An Approach Toward Stock Market Prediction and Portfolio Optimization in Indian Financial Sectors

In this article, we aim at predicting future stock price movements and recommending a profitable portfolio for the NIFTY-50 stocks. Stock market prediction is a challenging task due to multiple influencing factors, its nonlinear and volatile nature, and complex interdependencies. Recent approaches have neglected the interconnections between stocks and relied on predefined static relationships. The collection of relational data is difficult to access due to confidentiality and privacy agreements for emerging economies. Moreover, these predefined relationships lack the ability to explain the latent interactions between stocks. This work proposes a data-driven end-to-end framework, dynamic relation aware relational temporal network (DR2TNet), that learns the hidden intra- and intersector associations between stock pairs and temporal patterns. A financial knowledge graph is built from historical data and is updated dynamically during the training process to reflect the interactions between the stocks according to the current market situation. We have proposed a new loss function that considers prediction loss and directional movement loss to train a model. The applicability of prediction results obtained by DR2TNet is demonstrated in the portfolio optimization problem. The results show a higher return compared to other existing baseline models.