Learning to Trade: Deep Neural Networks for Robust Stock Market Price Forecasting

The work presented here closes a number of important gaps within the body of existing scholarship on financial time series forecasting by giving a thorough and equitable comparison between LSTM and Transformer architectures when tasked with forecasting noisy, scarce datasets which closely model actual conditions. The idea and experimental validation of a Hybrid LSTM–Transformer Network will demonstrate the effectiveness of true innovations, as opposed to optimizing “new objectives,” which are now made possible through such combinations and are effectively coupling memory and attention. Further, the emphasis placed on attention in order to have clearer analysis and better evaluation tools will carry the sub-field of deep learning forwards, more so as regards finance, where it comes nearer to transparency, which, despite its significance, remains a topic that has been neglected in many bibliographic examples, and which highlights the potential importance of model evaluation, in which indicators of success ultimately have more to do with something beyond projection, but the strength, sensibility, and effectiveness of these as well. As algorithmic decision-making becomes increasingly prevalent and enters the system of global trade, allocation, and protection, finance requires assurance of predictable access to forecasting models satisfying these requirements, and therefore this work, and these new paradigms within which it is situated, are central to new applications of finance and new applications of AI.