Volume: 19 Issue-01 (January-June) 2024

---

AN INVESTIGATION ON IMPROVED MULTI LEVEL GRAPH ATTENTION BASED GRAPH CONVOLUTION NETWORK FOR STOCK PREDICTION

ABSTRACT

The issued shares are moved, exchanged, and distributed on the stock market, a capitalistic paradise. Although the issue market serves as the foundation for stock pricing, the stock market's structure and trading activities are far more intricate than the issue market's. As a result, predicting the future accurately becomes a complex and challenging endeavour. Conversely, due to the potential rewards associated with stock prediction, it continues to draw in generations of academics and investors, who in turn continue to develop a wide range of prediction techniques from a variety of viewpoints, theories, investment approaches, and real-world experiences. The majority of current graph-based learning techniques ignore the complexity of stock linkages and instead manually construct stock relationships in order to produce stock graphs. In this study, a sentiment analysis module was created with natural language processes and Gated Recurrent Unit (GRU). The factor analysis module receives the outputs from the LSTM, GRU, BERT, and Relations models from wikidata. The sentiment data on the financial market is included into the factor analysis module. Subsequently, a subset of parameters were fed into the recently created Improved Multilevel Graph Attention network (IMGA) to forecast stock prices. Improved Multi level Graph Attention based Deep Stock prediction network (IMGA-DeSnet) is the name of the suggested technique. Graph Convolution Network (GCN) is used by IMGA to extract relational data. The attention encoded feature representation for certain classes is provided by the factor analysis models' extracted features and trained weights. Finally, the fully connected layer receives the concatenated feature vector and utilises a feature map matrix and input weights to forecast stock. The suggested architecture offers a solid and trustworthy stock prediction method.

FULL TEXT

PDF