Key Quantitative Finance Papers (Dec 6‑12 2025) – AI‑Driven Insights

Long‑only cryptocurrency portfolio management by ranking the assets: a neural network approach

Authors propose a novel machine‑learning‑based cryptocurrency portfolio management method that predicts the future return ranking of a set of assets at each timestep using a neural network and allocates weights accordingly. Back‑testing on daily market data from May 2020 to November 2023—covering a full bull, bear, and stagnant cycle—shows profitability, outperforming existing methods with a Sharpe ratio of 1.01 and an annual return of 64.26%, and robustness to higher transaction costs.

Local and Global Balance in Financial Correlation Networks: an Application to Investment Decisions

The paper investigates using local balance (a node‑level contribution to overall network balance) and deviation from global balance as criteria for selecting assets that outperform the market. During crises, most assets move together, maximizing global balance and limiting diversification benefits. Concentrating exposure on assets whose local balance strongly deviates from the network’s global balance can reduce risk. Empirical tests on real financial data support this intuition in both descriptive and predictive settings.

Exploratory Mean‑Variance with Jumps: An Equilibrium Approach

Reexamining the continuous‑time mean‑variance (MV) portfolio problem, the authors model market dynamics with a jump‑diffusion process and apply reinforcement learning (RL) to enable informed exploration. Recognizing the MV problem’s time inconsistency, they adopt a time‑inconsistent control (TIC) framework to derive an exploratory equilibrium strategy centered on a Gaussian distribution around the classic MV equilibrium. An Actor‑Critic RL algorithm is designed, with parameters converging to true values in simulation. Numerical experiments on 24 years of real market data show the RL model is profitable in 13 of 14 tests, demonstrating practical applicability.

Reinforcement Learning in Financial Decision Making: A Systematic Review of Performance, Challenges, and Implementation Strategies

This systematic review analyzes 167 papers published between 2017 and 2025 on reinforcement learning (RL) for market making, portfolio optimization, and algorithmic trading. It identifies key performance issues and challenges, noting that RL generally outperforms traditional methods in market‑making tasks. The authors propose a unified framework addressing interpretability, robustness, and deployment feasibility. Empirical evidence with synthetic data indicates that implementation quality and domain knowledge often matter more than algorithmic complexity. The study calls for interpretable RL architectures for regulatory compliance, enhanced robustness in non‑stationary environments, and standardized benchmarking protocols.

A Hybrid Model for Stock Market Forecasting: Integrating News Sentiment and Time Series Data with Graph Neural Networks

The authors present a multimodal approach that combines company news headlines with historical stock prices. Historical price sequences are encoded by an LSTM, while headlines are embedded via a language model; these embeddings become nodes in a heterogeneous graph processed by GraphSAGE to capture interactions among articles, companies, and industries. Experiments on U.S. stocks and a Bloomberg dataset show the GNN outperforms the LSTM baseline, achieving 53 % accuracy on a binary direction‑change task and a 4 % precision gain on an importance‑based task. Companies with more news coverage obtain higher prediction accuracy, and concise headlines prove more predictive than full articles.

Quantum Temporal Convolutional Neural Networks for Cross‑Sectional Equity Return Prediction: A Comparative Benchmark Study

To tackle noise, regime shifts, and limited generalization in stock prediction, the authors propose a Quantum Temporal Convolutional Neural Network (QTCNN). A classical time encoder extracts multi‑scale patterns from sequential technical indicators, and a parameter‑efficient quantum convolution circuit leverages superposition and entanglement to enrich feature representations and suppress overfitting. Benchmarking on the JPX Tokyo Stock Exchange dataset, QTCNN achieves an out‑of‑sample Sharpe ratio of 0.538—about 72 % higher than the best classical baseline—demonstrating the practical potential of quantum‑enhanced models for robust quantitative finance decisions.

Analysis of Contagion in China's Stock Market: A Hawkes Process Perspective

The study applies Hawkes processes to model autocorrelation and cross‑correlation in multivariate daily returns of the Shanghai Composite, Shenzhen Component, ChiNext, and industry indices (e.g., consumer, healthcare, finance). Fitting self‑exciting and inhibiting Hawkes processes reveals long‑term dependencies, including upward, downward, and oversold rebound trends. Results indicate that during high‑activity periods, industry indices tend to maintain their trends, whereas low‑activity periods exhibit strong sector rotation. The spatio‑temporal Hawkes framework deepens understanding of financial contagion by linking conditional intensity functions to sector‑rotation behavior.

Unveiling Hedge Funds: Topic Modeling and Sentiment Correlation with Fund Performance

Using a unique dataset of over 35,000 documents from more than 1,125 hedge‑fund managers, the authors apply three topic‑modeling methods—LDA, Top2Vec, and BERTopic. LDA with 20 topics provides the most interpretable results for humans and shows higher robustness to changes in topic number, while Top2Vec achieves superior classification performance. A novel quantitative framework links document sentiment to fund performance. Contrary to expectations, the generic DistilBERT model outperforms the finance‑specific FinBERT in sentiment scoring, better capturing diverse language patterns in hedge‑fund filings. Sentiment scores derived from DistilBERT combined with Top2Vec exhibit the strongest correlation with subsequent fund performance, suggesting that automatic topic modeling and sentiment analysis can furnish data‑driven decision‑support tools for investors.

Knowledge‑Augmented Large Language Model Agents for Explainable Financial Decision‑Making

The paper proposes a framework that augments large language model (LLM) agents with external knowledge retrieval, semantic representation, and reasoning generation to improve factual consistency and provide transparent reasoning chains. Financial text and structured data are encoded into semantic vectors; relevant information is retrieved from external knowledge bases via similarity search and fused with internal representations through weighted combination. A multi‑head attention mechanism constructs logical chains during generation, enabling causal traceability. Joint optimization of task performance and explanation consistency enhances both prediction accuracy and interpretability. Experiments on financial text processing and decision tasks demonstrate superiority over baselines in accuracy, generation quality, and factual support.

HiveMind: Contribution‑Guided Online Prompt Optimization of LLM Multi‑Agent Systems

HiveMind introduces an adaptive framework that optimizes collaboration among LLM‑based agents by quantifying each agent’s contribution using Shapley values. To address the computational burden of classic Shapley calculations, the authors devise DAG‑Shapley, an efficient attribution algorithm that exploits the directed‑acyclic‑graph structure of agent workflows, pruning infeasible coalitions and reusing intermediate outputs. In a multi‑agent stock‑trading scenario, HiveMind outperforms a static baseline, and DAG‑Shapley reduces LLM calls by over 80 % while preserving attribution accuracy comparable to full Shapley, establishing a new standard for credit allocation and large‑scale multi‑agent optimization.