Paper Summaries: Recent AI-Driven Finance Research (Sep 20‑26, 2025)

Increase Alpha: Performance and Risk of an AI‑Driven Trading Framework

Paper link: http://arxiv.org/pdf/2509.16707v1

Authors: Sid Ghatak, Arman Khaledian, Navid Parvini, Nariman Khaledian

The authors observe inefficiencies in U.S. equity markets and design a lightweight deep‑learning framework that uses feed‑forward and recurrent networks together with curated features to capture subtle patterns in noisy price, volume, and cross‑sectional data. The system maps daily directional signals for more than 800 U.S. stocks while keeping computational overhead minimal, making it suitable for large‑scale production. Performance is evaluated with industry‑standard metrics: cumulative returns, annualized Sharpe ratio (> 2.5), maximum drawdown (~ 3 %), and correlation with the S&P 500 (≈ 0). The framework is benchmarked against simple baselines and macro‑economic indicators. Robustness is demonstrated across market regimes, including the early‑2025 U.S. market turbulence, where the model maintains stable risk‑adjusted returns.

Financial Risk Relation Identification through Dual‑view Adaptation

Paper link: http://arxiv.org/pdf/2509.18775v1

Authors: Wei‑Ning Chiu, Yu‑Hsiang Wang, Andy Hsiao, Yu‑Shiang Huang, Chuan‑Ju Wang

To extract inter‑company risk relationships, the authors treat SEC Form 10‑K filings as a standardized textual source. They apply recent natural‑language‑processing advances to perform unsupervised fine‑tuning that leverages temporal and lexical patterns within the documents, thereby capturing implicit and abstract risk links. This yields a domain‑specific financial encoder and a quantitative risk‑relation scoring system that enhances transparency and interpretability. Extensive experiments on multiple evaluation settings show the method outperforms strong baselines.

Multimodal Language Models with Modality‑Specific Experts for Financial Forecasting from Interleaved Sequences of Text and Time Series

Paper link: http://arxiv.org/pdf/2509.19628v1

Authors: Ross Koval, Nicholas Andrews, Xifeng Yan

The paper proposes a unified neural architecture that processes interleaved sequences of news text and price time series using modality‑specific expert modules. Each expert learns patterns unique to its modality while a token‑weighting cross‑modal alignment module learns to focus on the most informative tokens across modalities. Experiments on large‑scale financial forecasting tasks demonstrate state‑of‑the‑art performance compared with strong single‑modal and multimodal baselines. An interpretability analysis shows that the time‑series context contributes significant predictive value and validates the alignment objective.

Predicting Credit Spreads and Ratings with Machine Learning: The Role of Non‑Financial Data

Paper link: http://arxiv.org/pdf/2509.19042v1

Authors: Yanran Wu, Xinlei Zhang, Quanyi Xu, Qianxin Yang, Chao Zhang

The authors construct a credit‑risk feature set of 167 indicators, combining macro‑economic, firm‑financial, bond‑specific variables with 30 large‑scale non‑financial attributes (e.g., corporate governance, ownership structure, disclosure quality). Seven machine‑learning models are trained to predict bond credit spreads and to assess rating performance. Results show the models explain spreads better than Chinese credit‑rating agencies. Adding non‑financial features more than doubles out‑of‑sample performance relative to traditional feature‑driven models. Feature‑importance analysis reveals that 7 of the top‑10 predictors are non‑financial, consistently identifying high‑risk characteristics such as deteriorating operations, short‑term debt pressure, and tighter financing constraints. Using predicted spreads as implicit rating signals yields accuracy, recall, and F1 scores exceeding 75 % across industry‑ and sub‑industry‑specific models.