Weekly AI‑Finance Paper Digest (Oct 18‑24 2025)

3S‑Trader: A Multi‑LLM Framework for Adaptive Stock Scoring, Strategy, and Selection in Portfolio Optimization

Paper link: http://arxiv.org/pdf/2510.17393v1

Authors: Kefan Chen, Hussain Ahmad, Diksha Goel, Claudia Szabo

3S‑Trader is a training‑free pipeline composed of three modules:

Scoring: For each candidate stock, recent multimodal signals are summarized into a concise report covering multiple scoring dimensions, enabling direct comparison across stocks.

Strategy: Historical strategy performance and overall market conditions are analyzed to iteratively generate an optimized selection strategy.

Selection: Stocks with high scores on the dimensions emphasized by the generated strategy are assembled into a portfolio.

Evaluation on four domains (DJIA constituents and three industry‑specific sets) shows the highest cumulative return of 131.83 % on DJIA, a Sharpe ratio of 0.31 and a Calmar ratio of 11.84, with strong results on the other sectors.

Denoising Complex Covariance Matrices with Hybrid ResNet and Random Matrix Theory: Cryptocurrency Portfolio Applications

Paper link: http://arxiv.org/pdf/2510.19130v1

Authors: Andres Garcia‑Medina

The authors observe power‑law scaling in covariance matrices estimated from short‑term, noisy, non‑Gaussian crypto time series. They propose a power‑law covariance model and a two‑step hybrid estimator:

RMT component: Regularizes the eigenvalue spectrum under high‑dimensional noise.

ResNet component: Learns data‑driven corrections to recover latent dependencies.

Monte‑Carlo simulations demonstrate consistent minimization of Frobenius and minimum‑variance (MV) losses across several covariance models. Empirical tests on 89 cryptocurrencies (2020‑2025) use a training period ending at Bitcoin’s local maximum in November 2021 and a testing period covering the subsequent bear market. The combined estimator yields the most profitable and balanced portfolio and remains robust to regime changes.

FinCARE: Financial Causal Analysis with Reasoning and Evidence

Paper link: http://arxiv.org/pdf/2510.20221v1

Authors: Alejandro Michel, Abhinav Arun, Bhaskarjit Sarmah, Stefano Pasquali

FinCARE integrates statistical causal discovery with two sources of domain knowledge:

A financial knowledge graph extracted from SEC 10‑K filings.

Concept‑level reasoning from large language models (LLMs).

Knowledge‑graph constraints are encoded into three causal discovery algorithms—constraint‑based PC, score‑based GES, and continuous‑optimization NOTEARS—while LLM reasoning generates hypotheses. On a synthetic dataset (500 companies, 18 variables) the KG+LLM augmentation improves F1 scores:

PC: 0.622 vs. 0.459 baseline (+36 %).

GES: 0.735 vs. 0.367 (+100 %).

NOTEARS: 0.759 vs. 0.163 (+366 %).

Counterfactual prediction error averages 0.003610 (MAE) and intervention‑effect direction is perfectly accurate.

Aligning Multilingual News for Stock Return Prediction

Paper link: http://arxiv.org/pdf/2510.19203v1

Authors: Yuntao Wu, Lynn Tao, Ing‑Haw Cheng, Charles Martineau, Yoshio Nozawa, John Hull, Andreas Veneris

The method applies optimal transport to align sentences in multilingual Bloomberg news articles, identifying semantically similar content across English and Japanese. Applied to >140,000 article pairs covering ~3,500 Tokyo Stock Exchange stocks (2012‑2024), the aligned sentences become sparser, more interpretable, and exhibit higher semantic similarity. Return scores derived from aligned sentences correlate more strongly with actual stock returns, and trading strategies built on these scores achieve Sharpe ratios up to 10 % higher than strategies using full‑text samples.

A Three‑Step Machine Learning Approach to Predict Market Bubbles with Financial News

Paper link: http://arxiv.org/pdf/2510.16636v1

Author: Abraham Atsiwo

The framework consists of three sequential steps:

Right‑tail unit root test (real‑time bubble detection) identifies bubble periods in the S&P 500.

Natural‑language‑processing extracts sentiment features from large‑scale financial news, capturing investor expectations.

Ensemble learning combines high‑sentiment foundations with macroeconomic predictors to forecast bubble occurrences.

K‑fold cross‑validation shows the ensemble markedly improves prediction accuracy and robustness compared with baseline machine‑learning models.

Fusing Narrative Semantics for Financial Volatility Forecasting (M2VN)

Paper link: http://arxiv.org/pdf/2510.20699v1

Authors: Yaxuan Kong, Yoontae Hwang, Marcus Kaiser, Chris Vryonides, Roel Oomen, Stefan Zohren

M2VN (Multimodal Volatility Network) unifies structured market time‑series features with unstructured news embeddings generated by Time Machine GPT, a point‑in‑time LLM that preserves temporal integrity. An auxiliary alignment loss encourages coherent fusion of the two modalities and mitigates look‑ahead bias. Extensive experiments demonstrate that M2VN consistently outperforms existing baselines in volatility forecasting, highlighting its utility for risk management.

Sentiment and Volatility in Financial Markets: A Review of BERT and GARCH Applications during Geopolitical Crises

Paper link: http://arxiv.org/pdf/2510.16503v1

Authors: Domenica Mino, Cillian Williamson

Using a fine‑tuned finance‑adapted BERT model, sentiment scores are extracted from news articles published between 2024‑01‑01 and 2024‑07‑17 on major US platforms. An enhanced Student‑t GARCH model captures heavy‑tailed return distributions. Empirical analysis reveals a statistically significant negative correlation between negative news sentiment (particularly war‑related coverage) and market stability, indicating that pessimistic sentiment is associated with increased S&P 500 volatility.

News‑Aware Direct Reinforcement Trading for Financial Markets

Paper link: http://arxiv.org/pdf/2510.19173v1

Authors: Qing‑Yu Lan, Zhan‑He Wang, Jun‑Qian Jiang, Yu‑Tong Wang, Yun‑Song Piao

The approach feeds LLM‑derived news sentiment scores together with raw price and volume data directly into reinforcement‑learning agents. Sequence models (RNNs or Transformers) process these observations to produce end‑to‑end trading actions. Experiments on cryptocurrency markets evaluate Double Deep Q‑Network (DDQN) and Group‑Relative Policy Optimization (GRPO); both algorithms outperform market benchmarks without relying on handcrafted features or manually designed rules.