The purpose of this study is to examines comparative functioning of linear versus nonlinear methods to forecasts stocks returns volatility by using daily data that extents from period January 2000 to June 2016.A range of naïve models to the comparatively difficult uncertain heteroskedastic models of the GARCH family in an out of sample forecasting for daily stock returns volatility. Return series of KSE-100 index is estimated by comparing linear versus nonlinear. Root Mean Squared Error (RMSE) criterion is used to check which model ranked first by getting minimum RMSE. The empirical result shows AR(1) model ranked at first in linear models dominating all other linear models while in nonlinear GARCHM(1,1) ranked first dominating Exponential-GARCH(1,1) and Power-ARCH(1,1). After comparing both linear and non-linear model we came to conclusion that GARCH-M (1, 1) proved as a best forecasting model of our study in both high and low node(50 &120) of out of sample forecasting used in stock prices volatility. Supervisor:- Dr. Abdul Rashid