We evaluate the effectiveness of 63 different features commonly used in the classification of actigraphy signals. We implement two feature selection techniques to rank the effectiveness of the features and select the “best” among them. Once the “best” feature(s) is (are) selected, a minimum distance classifier is used to classify the actigraphy signals into different types of activity. The minimum distance classifier uses class prototypes generated using either k-means or max-min clustering algorithm. Correct classification rates of 95%–100% were achieved using only 1–5 features.