Financial Time Series Prediction Using Machine Learning Algorithms: An Application on ADX

  • Fatima K. AlHarmoodi

Student thesis: Master's Thesis

Abstract

Financial applications using machine learning algorithms have been prevalent to demonstrate the power of regression and classification techniques within the industry. Forecasting time series data such as stock price prediction is a challenging yet necessitated application for investors, in order to develop a more advanced investment analysis tool in terms of technicality. A novel application of four algorithms on stocks from Abu Dhabi Securities Exchange (ADX) in the United Arab Emirates (UAE) was conducted in the thesis to predict stock prices as well as validate and compare the techniques, in addition to anticipating the stock price movement. The application uses multivariate linear regression (MVLR), simple exponential smoothing (SES), Holt's method and Holt Winter's method on three modes of prediction: one-step, two-step (two sequential trading days) and three-step (three sequential trading days) on three securities from ADX: Abu Dhabi Commercial Bank (ADCB), Al Dar Properties and Dana Gas Company on the basis that the symbols are actively trading and have generated enough trade log historical data to train and test the different sets of algorithms. Factors of low, high and open price were used in MVLR as predictors for the closing price, while an optimization technique was incorporated to determine the optimum parameters (alpha, beta and gamma) for the exponential smoothing techniques. Within the all the models, a novel integration of moving window analysis was applied to enhance the prediction accuracy of the techniques. Different values were tested to understand the important of historical data training within the models prior to testing the relative data. Results from MVLR and SES one step prediction show that as the moving window increases, the prediction accuracy increases as well using the percentage error as a validation measure, similarly in the two-step prediction mode. Holt's method and Holt Winter's results are similar in most moving window values. For one step prediction mode, the best prediction accuracy for ADCB was via Holt Winter's method with the percentage error value of 0.2358%, for Al Dar Properties was SES with the percentage error value of 0.0587% and for Dana Gas Co. was SES with the percentage error value of 0.1023%. For the two-step prediction mode, the lowest mean squared error (MSE) for ADCB was by means of Holt Winter's method, for Al Dar Properties was by SES and for Dana Gas Co. was by Holt Winter's method. As for the three-step prediction mode, the algorithm with the lowest MSE for ADCB was Holt's method with the MSE value of 0.0046, MVLR with the MSE value of 0.0000655 for Al Dar Properties and Holt Winter's method with the MSE value of 0.003612. A detailed analysis of the different methods was discussed within the study. To conclude the thesis, limitations and future work were considered to further understand the scope of the research as well as contemplate prospective improvements.
Date of AwardMar 2020
Original languageAmerican English

Keywords

  • Financial
  • Stock
  • Prediction
  • Machine
  • Learning.

Cite this

'