHC: Manifold-Constrained Hyper-Connections," authored by Zhenda Xie and 18 others and submitted on 31 December 2025, introduces the mHC framework to address challenges in Hyper-Connections (HC), such as compromised identity mapping, training instability, restricted scalability, and memory access overhead. mHC projects the residual connection space of HC onto a specific manifold to restore the identity mapping property while optimizing infrastructure for efficiency. Empirical experiments show that mHC improves training at scale with tangible performance gains and enhanced scalability, offering a flexible extension of HC for foundational model evolution.

🔗 Source: arxiv.org View Source | Found on Jan 01, 2026

🔹 Synthetic Financial Data Used to Improve Financial Modelling Accuracy

This paper, submitted on 25 December 2025 by Christophe D. Hounwanou, Yae Ulrich Gaba, and Pierre Ntakirutimana, introduces a unified multi-criteria evaluation framework for synthetic financial data generation and applies it to ARIMA-GARCH, Variational Autoencoders (VAEs), and Time-series Generative Adversarial Networks (TimeGAN) using historical S&P 500 daily data. The study evaluates fidelity via Maximum Mean Discrepancy (MMD), temporal structure through autocorrelation and volatility clustering, and practical utility in portfolio optimization and volatility forecasting. Results show TimeGAN achieved the lowest MMD (1.84e-3 average over 5 seeds) with superior realism and temporal coherence compared to other models.

🔗 Source: arxiv.org View Source | Found on Dec 29, 2025

🔹 Generative AI Used to Build Sector-Based Investment Portfolios

This paper by Alina Voronina and co-authors examines the application of Large Language Models (LLMs) from OpenAI, Google, Anthropic, DeepSeek, and xAI to quantitative sector-based portfolio construction using S&P 500 sector indices. Each LLM selected and weighted 20 stocks per sector, with portfolios evaluated over two out-of-sample periods: January-March 2025 (stable market) and April-June 2025 (volatile market). LLM-weighted portfolios often outperformed sector indices in stable conditions but underperformed during volatility. Combining LLM stock selection with classical optimization improved both performance and consistency, highlighting the potential of hybrid AI-quantitative investment strategies.

🔗 Source: arxiv.org View Source | Found on Jan 01, 2026

🔹 LLM Agents Enhance Combinatorial Efficient Frontier Analysis for Investment Portfolio Optimization

The article, "LLM Agents for Combinatorial Efficient Frontiers: Investment Portfolio Optimization" by Simon Paquette-Greenbaum and Jiangbo Yu, discusses the Cardinality Constrained Mean-Variance Portfolio Optimization (CCPO) problem, a mixed-integer quadratic programming challenge common in financial institutions. The authors present a novel agentic framework for CCPO that automates complex workflows and algorithm development. In benchmark tests, this framework matches state-of-the-art algorithms and alleviates workflow complexity; in the worst case, it reports lower but acceptable error. The study highlights the efficiency of agentic frameworks in combinatorial optimization tasks related to investment portfolios.

🔗 Source: arxiv.org View Source | Found on Jan 05, 2026

🔹 Deep Learning Applied to Art Market Valuation

The article "Deep Learning for Art Market Valuation" by Jianping Mei, Michael Moses, Jan Waelty, and Yucheng Yang examines how deep learning can enhance art market valuation by integrating visual content of artworks into predictive models. Using a large repeated-sales dataset from major auction houses, the study benchmarks classical hedonic regressions and tree-based methods against modern deep architectures, including multi-modal models that combine tabular and image data. The research finds that while artist identity and prior transaction history are primary predictors, visual embeddings significantly improve valuation accuracy for fresh-to-market works lacking historical data.

🔗 Source: arxiv.org View Source | Found on Dec 31, 2025

🔹 Multi-Retriever RAG Method Integrates Domain Knowledge with LLMs for Financial Question Answering

The research by Yukun Zhang, Stefan Elbl Droguett, and Samyak Jain presents a multi-retriever Retrieval Augmented Generator (RAG) system using the latest Large Language Models (LLMs) to improve financial numerical reasoning Question Answering tasks. Domain-specific training with the SecBERT encoder enabled their best neural symbolic model to surpass the FinQA paper's top baseline. Their prompt-based LLM generator achieved state-of-the-art performance with over 7% improvement but remained below human expert levels. The study found that external knowledge gains generally outweigh hallucination loss in larger models and confirmed enhanced numerical reasoning capabilities of LLMs optimized for few-shot learning.

🔗 Source: arxiv.org View Source | Found on Jan 02, 2026

🔹 Forward-Oriented Causal Observables Proposed for Analyzing Non-Stationary Financial Markets

The article by Lucas A. Souza, submitted on 31 December 2025, investigates short-horizon forecasting in financial time series under strict causal constraints, treating markets as non-stationary stochastic systems. Instead of proposing a direct price-forecast model, the study constructs an interpretable causal signal from micro-features such as momentum, volume pressure, trend acceleration, and volatility-normalized price location. The methodology involves causal centering, linear aggregation into a composite observable, causal stabilization using a one-dimensional Kalman filter, and an adaptive forward-like operator. Application to high-frequency EURUSDT (1-minute) data demonstrates economic relevance in specific regimes but shows degradation under regime shifts.

🔗 Source: arxiv.org View Source | Found on Jan 01, 2026

🔹 Vanilla LSTMs Surpass Transformer Models in Stock Price Forecasting, Study Finds

The article "StockBot 2.0: Vanilla LSTMs Outperform Transformer-based Forecasting for Stock Prices" by Shaswat Mohanty, submitted on 1 Jan 2026, presents an enhanced StockBot architecture that systematically evaluates attention-based, convolutional, and recurrent time-series forecasting models within a unified experimental setting. The study finds that a carefully constructed vanilla LSTM consistently achieves superior predictive accuracy and more stable buy/sell decision-making compared to transformer-inspired models when trained under default hyperparameters. The results emphasize the robustness and data efficiency of recurrent sequence models for financial time-series forecasting, especially with limited data discretized to single-day intervals.

🔗 Source: arxiv.org View Source | Found on Jan 05, 2026

🔹 MASFIN Introduces Multi-Agent System for Decomposed Financial Reasoning and Forecasting

MASFIN is a modular multi-agent system developed by Marc S. Montalvo and Hamed Yaghoobian that integrates large language models, specifically GPT-4.1-nano, with structured financial metrics and unstructured news for decomposed financial reasoning and forecasting. The system includes explicit bias-mitigation protocols and generates weekly portfolios of 15-30 equities with allocation weights optimized for short-term performance. In an eight-week evaluation, MASFIN achieved a 7.33% cumulative return, outperforming the S&P 500, NASDAQ-100, and Dow Jones benchmarks in six out of eight weeks but exhibited higher volatility. The work was accepted to the NeurIPS 2025 Workshop on Generative AI in Finance.

🔗 Source: arxiv.org View Source | Found on Dec 29, 2025