The forecast performances of the classical time series model and machine learning algorithms on BIST-50 price index using exogenous variables

Date

2023-8

Author

Parlak, Fatma

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

129
views

97
downloads

Cite This

Time series analysis aids in the prediction of future values using a collection of data over a period of time such as hours, days, months and years. Time series analysis gives researchers an opportunity to make inferences from data and it sheds light on the internal structures of the time dependent dataset. The main purpose of this thesis is to forecast the BIST-50 Price Index by using several macroeconomic and financial indicators. Autoregressive Integrated Moving Average (ARIMA), Prophet, Random Forest, Support Vector Machines, XGBoost and Bayesian Regularized Neural Network algorithms are applied for predicting the dataset. Moreover, stacking method which is an ensemble model technique is constructed in order to improve the performance of the models. Furthermore, a comparison of the forecast performance of all models is made. The Mean Absolute Percentage Error, Root Mean Square Error and Mean Absolute Error are calculated for measuring the forecast accuracy. In addition to comparing the model performances, SHAP analysis is also implemented to determine which exogenous variables are related with the response variable negatively or positively. Consumer Price Index, Official Reserve Assets and MSCI Index are concluded as the most significant variables in the XGBoost Model. The most important variables in the Random Forest model are Consumer Price Index, M2 Money Supply and USD/TRY Exchange Rate. Consumer Price Index, the 4th lag of BIST50 dataset and MSCI Index are found as the most significant variables in the Support Vector Machine model. The findings present that the best models are ARIMA, Prophet and ensemble model composed of Random Forest method created by using Random Forest, Support Vector Machine and Bayesian Regularized Neural Network forecasts.

Subject Keywords

BIST-50, Time series analysis, Forecasting stock prices, Machine learning, Ensemble model

URI

https://hdl.handle.net/11511/105148

Collections

Graduate School of Natural and Applied Sciences, Thesis