A frequency count based two stage classification approach is proposed by combining generative and discriminative modeling principles for online handwritten character recognition. The first stage classifier based on Hidden Markov Model (HMM) returns top-K ranking characters out of the total N classes. In the second stage, pairwise classifiers for K(K - 1)/2 unique combinations of top-K characters using Support Vector Machine (SVM) are developed. Usually pairwise classifiers are trained for most confused character pairs, by analyzing the confusion matrix. Alternatively, in this work, the pairwise classifiers are trained by analyzing the frequency count of top-K labels. The frequency count helps the pairwise classifiers trained not only for confused, but also for similar shape characters. The proposed frequency count based two stage approach achieves 95.1% recognition accuracy as opposed to 94.4% using confusion matrix information, and 93.8% using single stage HMM based approach on 113 classes dataset. © 2016 IEEE.