In order to provide basic communication abilities to people with motor disability, motor imagery brain computer interface is one of most widely used technique. In this paper, we present a novel algorithm (Composite Filter bank based stationary CSP) for determining subject as well as task specific discriminative frequency bands for classification of motor imagery tasks. It is noted in the literature that while performing any motor imagery tasks, two major frequency band of EEG spectrum i.e mu (7-12 Hz) as well as beta (12-30 Hz) bands are actively involved. Hence, in most of the literature work EEG signals were filtered using a frequency band of 7-30 Hz usually before using CSP transformation. However, it is possible that some of the frequencies may not provide useful features to distinguish motor imagery tasks. In this paper, we propose a novel approach to select a subset of relevant frequency bands using sequential forward feature selection method from a composite filter bank which consists of Prior-known EEG frequency bands and a set of variable size overlapping frequency bands to improve the performance of motor imagery tasks classification. Experimental results of the proposed work on publicly available datasets validate the effectiveness of the proposed method. Friedman statistical test conducted further shows that the proposed approach significantly outperforms the existing methods. © 2018 IEEE.