Weekly Quantitative Finance Paper Roundup (Mar 21‑27, 2026)

Designing Agentic AI‑Based Screening for Portfolio Investment

Paper link: http://arxiv.org/pdf/2603.23300v1

Authors: Mehmet Caner, Agostino Capponi, Nathan Sun, Jonathan Y. Tan

Abstract: The authors propose a three‑layer agentic AI platform for portfolio management. Two large‑language‑model (LLM) agents are assigned distinct tasks: one screens companies with desirable fundamentals, the other screens companies with favorable news sentiment. The agents negotiate to generate and agree on buy‑sell signals, dramatically shrinking the candidate asset pool. A high‑dimensional precision matrix estimator then determines optimal portfolio weights. A defining theoretical feature is that the number of assets in the portfolio is itself a random variable produced by the screening process. The concept of “reasonable screening” is introduced, and the authors prove that, under small screening errors, the squared Sharpe ratio of the screened portfolio consistently estimates the target. Empirically, the method outperforms an unfiltered baseline and traditional screening methods on S&P 500 data from 2020‑2024, achieving a higher Sharpe ratio.

Hybrid Wavelet‑Based Physics‑Informed Neural Network for Portfolio Management

Paper link: http://arxiv.org/pdf/2603.21834v1

Authors: Bahadur Yadav, Mahaprasad Mohanty, Ratikanta Behera, Sanjay Kumar Mohanty

Abstract: The authors introduce a hybrid wavelet‑based physics‑informed neural network (HW‑PINNs) framework for portfolio management, offering an alternative to conventional PINNs. They first discuss the Merton jump‑diffusion model and the general HW‑PINNs framework, then focus on the one‑dimensional European option case. The HW‑PINN is adapted to the Merton model using a simplified direct coefficient optimization strategy, a mathematically corrected log‑space formulation, and an FFT‑based integral‑differential operator for efficient computation. Numerical experiments on realistic market scenarios demonstrate high accuracy and robustness; under low jump intensity, the average relative error versus a high‑fidelity benchmark is 0.27%. The results confirm that the HW‑PINN implementation is computationally efficient and reliable for pricing derivatives in high‑jump‑risk markets. The paper also discusses risk analysis using Value‑at‑Risk (VaR) and Conditional VaR (CVaR), providing insights into downside risk across market conditions.

FinRL‑X: An AI‑Native Modular Infrastructure for Quantitative Trading

Paper link: http://arxiv.org/pdf/2603.21330v1

Code link: https://github.com/AI4Finance-Foundation/FinRL-Trading

Authors: Hongyang Yang, Boyu Zhang, Yang She, Xinyu Liao, Xiaoli Zhang

Abstract: The authors present FinRL‑X, a modular and deployment‑consistent trading architecture that unifies data processing, strategy construction, backtesting, and brokerage execution under a weight‑centric interface. Existing open‑source platforms are typically backtest‑ or model‑centric and lack system‑level consistency between research evaluation and real‑time deployment. FinRL‑X addresses this gap with a composable strategy pipeline that integrates stock selection, portfolio allocation, timing, and portfolio‑level risk coverage within a unified protocol. The framework supports both rule‑based and AI‑driven components, including reinforcement‑learning allocators and LLM‑based sentiment signals, without altering downstream execution semantics. FinRL‑X thus provides a scalable foundation for reproducible, end‑to‑end quantitative trading research and deployment.

S³G: Stock State Space Graph for Enhanced Stock Trend Prediction

Paper link: http://arxiv.org/pdf/2603.24236v1

Authors: Yao Lu, Kaiyi Hu, Luyan Zhang

Abstract: Stock trend prediction attracts attention due to its potential investment returns. Researchers recognize that inter‑stock relationships (e.g., industry membership, upstream‑downstream ties) are important for capturing market dynamics. Prior work often relies on static industry graphs or per‑time‑step similarity‑based graphs, overlooking the evolving liquidity of stock relationships. The authors observe that competitive and cooperative interactions cause fine‑grained, time‑varying dependencies that static or snapshot graphs cannot capture. To address this, they introduce the Stock State Space Graph (S³G) framework. First, a wavelet transform denoises financial series and extracts salient patterns. Then, at each time point, a data‑dependency graph is constructed and modeled with a state‑space model to capture its dynamic evolution. Finally, graph aggregation yields predictive returns. Extensive experiments on historical CSI 500 data show that S³G achieves state‑of‑the‑art performance, delivering higher annualized returns and Sharpe ratios than baseline methods.