Використання еволюціонуючої графової нейронної мережі для прогнозування цін на фондовому ринку

Abstract

An evolving graph neural network is proposed for stock price forecasting on financial markets, in which the relationship graph structure is generated and evolved using a Dendritic Artificial Immune Network (DaiNet). Price sequences are encoded via transformers and aggregated through a temporal graph attention layer, with final prediction performed by an MLP. Experiments on 16 technology companies show the evolving graph approach outperforms static methods by 5–10%, achieving ~65% accuracy