Modeling Swedish Stock Data Using Regime Switching ARCH-Processes

Erik Henricsson

Handledare: Jan Holst

Centre for Mathematical Sciences
Mathematical Statistics
Lund Institute of Technology,
Lund University,

Conventional time-series models assume a single structure for the conditional mean and variance throughout the entire study period. Such assumptions cause problems that make regime-switching models more appropriate, since they allow the model parameters to change. The objective of this paper is to evaluate whether regime switching ARCH models can produce a better description of stock volatility than the ordinary single regime models. A regime switching ARCH model with only one switching parameter was compared with the ordinary single regime ARCH and GARCH models. All models were estimated and evaluated on daily data that come from the Swedish stock index Affärsvärldens generalindex. The model evaluation was based on loss functions for the one-step predictions and hypothesis tests for the normalized residuals, as well as graphical interpretation of different plots. After having evaluated and compared all models, Student-t distributed GARCH(1,1)-L seemed to be the best model, in particular in terms of loss and likelihood functions. Generally, Student-t distributed models appeared to yield better-looking residuals than the models that were driven by Gaussian noise. Interestingly, switching ARCH imposed less persistence than the ordinary ARCH, which could be taken as evidence that regime-switching models can take care of outliers and instationarities.