Stock price prediction and portfolio optimisation using deep learning techniques

Abstract

Stock market analysis is a well-explored field, with accurate prediction becoming newlineincreasingly essential. Forecasting stock market behavior is a complex task that requires a detailed examination of data patterns. Deep learning techniques represent the latest advancements in this domain, and their application in stock market prediction and portfolio optimisation is gaining significant traction in research. This study focuses on stock market dynamics in developing economies, which are generally considered less stable than their developed counterparts. The research is structured in two stages. In the first stage, a novel multi-output stacked LSTM model is developed for daily predictions of NIFTY stocks. While previous studies primarily employed LSTM for single-output stock market forecasting, this study overcomes those limitations by predicting the entire OHLC price range, delivering a more holistic outlook of price dynamics. The model integrates historical prices and STIs as input features and applies PCA for dimensionality reduction. To evaluate performance, the model is benchmarked against existing studies using metrics such as MAE, MSE, RMSE, and accuracy rate. In the second stage, stocks selected based on predicted returns from the first stage are used to construct 30 different portfolios, each consisting of the top 7, 8, 9, and 10 return newlinegenerating NIFTY stocks. These portfolios are then assessed based on risk and returns newlinemetrics. The results indicate that portfolios with five stocks yield the highest returns, newlinewhile increasing the portfolio size beyond nine stocks leads to excessive liversification and complexity. Consequently, the findings suggest that the proposed two-stage portfolio optimisation approach effectively balances historical and predictive asset information, making it a promising investment strategy.