GARCH accounts for stochastic volatility in a time series of returns but the returns may have components other than that can be explained by stochastic vol, such as trends or moving average. Thus they may be inappropriate when an asymmetric effect is observed as a different instability. You can have more information through the link below.

https://mpra.ub.uni-muenchen.de/15143/1/MPRA_paper_15143.pdf

Volatility by their very nature is unpredictable. Perhaps a more manageable task is scenario building rather forecasting the path of the volatility.

The garch family of models are variance models. The capture variation in the variance of the time series. That is why they are suited for modelling series that exhibit volatility. Smoother series may not be modelled by the garch approach but by the Arima approach.

Dear Bhavana

Here are some readings that address your question:

http://www.science.unitn.it/AnalisiInfoTLC/SSP/SSP14_15/WEBSITE/courses/Ssp1415_2014/Slides_Sources/27_20_2014-12-15/lecture20Prev.pdf

http://www.ipe.ro/rjef/rjef4_09/rjef4_09_3.pdf

https://mpra.ub.uni-muenchen.de/88309/1/MPRA_paper_88309.pdf

http://www.actuaries.org/ASTIN/Colloquia/Berlin/Tagliafichi.pdf

Thanks everyone for your valuable inputs.

i can assume two limitations in application of GARCH models.

one is underlying probability distribution, it can be a limitation of any regression model including garch.

two, there are plenty of garch models. selecting the appropriate model can also create problems

volatility

GARCH models assume deterministic volatility based on past returns and conditional variances. In reality it is not true. Besides they are mainly based on closing prices and do not include prices during the day. Read about the range-based volatility models which are better and do not require additional intraday data.