Weekly Quantitative Finance Paper Summary (Jan 31–Feb 6 2026)

A Novel approach to portfolio construction

Paper link: https://arxiv.org/pdf/2602.03325v1

Authors: T. Di Matteo, L. Riso, M. G. Zoia

Abstract: This paper proposes a machine‑learning‑based asset selection and portfolio construction framework called Best‑Path‑Algorithm Sparse Graph Model (BPASGM). By mapping linear and nonlinear dependencies among many financial assets to a sparse graph satisfying structural Markov properties, BPASGM extends the Best‑Path‑Algorithm (BPA). The representation enables dependency‑driven screening that removes assets with forward or redundant connections, isolating conditionally independent or negatively correlated subsets to enhance diversification and reduce estimation error in high‑dimensional portfolios. Standard mean‑variance optimization is then applied to the selected subset. BPASGM aims not to improve the theoretical mean‑variance optimum under known parameters, but to achieve better out‑of‑sample performance where Markowitz portfolios are sensitive to estimation error. Monte‑Carlo simulations show that BPASGM portfolios exhibit more stable risk‑return profiles, lower realized volatility, and superior risk‑adjusted performance compared with standard mean‑variance portfolios. Empirical results on U.S. stocks, global equity indices, and FX rates from 1990‑2025 confirm these findings and demonstrate a substantial reduction in portfolio cardinality.

The Enhanced Physics‑Informed Kolmogorov‑Arnold Networks: Applications of Newton's Laws in Financial Deep Reinforcement Learning (RL) Algorithms

Paper link: https://arxiv.org/pdf/2602.01388v2

Authors: Trang Thoi, Hung Tran, Tram Thoi, Huaiyang Zhong

Abstract: Deep reinforcement learning (DRL) has become a powerful method for financial trading, generating discrete signals or continuous portfolio allocations. This work introduces a novel RL framework for portfolio optimization that integrates physics‑informed Kolmogorov‑Arnold networks (PIKANs) into several DRL algorithms. The approach replaces traditional multilayer perceptrons with Kolmogorov‑Arnold networks (KANs) that use learnable B‑spline univariate functions in the actor and critic, achieving parameter efficiency and improved interpretability. During actor updates, a physics‑informed regularization loss enforces second‑order temporal consistency between observed return dynamics and action‑induced portfolio adjustments. The framework is evaluated on three stock markets—China, Vietnam, and the United States—covering emerging and developed economies. Across all markets, PIKAN‑based agents consistently deliver higher cumulative and annualized returns, better Sharpe and Calmar ratios, and more favorable drawdown characteristics compared with traditional DRL baselines and classic online portfolio selection methods. The method also yields more stable training and higher Sharpe ratios, particularly valuable in highly dynamic and noisy financial markets where conventional DRL suffers from instability and poor generalization.

GAPNet: Plug‑in Jointly Learning Task‑Specific Graph for Dynamic Stock Relation

Paper link: http://arxiv.org/pdf/2602.00888v1

Authors: Yingjie Niu, Lanxin Lu, Changhong Jin, Ruihai Dong

Abstract: The emergence of networks has transformed financial relational paradigms, with real‑time propagation of news, social discourse, and financial documents reshaping prediction. Existing methods rely on pre‑defined graphs to capture stock relationships, but stock‑related network signals are noisy, asynchronous, and hard to obtain, leading to poor generalization and inconsistency with downstream tasks. To address this, the authors propose GAPNet, a graph‑adaptation plug‑in network that jointly learns task‑specific topology and representations in an end‑to‑end manner. GAPNet attaches to existing pairwise‑graph or hypergraph backbones and dynamically adapts and reconstructs edge topology via two complementary components: a spatial‑aware layer capturing short‑term co‑movement among assets, and a temporal‑aware layer preserving long‑term dependencies under distribution shift. Experiments on two real‑world stock datasets show that GAPNet consistently enhances profitability and stability over state‑of‑the‑art models, achieving annualized cumulative returns up to 0.47 (RT‑GCN) and 0.63 (CI‑STHPAN), with peak Sharpe ratios of 2.20 and 2.12 respectively. Its plug‑and‑play design ensures broad applicability across various GNN‑based architectures.

The article also provides a curated list of additional recent papers on AI‑driven quantitative finance, each with a link to the full text.