All social groups in community find traffic congestion to be a major cause for concern, and it either directly or informally affects the country's economy. Hence, traffic modelers are concerned with providing precise forecasts of traffic flow. This study proposes a novel multivariate-tree structured Machine learning hybrid modeling framework for improved traffic flow prediction. In this method, firstly Multivariate Empirical model decomposition (MEMD) is utilized for the breakdown of datasets into a collection of Intrinsic mode functions and final residue. The entire set of modes is grouped into three components, by retaining the first mode (C1), residue (R) and aggregated component (AC) by aggregating all the intermediate IMFs. Each component is separately modeled using three tree-based methods-M5P Model Tree (M5P), Random Forest (RF), and Reduced Error Pruning tree (REPTree). The final prediction results are recombined to get the modeled traffic flow. Performances of predictions are evaluated by a set of statistical and graphical measures after developing prediction models using the three tree-based algorithms using raw data inputs. All the hybrid models are performing better than their standalone counterparts in terms of both correlation and error measures, while the MEMD-REPTree framework is found to be performing best among the six methods considered in this study. The proposed MEMD-based framework reduces the effects of noise in traffic data and increases the accuracy of predictions while accounting for the use of multiple inputs while overcoming the challenges offered by traditional decomposition strategies in terms of number of modes, accuracy and computational expenses.