Abstract/Summary

Effective investment requires adeptly managing trading activities that can adapt to diverse market changes and risks. Key stock market metrics such as volatility, daily returns, and trading volumes play a crucial role in numerous pricing models and trading strategies, including risk management. However, existing tools used to predict these metrics have shown limitations, as evident from recent financial crises. As a result, researchers are examining new methods that leverage machine learning and artificial intelligence to address these weaknesses. This research makes a significant contribution to the existing body of work on using Deep Learning (DL) for predicting stock market metrics. Different multivariate and multitasking DL architectures are introduced to enhance the prediction accuracy for various future prediction time horizons and different market conditions. For comparison and benchmarking, the models have been assessed against traditional statistical methods and commonly used DL networks. The results validate the effectiveness of deep learning models in modelling stock market volatility, demonstrating their predictive capability in both upward and downward market conditions across short and long datasets. Furthermore, the study illustrates that deploying multivariate deep learning enhances stock market volatility prediction compared to single-input models. This improvement arises from the utilisation of positively correlated input data and larger datasets, enabling the models to extract crucial information during training and thereby enhancing prediction accuracy. The deployment of various multi-task deep learning models with shared input layers resulted in significant enhancements in predicting stock market volatility, daily returns, and trading volumes. This improvement stems from the optimisation of the loss function across all output tasks simultaneously. Determining the optimal combination of weights and inputs required an initial step of assigning equal weights to two inputs, followed by an iterative testing process. This iterative testing included the adjustment of the number of inputs and the allocation of weights. The multi-task models demonstrated superior performance relative to both statistical models and single-task deep learning models, including those with multivariate input.