Real Time Distributed Stock Market Forecasting using Feed-Forward Neural Networks, Market Orders, and Financial indicators

Abstract

Machine learning and mathematical models are two tools used in prior research of stock predictions. However, the stock market provides enormous data sets, making machine learning an expensive and slow task, and a solution to this is to distribute the computations. The input to the machine learning in this thesis uses market orders, which is a different way to make short-term predictions than previous work. Distributing machine learning in a modular configuration is also implemented in this thesis, showing a new way to combine predictions from multiple models. The models are tested with different parameters, with an input base consisting of a list of the latest market orders for a stock. The distributed system is divided into so-called node-boxes and tested based on latency. The distributed system works well and has the potential to be used in large systems. Unfortunately, making predictions with market orders in neural networks does not provide good enough performance to be viable. Using a combination of predictions and financial indicators, however, shows better results.