Weekly Quantitative Paper Digest (Oct 25‑31 2025)

ChatGPT in Systematic Investing – Enhancing Risk‑Adjusted Returns with LLMs investigates whether large language models (LLMs) can improve cross‑sectional momentum strategies by extracting predictive signals from company‑specific news. The authors combine daily S&P 500 returns with high‑frequency news and use prompt engineering to ask ChatGPT if a stock will enter a momentum portfolio. The LLM scores recent news for continuation support, influencing stock selection and portfolio weights. The LLM‑enhanced momentum outperforms a long‑only benchmark, delivering higher Sharpe and Sortino ratios both in‑sample and in a true out‑of‑sample period after the model’s pre‑training cutoff. Gains are robust to transaction costs, prompt design, and portfolio constraints, and are strongest for concentrated, high‑confidence portfolios, showing that LLMs can serve as real‑time interpreters of financial news.

Effectiveness of Cardinality‑Return Weighted Maximum Independent Set Approach for Financial Portfolio Optimization addresses fundamental limits of the Markowitz mean‑variance model, such as the normal‑return assumption and sensitivity to estimation errors. The authors propose a graph‑theoretic CR‑WMIS model that combines a maximum independent set (MIS) to select a large number of weakly correlated stocks with a weighted MIS (WMIS) that incorporates expected returns. Using real S&P 500 data from April 2019 to March 2024, a five‑year backtest with a simulation‑fork solver demonstrates the method’s effectiveness. Comprehensive risk assessment, absent in prior MIS/WMIS studies, shows CR‑WMIS outperforms traditional MIS, WMIS, and the market index on both return and risk metrics.

Learning to Manage Investment Portfolios beyond Simple Utility Functions notes that disclosed fund objectives hide complex, multi‑objective manager goals that traditional utility‑based models struggle to specify. The paper introduces a generative framework that learns a latent representation of manager strategies without explicit utility functions. It models the conditional probability of portfolio weights given stock features, historical returns, previous weights, and latent variables, using a GAN‑based architecture that learns directly from observed holdings and market data. Validation on a dataset of 1,436 U.S. mutual funds shows the learned representations capture known styles (e.g., growth, value) and reveal heterogeneity in turnover, concentration, and latent factors. The authors also develop interpretability tests that embed expert‑label benchmarks in a linear, explainable form.

H3M‑SSMoEs: Hypergraph‑based Multimodal Learning with LLM Reasoning and Style‑Structured Mixture of Experts tackles stock‑trend prediction’s challenges of temporal dependence, heterogeneous modalities, and dynamic inter‑stock relations. The proposed H3M‑SSMoEs architecture introduces three innovations: (1) a multi‑context hypergraph with local (LCH) and global (GCH) layers that hierarchically capture fine‑grained spatio‑temporal dynamics, using shared cross‑modal hyperedges and Jensen‑Shannon‑divergence weighting for adaptive relation learning; (2) an LLM‑enhanced reasoning module that freezes a large language model and adds lightweight adapters to fuse quantitative and textual modalities, enriched with domain‑specific financial knowledge; (3) a style‑structured mixture of experts (SSMoEs) that combines shared market experts with industry‑specialized experts parameterized by learnable style vectors, enabling sparse activation and institutional awareness. Extensive experiments on three major stock markets show H3M‑SSMoEs surpass state‑of‑the‑art baselines in prediction accuracy and investment performance while maintaining effective risk control.

GroupSHAP‑Guided Integration of Financial News Keywords and Technical Indicators for Stock Price Prediction observes that recent FinBERT advances enable sentiment quantification but collapsing diverse language signals into a single index loses contextual nuance and hampers interpretability. While SHAP explains feature importance, its computational cost grows exponentially with feature count. The authors propose a GRU‑based prediction framework enhanced by GroupSHAP, which quantifies contributions of semantically coherent keyword groups rather than individual tokens, reducing computation while preserving explainability. Using FinBERT embeddings of news from 2015‑2024, articles are clustered into semantic groups; GroupSHAP measures each group’s impact on price movements. Group‑level SHAP variables feed the predictor. One‑day S&P 500 forecasts for 2024 show a 32.2% reduction in MAE and a 40.5% reduction in RMSE compared with a baseline lacking GroupSHAP, marking the first application of GroupSHAP to news‑driven financial forecasting and demonstrating enhanced interpretability and performance.

P1GPT: a Multi‑Agent LLM Workflow Module for Multi‑Modal Financial Information Analysis notes that recent LLM progress enables collaborative multi‑agent reasoning, yet most financial frameworks remain single‑agent predictors or loosely coupled analyst ensembles, lacking a unified reasoning pipeline across data modalities. P1GPT introduces a hierarchical multi‑agent LLM system that structures a reasoning pipeline through coordinated agent communication and temporal synthesis, systematically integrating technical, fundamental, and news‑driven insights. Backtesting on major U.S. stocks shows P1GPT achieves higher cumulative and risk‑adjusted returns, low drawdown, and provides transparent causal explanations for its decisions. The results suggest that a structured reasoning workflow, rather than mere role‑play imitation, offers a scalable route to explainable and trustworthy financial AI systems.