Weekly Quantitative Finance Paper Digest (Sep 13‑19, 2025)

Adaptive and Regime‑Aware RL for Portfolio Optimization (arXiv:2509.14385v1) by Gabriel Nixon Raj proposes an adaptive reinforcement‑learning framework for long‑term portfolio optimization. The environment lets the agent dynamically reallocate capital in response to macro‑economic regime shifts, using mixed observations and a constrained reward that penalizes volatility, includes capital‑reset terms, and incorporates tail‑risk shocks. Benchmarks compare PPO, LSTM‑PPO and Transformer‑PPO against equal‑weight and Sharpe‑ratio baselines; Transformer PPO achieves the highest risk‑adjusted return, while the LSTM variant offers a favorable trade‑off between interpretability and training cost.

DeltaHedge: A Multi‑Agent Framework for Portfolio Options Optimization (arXiv:2509.12753v1) by Feliks Bańka and Jarosław A. Chudziak introduces DeltaHedge, a multi‑agent system that integrates options trading with AI‑driven portfolio management. By combining advanced reinforcement‑learning techniques with option‑based hedging strategies, the framework improves risk‑adjusted returns and stabilizes portfolio performance across market conditions. Experiments show DeltaHedge outperforms traditional equity‑only strategies and independent models, highlighting its potential to transform quantitative finance.

Dynamic Factor Models with Forward‑Looking Views (arXiv:2509.11528v1) by Anas Abdelhakmi and Andrew E. B. Lim develops a method that augments historically calibrated dynamic factor models with forward‑looking expert views of factor/ covariate values. The authors derive closed‑form dynamics for factors and asset prices conditioned on these views, revealing a novel connection between the conditional factor process and a mean‑reverting bridge (MrB) extending the classic Brownian bridge. Optimal portfolio strategies are derived, showing that views affect mean‑variance terms and inter‑temporal hedging, and the framework accommodates online learning of uncertain long‑run means.

QuantAgent: Price‑Driven Multi‑Agent LLMs for High‑Frequency Trading (arXiv:2509.09995v2) by Fei Xiong et al. presents QuantAgent, the first multi‑agent large‑language‑model system designed for high‑frequency algorithmic trading. Four specialized agents (metrics, patterns, trends, risk) are equipped with domain‑specific tools and structured reasoning to capture short‑term market dynamics. Zero‑shot evaluation on ten instruments (including Bitcoin and Nasdaq futures) demonstrates superior prediction accuracy and cumulative returns over random baselines, illustrating the benefit of combining structured financial priors with LLM reasoning for real‑time trading.

Trading‑R1: Financial Trading with LLM Reasoning via Reinforcement Learning (arXiv:2509.11420v1) by Yijia Xiao et al. proposes Trading‑R1, a finance‑aware model that aligns strategic planning, argument construction, factual analysis, and volatility‑adjusted decision‑making. Trained with supervised fine‑tuning and reinforcement learning on the 100k‑sample Tauric‑TR1‑DB (covering 18 months, 14 stocks, and 5 heterogeneous data sources), the system achieves higher risk‑adjusted returns and lower drawdowns than open‑source and proprietary instruction‑following models across six major stocks/ETFs. Trading‑R1 generates structured, evidence‑based investment arguments to support disciplined, explainable trading.

Prediction of Stocks Index Price using Quantum GANs (arXiv:2509.12286v1) by Sangram Deshpande et al. investigates quantum generative adversarial networks (QGANs) for stock price prediction. Implemented on the AWS Braket SV1 simulator, the QGAN model generates synthetic market data that closely matches real dynamics, improving prediction accuracy over LSTM and classical GAN baselines. Experiments on index price data show faster convergence and higher forecasting precision, suggesting quantum‑enhanced models can offer speed and accuracy advantages in financial time‑series prediction.

FinSentLLM: Multi‑LLM and Structured Semantic Signals for Enhanced Financial Sentiment Forecasting (arXiv:2509.12638v1) by Zijian Zhang et al. introduces FinSentLLM, a lightweight multi‑LLM framework that aggregates expert LLM sentiment predictions with structured semantic financial signals via a compact meta‑classifier. Without costly retraining, the system captures complementary expert knowledge, semantic reasoning, and agreement/disagreement patterns, yielding 3‑6% improvements in accuracy and F1 over strong baselines. Econometric analysis (DCC‑GARCH and Johansen cointegration) on the FNSPID dataset confirms statistically significant long‑run comovement between financial sentiment and stock market returns.