Monitoring and prediction of high frequency trading in stock markets using deep reinforced learning, self-organising Machine Learning Framework and feature engineering

Abstract

Forecasting of the stock market is an immanently arduous problem. The Efficient Market Hypothesis (EMH) suggests that financial prices are unpredictable, however recent conceptual and engineering breakthroughs in Machine Learning (ML) have obtained remarkable results on antedating stock movements. Today, many industries covering hedge funds, banks, retail to oil have dedicated divisions to develop financial modelling and decision tools using ML tools and techniques. Stock prices of different industries in general do not follow the same trend, which is a contrast to the trends of prices for stocks belonging to the same industry [211]. Moreover, an accurate prediction of stock price movements is a complex assignment, since it involves multiple pre-conditions, events or factors for a change in the direction of a stock. As suggested by Lopez de Pardo [1], the solution to the most challenging problems in using ML aided financial models or decision tools will require not only cutting-edge ML engineering but also other non-standard and atypical approaches. Hence, it is quintessential to select the optimal financial model which will capture as many pre-conditions or significant factors which effect the stock movements and, is suitable for accurately encapsulating the price movements of stocks irrespective of the industry. This will help to mitigate the need for training multiple ML models and shrink the overhead computational burden, latency and resource utilisation. However, to our best knowledge and an extensive literature review carried out during the research no previous work has addressed all the above challenges. In this thesis, a novel Artificial Intelligence (AI) enabled High Frequency Trading Filter (AI-HFTF) is developed, which functions as a two-staged filter to improve overall Automated Trading system performance (maximum accuracy and minimum error) and efficiency (reduced latency) while predicting stock price movements and making trading decisions. The stage 1 filter is designed to derive an optimal feature set construction by leveraging frequency and meaningful patterns present in the high frequency financial time series. The derived information from stage 1 is then fed as the most significant features sets to stage 2 modelling filters layer 1 regressors and layer 2 deep reinforcement learning agent. The main purpose of the thesis is to attain the ability to derive macroscopic patterns present in financial time series, which will help automatic trading systems to accurately predict stock movement, speed up processing time and make reliable HF trading decisions irrespective of the stock industry. In Step 1, FTSEUK 100 index stocks are chosen, among which the top 10 HF trading companies covering industries such as finance, retail, oil and energy etc are selected for modelling purpose. The research spans over an HFT dataset collected from February 2018 through September 2018. Step 2 performs a few fundamental analysis (for example: seasonality trend analysis, a non-stationarity test, a correlation matrix etc) as part of data exploration prior to pre-processing the dataset. In step 3, stage 1 filtering commences by generating the following features: a) Inter-Company Correlations; b) Stock Market Technical Indicators; c) the ARIMA model and d) the Fourier model as features set (input) for the stage 2 modelling filter. In Step 4, the most significant features from (step-3) are selected using XGBoost, to avoid the curse of dimensionality and attain an accurate prediction of the stock price movement and making better market marking (buy/sell/no action) predictions for the HF times-series. The optimal feature set generated by the stage 1filter is then fed to the stage 2 modelling filter, which selects layer 1or layer 2 depending on the requirement of trading. Layer 1 is selected for predicting stock movements and Layer 2 is selected for buying, selling or no action trading decision making respectively. Since, this research work deal with a regression problem, mean square error (mse), root mean square error (rmse), mean absolute error(mae) and r2-score metrics were used instead of an accuracy for performance analysis. Layer 2 is created with widely used deep reinforcement learning agents which generate trading signals and choose the ideal agent for making reliable trading decisions. In Step 5, the modelling filter was tested using scenario 1-which used the most significant features as input and then the higher performing model was tested under scenarios 2- hyperparameter tuning with significant features and scenarios 3- all the features to draw a comparison for the developed filter .In Step 6, it was concluded that LSTM was the optimal stock price movement prediction trading model for the time-series data and the layer 2 learner policy gradient agent offered optimal improved decision-making ability. For all of the selected companies the filter accurately predicted (maximum 100% and minimum 0.82 %) the stock price movement with minimal error (RMSE = below 1.6, MSE = below 2, MAE = below 1.08) and attained better profit (7.24 % on an average) with a fast execution time (0.024 seconds maximum and 0.016 seconds minimum) for the HF time series. Finally, the results were concluded keeping a balance between efficiency and performance and suggestions for future work scope is given in Step 8.