Intraday machine learning for the securities market

A major issue in financial market trading is knowing when to undertake a transaction for the purpose of generating profit. Trading on large national/international financial systems can be analysed by various sophisticated techniques including neural networks. This work focuses on using deep neural networks and other machine learning tools to analyse financial markets (such as Forex which specialises in trading currencies and is the focus of this work with respect to EUR/US Dollar rates) by identifying patterns in the behaviour of major financial market participants: funds and market makers. Current techniques have drawbacks in that market uncertainty limits the confidence traders have in such predictive aides. This research investigates the use of convolutional neural networks to identify subtle patterns that precede significant movements in financial markets. A new approach is taken which focuses on intraday trading features in order to reduce the risks associated with overnight price gaps which have increased in recent years due to financial instability and the COVID-19 pandemic. Particular emphasis is placed on the advanced preliminary analysis of big financial data, including all minimal price changes (ticks) and all transaction volumes, before feeding them into various neural network architectures. An innovative approach to predicting financial markets is described based on the vector of the probability of significant price movements. This makes it possible for the analysis to easily transition from a standard regression task that predicts prices to a classification task, partly mitigating a common issue of balancing re-training frequency versus application. Critical to this approach is the ability to identify the intensity and intraday volatility based on time intervals between each trade. So time analysis is added to the commonly analysed variables of price and volume to reduce the probability of received losses due to stop-loss orders. The big data processing uses open platforms with GPU processors, and the current work also presents a novel method for reducing the amount of data for training neural networks.