High-Frequency trading strategy based on deep neural networks

Recent conceptual and engineering breakthroughs in Machine Learning (ML), particularly in Deep Neural Networks (DNN), have revolutionized the Computer Science field and have been responsible for astonishing breakthroughs in computer vision, speech recognition, facial recognition, transaction fraud detection, automatic translation, video object tracking, natural language processing, and robotics, virtually disrupting every aspect of our lives. The financial industry has not been oblivious to this revolution; since the introduction of the first ML techniques, there have been efforts to use them as financial modeling and decision tools rendering in some cases limited and other in cases useful results, but overall, not astonishing results as in other areas. Some of the most challenging problems for ML come form finance, for instance, price prediction whose solution will require not only the most advanced ML techniques but also other non-standard and uncommon methods and techniques, giving the origin of a new field called Financial ML, whose name has been coined by Lopez de Prado last year. Today, many hedge funds and investment banks have ML divisions, using all kinds of data sources and techniques, to develop financial modeling and decision tools. Consequently, ML is a part of the present and probably will be the future of the financial industry. In this thesis, we use the Deep Neural Networks (DNN) and Recurrent Neural Networks (RNN), two of the most advanced ML techniques, whose learning capabilities are enhanced using the representational power of the Discrete Wavelet Transform (DWT), to model and predict short-term stock prices showing that these techniques allow us to develop exploitable high-frequency trading strategies. Since high-frequency financial (HF) data are expensive, difficult to access, and immense (Big Data), there is no standard dataset in Finance or Computational Finance. Therefore, the chosen testing dataset consists of the tick-by-tick data of 18 well-known companies from the Dow Jones Industrial Average Index (DJIA). This dataset has 348.98 millions of transactions (17 GB) from January 2015 to July 2017. After a long iterative process of data exploration and feature engineering, several features were tested and combined. The tick-by-tick data are preprocessed and transformed using the DWT with a Haar Filter. The final features consist of the sliding windows of two variables: one-minute pseudo-log-returns (the logarithmic difference between one-minute average prices) and the features generated by the DWT. These transformations, which are non-standard data transformations in finance, will better represent the high-frequency behavior of Financial Time Series (FTS). Moreover, the DNN predicts the next one-minute pseudo-log-return that can be transformed into the next predicted one-minute average price. These prices will be used to build a high-frequency trading strategy that buys (sells) when the next one-minute average price prediction is above (below) the last one-minute closing price. Results show that (i) the proposed DNN achieves a highly competitive prediction performance in the price prediction domain given by a Directional Accuracy (DA) ranging from 64% to 72%. (ii) The proposed strategy yields positive profits, a max draw-down less or equal to 3%, and an annualized volatility ranging from 3% to 9% for all stocks. The main contribution is the innovative approach for predicting FTS. It includes the combination of the advanced learning capabilities of the Deep Recurrent Neural Networks (DRNNs), the representational power in frequency and time domains of the DWT, and the idea of modeling time series through average prices.