Graph-based approaches are revolutionizing the analysis of different real-life systems, and the stock market is no exception. Individual stocks and stock market indices are connected, and interesting patterns appear when the stock market is considered as a graph. Researchers are analyzing the stock market using graph-based approaches in recent years, and there is a need to survey those works from multiple perspectives. We discuss the existing graph-based works from five perspectives: (i) stock market graph formulation, (ii) stock market graph filtering, (iii) stock market graph clustering, (iv) stock movement prediction, and (v) portfolio optimization. This study contains a concise description of major techniques and algorithms relevant to graph-based approaches for the stock market.

Traditional stock movement prediction tasks are formulated as either classification or regression task, and the relation between stocks are not considered as an input of prediction. The relative order or ranking of stocks is more important than the price or return of a single stock for making proper investment decisions. Stock ranking performance can be improved by incorporating the stock relation information in the prediction task. We employ a graph-based approach for stock ranking prediction and use the stock relation information as the input of the machine learning model. Investors might be interested in the prediction performance of top-k stocks as they would be more profitable than the others. Thus, the performance measure for stock ranking prediction should be top-weighted and bounded for any value of k. Existing evaluation measures lack these properties, and we propose a new measure named normalized rank biased overlap for top-k (N RBO@k) stocks for stock ranking prediction. N RBO@k-based investment strategy generates 0.281% to 4.928% higher relative investment gain than the topmost stock-based strategy. We show that the list-wise loss function can improve the stock ranking performance significantly in a graphbased approach. It generates better N RBO@10 than the combination of point-wise and pair-wise loss in three out of four cases. Node embedding techniques such as Node2Vec can reduce the training time of graph-based approaches for stock ranking prediction significantly. Additionally, we improve the prediction performance through hyperparameter tuning of Node2Vec when a sparse stock relation graph is applied.

No abstract

No abstract

No abstract