This study attempts finding the best forecasting model for multivariate time series in Pakistan by using the variable selection models like, Regression Tree, Bagging, Random Forest, Boosting, and Adaptive LASSO. The monthly time series data on 31 Macroeconomic variables is collected from State Bank of Pakistan and International Financial Statistics for the period 1990-2017. The performance of each technique is gauged on the basis of forecast error. Further, performance is checked by varying the number of predictors in the model i-e, information set ranging from four, fourteen and thirty one. In this study we evaluate the empirical performance of the competition forecasting method with nested subsets of predictor with different value of k, specifically I4 ⊂ I14 ⊂ I31. This allows us to investigate the impact of utilizing information set of different sizes. In this study we have found that when we increase the numbers of predictors form four to fourteen the predications found are better from four variables and if the number of variables increased from fourteen to thirty one the results showed prediction improved even further. If there are more than 20-31 variables then there is no improvement in the result. Supervisor:- Dr. Amena Urooj Co-Supervisor:- Ms. Uzma Zia