In this work, we propose a feature exploration method for learning-based cuffless blood pressure measurement. More specifically, to efficiently explore a large feature space from the photoplethysmography signal, we have applied several analytical techniques, including random error elimination, adaptive outlier removal, maximum information coefficient and Pearson's correlation coefficient based feature assessment methods. We evaluate fifty-seven possible feature candidates and propose three separate feature sets with each containing eleven features to predict the systolic blood pressure (SBP), diastolic blood pressure (DBP) and mean blood pressure (MBP), respectively. From our experimental results on a realistic dataset, this work achieves 4.77±7.68, 3.67±5.69 and 3.85±5.87 mmHg prediction accuracy for SBP, DBP and MBP. In summary, using the proposed light-weight features, the proposed predictors can successfully achieve a Grade A in two standards proposed by the American National Standards of the Association for the Advancement of Medical Instrumentation (AAMI) and British Hypertension Society (BHS).