Forecasting stock returns is considered one of the hardest tasks for every potential investor. This paper attempts to predict the movement of Croatian stock market index Crobex on Zagreb Stock Exchange. Main aim of this paper was to empirically examine the best univariate Autoregressive Integrated Moving Average model for forecasting. This research examined ARIMA (p;d;q) model on weekly closed prices of Crobex from 01/01/2011 to 01/01/2013. First it was necessary to meet the stationary condition. While checking the conditions of stationarity, data series were observed by ACF, PACF plots and by Ljung- Box Q statistic and Augmented Dickey-Fuller test statistic. After differencing, statistic showed that the data is stationary and the next step was to find the best ARIMA model. The most important criteria that were used are: R-squared, Adjusted R-squared, Akaike information criterion, Schwarz criterion and Hannan-Quinn information criterion. After checking the exceptionally large number of models it was found the model that suits best, according to the criteria.