Lightweight dynamic graph reinforcement learning for adaptive stock portfolio optimization

Traditional graph neural network based portfolio optimization methods rely on fixed topological structures to model stock correlations, which fail to capture the structural changes in stock linkages during industry rotation and market style shifts. Moreover, full-scale dynamic graph approaches incur prohibitive computational overhead, making them inefficient for the iterative training of deep reinforcement learning on large-scale asset pools. To address these limitations, this paper proposes a lightweight dynamic graph reinforcement learning (LDGRL) framework for adaptive stock portfolio optimization. The framework first constructs a dynamically evolving stock correlation graph based on rolling return correlations, capital flow patterns, and market sentiment signals. A dual pruning mechanism including adaptive edge filtering and node screening is adopted to generate sparse topology, enabling dynamic updates of graph connections. Notably, the global asset pool remains fixed to ensure consistent reinforcement learning action dimensions. We propose a unified triple lightweight design strategy: dynamic topology dual pruning, local consistent subgraph sampling, and shallow sparse graph convolution combined with principal component analysis feature dimensionality reduction. This design reduces computational complexity while preserving critical correlation information. The dynamic graph embeddings are then fed into an improved twin trust region policy optimization agent to generate dynamic position allocation decisions. A unique topology-decision closed-loop mechanism is introduced to dynamically adjust key hyperparameters of graph reconstruction using investment returns and risk feedback. Extensive experiments on the CSI 300 index constituent stocks show that LDGRL achieves higher cumulative return and lower maximum drawdown than state-of-the-art deep reinforcement learning approaches, with a competitive Sharpe ratio.

Authors

Institutions

Publication Details

Journal
Scientific Reports
Published
2026-11-17
DOI
https://doi.org/10.1038/s41598-026-80621-y
Primary Topic
Stock Market Forecasting Methods
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Lightweight dynamic graph reinforcement learning for adaptive stock portfolio optimization

Shaolong Chen, Xuan Chen
3 links
Scientific Reports
Stock Market Forecasting Methods
article

Lightweight dynamic graph reinforcement learning for adaptive stock portfolio optimization

Shaolong Chen, Xuan Chen
article en

Abstract

Traditional graph neural network based portfolio optimization methods rely on fixed topological structures to model stock correlations, which fail to capture the structural changes in stock linkages during industry rotation and market style shifts. Moreover, full-scale dynamic graph approaches incur prohibitive computational overhead, making them inefficient for the iterative training of deep reinforcement learning on large-scale asset pools. To address these limitations, this paper proposes a lightweight dynamic graph reinforcement learning (LDGRL) framework for adaptive stock portfolio optimization. The framework first constructs a dynamically evolving stock correlation graph based on rolling return correlations, capital flow patterns, and market sentiment signals. A dual pruning mechanism including adaptive edge filtering and node screening is adopted to generate sparse topology, enabling dynamic updates of graph connections. Notably, the global asset pool remains fixed to ensure consistent reinforcement learning action dimensions. We propose a unified triple lightweight design strategy: dynamic topology dual pruning, local consistent subgraph sampling, and shallow sparse graph convolution combined with principal component analysis feature dimensionality reduction. This design reduces computational complexity while preserving critical correlation information. The dynamic graph embeddings are then fed into an improved twin trust region policy optimization agent to generate dynamic position allocation decisions. A unique topology-decision closed-loop mechanism is introduced to dynamically adjust key hyperparameters of graph reconstruction using investment returns and risk feedback. Extensive experiments on the CSI 300 index constituent stocks show that LDGRL achieves higher cumulative return and lower maximum drawdown than state-of-the-art deep reinforcement learning approaches, with a competitive Sharpe ratio.

Scientific Reports
Universiti Sains Malaysia (MY), Sun Yat-sen University (CN)
Department of Education of Guangdong Province
Stock Market Forecasting Methods
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon