Forecasting stock prices using hidden Markov models and support vector regression with firefly algorithm

Many models have been proposed for forecasting stock prices. One is the Autoregressive Integrated Moving Average (ARIMA) model which is extensively used in the fields of economics and finance (Ariyo et al., 2014). In the Philippines, the Philippine Stock Exchange (PSE) forecasts future stock price movement by applying the ARIMA model on vast amounts of historical data (Trading Economics, n. d.). Although generally used, is ARIMA really the best model for the job? In this paper, two new forecasting techniques are introduced: Hidden Markov Models (HMM) and Support Vector Regression with Firefly Algorithm (SVR FA). Both methods are compared to ARIMA in analyzing closing stock prices of five selected Philippine companies: SM, Ayala Corporations (AC), Philippine Long Distance Telephone Company (TEL), JG Summit (JGS), and Meralco (MER). All of these five companies present closing stock price movements which cannot be analyzed easily. Thus, such companies will present a challenge to the proposed models as well as for ARIMA. HMM is a tool for time series data modelling that has a solid statistical foundation (Nath & Hassan, 2005). The HMMs are trained using observable states that can emit possible movement of hidden states. In stock price forecasting, we may assume an underlying hidden movement that governs the actual increases (decreases) in stock prices for model estimation. The SVR FA model uses the ε-sensitive loss function and a kernel function for analyzing the best regression hyperplane that relates the current day\u27s closing stock prices with the future closing stock prices. Forecasting accuracies of the two proposed models are compared to ARIMA using the mean absolute percentage error (MAPE) and mean absolute deviation (MAD). Results show that the use of SVR-FA and HMM performed better in forecasting closing stock prices of the selected companies compared to ARIMA with SVR-FA having the best forecasts, yielding the lowest MAPEs and MADs