Underfitting is a problem, when we have low error in both training set and the testing set. Few algorithms works better for interpretations. But fails for better predictions.