Predicting the next symbol of a sequence of symbols is a task with wide applications. The Compact Prediction Tree (CPT) is a recently proposed prediction model that provide more accurate predictions than several state-of-the-art prediction models. In this paper, we introduce new strategies to reduce the size of CPT and its prediction time. Experimental results on seven datasets shows that the resulting model is up to 98% more compact than CPT and 4.5 times faster, and remains on overall much more accurate than state of the art predictions models All-K-order Markov, DG, Lz78, PPM and TDAG.