Sarima-arch versus genetic programming in stock price prediction

In financial time series, one of the most challenging problems is predicting stock prices since the data generally exhibit deviation from the assumptions of stationary and homoscedasticity. For homogenous non-stationary time series, the Autoregressive Integrated Moving Average (ARIMA) model is the most commonly used linear class including some transformation such as differencing and variance stabilizing process. However, stockmarket data is often nonlinear, which indicates that more advanced methods are necessary. Genetic Programming (GP) is one of the evolutionary computational methods that could capture both linear and nonlinear patterns in time series data. The present study aims to build a machine learning tool using GP for prediction The Istanbul Stock Exchange National 100 (XU100) index and compare the obtained results with conventional seasonal ARIMA(SARIMA) and ARCH models. In order to achieve this goal, it was first modeled with the SARIMA model after appropriate transformations were made to the stock price series and the diagnostic control result showed that the residual of the SARIMA model have the heteroscedasticity problem. Then, the ARCH model was applied to SARIMA residuals to eliminate this effect and an integrated SARIMA-ARCH model is obtained. Since it is possible and capable to model nonlinear and non-stationary time series using GP without any pre-assumptions, we proposed GP to predict the stock price series. The function set of GP consists of not only arithmetic but also trigonometric functions. To the best of our knowledge, this study is the first to predict XU100 stock price data using GP. In this experiment, the data set consists of the daily closing prices of the XU100 index over 775 days from the beginning of 2017 until the end of January 2020. The experimental results obtained show that the accuracy metrics used in the study are lower in the proposed GP model compared to other models. These results reveal that the GP method provides better predictive results for the financial time series data of the XU100 index than traditional methods.