Weekly Quantitative Finance Paper Digest (Aug 23‑29, 2025)

Achieving High-Quality Portfolio Optimization with the Variational Quantum Eigensolver

Paper link: http://arxiv.org/pdf/2508.18625v1

Authors: Anbang Wang, Zhonggang Lv, Zhenyuan Ma, Dunbo Cai, Zhihong Zhang

Abstract: Portfolio optimization is formulated as an unconstrained binary quadratic optimization (QUBO) problem, which is NP‑hard. The authors apply the Variational Quantum Eigensolver (VQE) to solve this problem. To increase the likelihood of converging to high‑quality solutions, they use Weighted Conditional Value‑at‑Risk (WCVaR) as the cost function and the Covariance Matrix Adaptation Evolution Strategy (CMA‑ES) as the optimizer. Experiments on the Wuyue QuantumAI platform using classical simulation show that the combination of WCVaR and CMA‑ES improves performance.

THEME: Enhancing Thematic Investing with Semantic Stock Representations and Temporal Dynamics

Paper link: http://arxiv.org/pdf/2508.16936v1

Authors: Hoyoung Lee, Wonbin Ahn, Suhwan Park, Jaehoon Lee, Minjae Kim, Sungdong Yoo, Taeyoon Lim, Woohyung Lim, Yongjae Lee

Abstract: The authors construct a Theme Representation Set (TRS) by starting from real‑world thematic ETFs and expanding them with industry classifications and financial news. TRS provides explicit mappings from themes to constituent stocks and rich textual information for each theme. Based on TRS, they introduce THEME, a hierarchical contrastive learning framework. Themes and stocks are embedded as text; THEME first aligns them semantically using hierarchical relations, then refines the embeddings with a temporal stage that incorporates individual stock returns. Empirical results show that THEME outperforms strong baselines on multiple retrieval metrics and significantly improves portfolio construction performance.

QTMRL: An Agent for Quantitative Trading Decision‑Making Based on Multi‑Indicator Guided Reinforcement Learning

Paper link: http://arxiv.org/pdf/2508.20467v1

Code link: https://github.com/ChenJiahaoJNU/QTMRL.git

Authors: Xiangdong Liu, Jiahao Chen

Abstract: The authors build a comprehensive multi‑indicator dataset for 16 representative S&P 500 stocks using 23 years of daily OHLCV data (2000‑2022), covering five sectors and enriching raw data with trend, volatility, and momentum indicators. They design a lightweight reinforcement‑learning framework based on the Advantage Actor‑Critic (A2C) algorithm, comprising data processing, the A2C algorithm, and a trading‑agent module to support policy learning and actionable decisions. Extensive experiments compare QTMRL with nine baselines (e.g., ARIMA, LSTM, moving‑average strategies) across various market conditions, demonstrating superior profitability, risk‑adjusted returns, and downside‑risk control.

Is attention truly all we need? An empirical study of asset pricing in pretrained RNN sparse and global attention models

Paper link: http://arxiv.org/pdf/2508.19006v1

Author: Shanyan Lai

Abstract: The study evaluates pretrained RNN attention models—including additive attention, three Luong variants, global self‑attention, and sliding‑window sparse attention—on asset‑pricing experiments for the top 420 U.S. large‑cap stocks. Causal masking prevents future‑data leakage. The models also address temporal sparsity and mitigate over‑fitting via simplified architectures. Evaluations span three periods (pre‑COVID‑19, COVID‑19, post‑COVID) to test stability under extreme market conditions. In market‑cap‑weighted portfolio backtests, Self‑Attention and Sparse‑Attention achieve annualized Sortino ratios of 2.0 and 1.80 respectively during COVID‑19, with Sparse‑Attention showing more stable absolute returns across market‑cap sizes.

Identifying Risk Variables From ESG Raw Data Using A Hierarchical Variable Selection Algorithm

Paper link: http://arxiv.org/pdf/2508.18679v1

Authors: Zhi Chen, Zachary Feinstein, Ionut Florescu

Abstract: The authors propose a Hierarchical Variable Selection (HVS) algorithm to extract a concise set of risk‑relevant variables from raw ESG data, which typically has many more variables than observations and a tree‑structured organization. Experiments show that HVS outperforms models using aggregated ESG scores and achieves superior explanatory power with a more compact variable set compared to traditional variable‑selection methods.

Forecasting Probability Distributions of Financial Returns with Deep Neural Networks

Paper link: http://arxiv.org/pdf/2508.18921v1

Author: Jakub Michańków

Abstract: One‑dimensional convolutional neural networks (CNN) and long short‑term memory (LSTM) architectures are used to predict parameters of three distributions—normal, Student‑t, and skewed Student‑t—via a custom negative log‑likelihood loss. Models are evaluated on six major equity indices (S&P 500, BOVESPA, DAX, WIG, Nikkei 225, KOSPI) using Log‑Predictive Score (LPS), Continuous Ranked Probability Score (CRPS), and Probability Integral Transform (PIT). Results indicate that deep learning provides accurate distribution forecasts and competes with classic GARCH models in Value‑at‑Risk estimation. Across metrics, the LSTM with skewed Student‑t distribution performs best, capturing heavy tails and asymmetry.

FinCast: A Foundation Model for Financial Time‑Series Forecasting

Paper link: http://arxiv.org/pdf/2508.19609v1

Authors: Zhuohang Zhu, Haodong Chen, Qiang Qu, Vera Chung

Abstract: Financial time‑series forecasting faces challenges from temporal non‑stationarity, multi‑domain diversity (stocks, commodities, futures), and varying time resolutions. Existing deep‑learning methods often overfit and require extensive domain‑specific fine‑tuning. The authors introduce FinCast, the first foundation model dedicated to financial time‑series forecasting, trained on a massive multi‑domain dataset. FinCast demonstrates strong zero‑shot performance, capturing diverse patterns without domain‑specific fine‑tuning, and empirically outperforms existing state‑of‑the‑art methods, highlighting robust generalization.

TradingGroup: A Multi‑Agent Trading System with Self‑Reflection and Data‑Synthesis

Paper link: http://arxiv.org/pdf/2508.17565v1

Authors: Feng Tian, Flora D. Salim, Hao Xue

Abstract: The system comprises specialized agents for news sentiment analysis, financial report interpretation, stock trend prediction, trading‑style adaptation, and a central decision‑making agent that aggregates signals to output buy, sell, or hold actions. Self‑reflection mechanisms distill past successes and failures to guide future reasoning, while a dynamic risk‑management model provides configurable stop‑loss and take‑profit controls. An automated data‑synthesis and annotation pipeline generates high‑quality post‑training data, improving agent performance after fine‑tuning. Backtests on five real‑world stock datasets show that TradingGroup outperforms rule‑based, machine‑learning, reinforcement‑learning, and existing LLM‑based trading strategies.

Combined Machine Learning for Stock Selection Strategy Based on Dynamic Weighting Methods

Paper link: http://arxiv.org/pdf/2508.18592v1

Authors: Lin Cai, Zhiyang He, Caiya Zhang

Abstract: The authors develop a stock‑selection framework that combines three representative machine‑learning models with two weighting schemes: a static weight based on model evaluation metrics and a dynamic weight based on the information coefficient (IC). Using CSI 300 index data, empirical evaluation shows that (1) the ensemble strategy significantly outperforms single‑model approaches in back‑test returns; (2) IC‑based weighting (especially IC_Mean) is more competitive than metric‑based weighting in both returns and prediction performance; (3) factor selection markedly improves the ensemble strategy.