Quantitative Finance Paper Digest: Dec 13‑19 2025 Highlights

LLM‑based Personalized Portfolio Recommender: Integrating Large Language Models and Reinforcement Learning for Intelligent Investment Strategy Optimization

Paper link: https://arxiv.org/pdf/2512.12922v1

Authors: Bangyu Li, Boping Gu, Ziyang Ding

Abstract: In modern financial markets, investors increasingly seek personalized and adaptive portfolio strategies that reflect individual risk preferences and respond to dynamic market conditions. Traditional rule‑based or static optimization methods often fail to capture the nonlinear interactions among investor behavior, market volatility, and evolving financial goals. To address these limitations, this paper introduces an LLM‑based personalized portfolio recommender, an integrated framework that combines large language models, reinforcement learning, and personalized risk‑preference modeling to support intelligent investment decisions.

Deep Hedging with Reinforcement Learning: A Practical Framework for Option Risk Management

Paper link: https://arxiv.org/pdf/2512.12420v1

Authors: Travon Lucius, Christian Koch, Jacob Starling, Julia Zhu, Miguel Urena, Carrie Hu

Abstract: The authors propose a reinforcement‑learning framework for dynamically hedging equity‑index option exposure under realistic transaction costs and position limits. By hedging a standardized option‑implied stock exposure (one unit of underlying delta offset by SPY) using only option surface and macro variables as state information, the approach avoids direct pricing engines. Building on the “deep hedging” paradigm, a leak‑free environment, cost‑aware reward function, and lightweight stochastic actor‑critic agents are designed. Agents are trained on end‑of‑day panel data constructed from SPX/SPY implied volatility term structures, skew, realized volatility, and macro‑rate environments. On a fixed train/validation/test split, the learned strategies outperform no‑hedge, momentum, and volatility‑target baselines in risk‑adjusted performance (higher point‑estimate Sharpe ratios). Only the GAE strategy’s test‑sample Sharpe ratio is statistically distinct from zero, though its confidence interval overlaps the long‑run SPY benchmark, so the authors do not claim formal superiority. Turnover robustness is demonstrated under doubled transaction costs. The modular codebase includes data pipelines, simulators, and training scripts, intended to extend to multi‑asset coverage, alternative objectives (e.g., drawdown or CVaR), and intraday data. From a portfolio‑management perspective, the learned overlay aims to sit atop existing SPX or SPY allocations, improving the mean‑variance trade‑off by controlling turnover and drawdown.

Stochastic Volatility Modelling with LSTM Networks: A Hybrid Approach for S&P 500 Index Volatility Forecasting

Paper link: https://arxiv.org/pdf/2512.12250v1

Authors: Anna Perekhodko, Robert Ślepaczuk

Abstract: Accurate volatility prediction is crucial for banking, investment, and risk management because expectations of future market movements directly influence current decisions. This study proposes a hybrid modelling framework that combines stochastic volatility (SV) models with long‑short‑term memory (LSTM) neural networks. The SV component improves statistical precision and captures latent volatility dynamics, especially in response to unexpected events, while the LSTM enhances the model’s ability to detect complex nonlinear patterns in financial time series. Forecasts are generated using daily S&P 500 data spanning 1 Jan 1998 to 31 Dec 2024. A rolling‑window training scheme produces one‑step‑ahead volatility forecasts. Performance is evaluated via statistical tests and investment simulations. Results show the hybrid SV‑LSTM method outperforms standalone SV and LSTM models, contributing to volatility modelling techniques and providing a foundation for improved risk assessment and strategic investment planning for the S&P 500 index.

Dynamic Stacking Ensemble Learning with Investor Knowledge Representations for Stock Market Index Prediction Based on Multi‑Source Financial Data

Paper link: https://arxiv.org/pdf/2512.14042v1

Authors: Ruize Gao, Mei Yang, Yu Wang, Shaoze Cui

Abstract: Significant pattern differences exist across diverse financial data sources, and investors exhibit heterogeneous cognitive behaviours when processing information. To capture these varied patterns, the authors propose a novel two‑stage dynamic stacking ensemble model based on investor‑knowledge representations, designed to efficiently extract and integrate features from multi‑source financial data. In stage one, distinct financial attributes from global equity indices, industrial indices, and financial news are identified, and neural network architectures tailored to each attribute generate effective feature representations. Based on the learned representations, multiple meta‑classifiers are constructed, and the optimal one is dynamically selected for each time window, enabling the model to capture and learn prominent patterns that emerge over different periods. Evaluation on daily movements of China’s SSE, SZEC, and GEI indices demonstrates that the model improves prediction performance, outperforming the best competing models by 1.42 %, 7.94 %, and 7.73 % respectively on accuracy metrics. A trading strategy built on the proposed model also yields superior cumulative returns and Sharpe ratios compared with competing strategies.

Adaptive Weighted Genetic Algorithm‑Optimized SVR for Robust Long‑Term Forecasting of Global Stock Indices for Investment Decisions

Paper link: https://arxiv.org/pdf/2512.15113v1

Author: Mohit Beniwal

Abstract: Long‑term price forecasting remains a major challenge due to inherent uncertainty over extended horizons, despite successes in short‑term prediction. Accurate long‑term forecasts are vital for high‑net‑worth individuals, institutional investors, and traders. This paper presents an Improved Genetic Algorithm‑optimized Support Vector Regression (IGA‑SVR) model specifically for long‑term price forecasting of global indices. IGA‑SVR’s performance is rigorously evaluated against state‑of‑the‑art baselines—Long Short‑Term Memory (LSTM) and Forward‑Validated GA‑optimized SVR (OGA‑SVR). Over 2021‑2024, five global indices (India Nifty, DJIA, Germany DAX, Japan N225, and China SSE) undergo a year‑long daily price forecasting test. Overall, IGA‑SVR reduces Mean Absolute Percentage Error (MAPE) by 19.87 % relative to LSTM and by 50.03 % relative to OGA‑SVR, demonstrating superior accuracy and computational efficiency (LSTM’s runtime is ~20× that of IGA‑SVR). The genetic algorithm selects optimal SVR hyper‑parameters by minimizing the arithmetic mean of MAPE computed on the full training set and the most recent five‑year subset, preserving long‑term trend information while adapting to recent trends, thereby achieving better generalization than LSTM and OGA‑SVR’s rolling‑forward validation, which forgets long‑term trends and suffers recent‑bias.

Interpretable Deep Learning for Stock Returns: A Consensus‑Bottleneck Asset Pricing Model

Paper link: https://arxiv.org/pdf/2512.16251v1

Authors: Bong‑Gyu Jang, Younwoo Jeong, Changeun Kim

Abstract: The authors introduce a partially interpretable neural network—the Consensus‑Bottleneck Asset Pricing Model (CB‑APM)—which captures how dispersed investor beliefs are compressed into asset prices through a consensus‑formation process, thereby emulating sell‑side analyst reasoning. By simulating this “bottleneck,” the model aggregates company‑ and macro‑level information in a structured, interpretable manner. CB‑APM not only predicts future risk premia of US stocks but also links belief aggregation to expected returns in a transparent way. The model improves long‑term return prediction performance and surpasses standard deep learning methods in both predictive accuracy and interpretability. Comprehensive portfolio analysis shows that CB‑APM’s out‑of‑sample predictions translate into economically meaningful gains, exhibiting monotonic return differentials and stable long‑short performance under regularization. Empirically, CB‑APM uses consensus as a regularizer to enhance long‑term predictability and yields interpretable consensus‑based components that clarify how information is priced into returns. Regression and GRS pricing diagnostics reveal that the learned consensus representation only partially overlaps with traditional factor model price variations, indicating that CB‑APM uncovers belief‑driven return structures beyond the typical factor space. Overall, CB‑APM provides an interpretable, empirically grounded framework for understanding belief‑driven return dynamics.