This study developed and evaluated machine learning algorithms to predict children's physical activity category from raw accelerometer data collected at the hip. Fifty participants (mean age = 13.9 ± 3.0 y) completed 12 activity trials that were categorized into 5 categories: sedentary (SED), light household activities and games (LHHAG), moderate-vigorous games and sports (MVGS), walking (WALK), and running (RUN). Random Forest (RF) and Logistic Regression (LR) classifiers were trained with features extracted from the vector magnitude using 10 s non-overlapping windows. Classification accuracy was evaluated using leave-one-subject-out cross validation. Overall accuracy for the RF and LR classifiers was 95.7% and 94.3%, respectively. Classification accuracy was excellent for SED (96.3%-98.1%), LHHAG (92.3%-95.2%), WALK (94.5%-97.1%), RUN (99.5%-99.6%); and MVGS (87.5%-92.7%). The results indicate that classifiers trained on features in the raw acceleration from the hip can be used for activity recognition in young people.