EFFECTIVE STOCK PRICE FORECASTING USING MACHINE LEARNING TECHNIQUES WHILST ACCOUNTING FOR THE STATE OF THE MARKET

Abstract

Machine learning methods have been successfully applied to stock price forecasting. Although finance practitioners and academics have advocated for the benefits of using fundamental and technical analyses together, the machine learning research has been focused on using the technical analysis based indicators almost exclusively. The main target for prediction by machine learning researchers have been forecasting of next day’s price for a market index or a firm’s stock. Another challenge presented in stock price forecasting is the impact of the overall stock market volatility on the individual stock prices. The aim of this thesis was to investigate into the impact on machine learning-based stock price forecasting by using various inputs (technical, fundamental, and combined) and also by accounting for the states of stock market. A framework is proposed which enables the selection of the best performing model with relevant inputs and which can also factor insensitivity of the stock’s price to various states of the market. The initial simulations were run for 147 companies with 252 days out stock price forecasting, and further simulations were undertaken for 85 companies with 126 days out stock price forecasting. We show the importance and relevance of using the fundamental indicators and combination of the technical and fundamental indicators when forecasting financial time-series into the horizons of 126 and 252 days. The proposed approach for integrating the moods exhibited by the stock market is embedded into the forecasting process. The explicit identification and inclusion of the market states were more effective for 126 days than for 252 days, but also when the combined indicator set was not being used as the input. Another contribution of the thesis was the framework that provided an improved structured approach for conducting financial time series forecasting (RMSE of 0.0614 vs. 0.3175 of the Random Walk model).

Date of Award	31 Dec 2019
Original language	English
Awarding Institution	The University of Manchester
Supervisor	John Keane (Supervisor)