Ameliorated graph sample and aggregate network and convolutional neural network for stock trading decisions

Abstract

Both graph neural network and convolutional neural network (CNN) perform well in predicting the trading signals of time series data. However, these models often encounter challenges in profitability due to either inadequate information or algorithmic limitations. A model AG-MCNN is proposed by combining ameliorated graph sample and aggregate network (GraphSAGE) and multi-CNN for stock trading decisions. Ameliorated GraphSAGE (AG) is used to learn graph embedding in daily market network constructed from comprehensive market information, with neighbors having higher correlation exerting stronger influence on the target node. Moreover, a more reasonable graph modeling approach is adopted to preserve the important features of time series while filtering out noise. Multi-scale time series and comments with sentiment information are collected to analyze short-term and long-term trends. The short-term features of the weighted graph are extracted by a dual concatenated CNN. The pre-trained model is fine-tuned based on transfer learning to address the issue of insufficient sentiment information in the long-term trend. Experimental results demonstrate that AG-MCNN achieves an average profit rate of 68.4094%, compared to the highest value of 46.1091% for comparison models and 18.2234% for the buy-and-hold strategy. AG-MCNN’s high level of profitability in stock investments makes it a promising choice for investors.