Modeling and Forecasting the Dynamics in Romanian Stock Market Indices Using Threshold Models

We investigate the existence of nonlinear patterns in the dynamics of the main stock index returns in Romania. We use daily closing data of the BET stock index series from 2004 to early 2010. Based on several tests for nonlinearity we reject the null hypothesis of linearity. We use several types of threshold models and compare their fitness and forecasting performance with basic AR models. We found that the LSTAR and SETAR models fit best the data; however, they cannot outperform the simpler AR models in forecasting. These results suggest that although there are nonlinear features in data, the threshold models are not complex enough to reveal the data complexity.