Dim light melatonin onset, or DLMO, is the gold standard for circadian phase assessment in humans, but it is limited in being time and resource intensive. Numerous studies have attempted to predict circadian phase from actigraphy data with some success in more controlled populations, where mean errors are reported between .5 and 1 hours. We find that such algorithms are less successful in predicting DLMO in a population of college students with more irregular schedules, where mean errors in predicting DLMO directly are approximately 1.5 hours. By reframing the problem of predicting DLMO as a classification problem, where we use a neural network to predict whether a timepoint falls before or after DLMO, we find high classification accuracy of about 90%, which decreases the mean error to approximately 1.2 hours. To test the generalizability of this approach, we apply the same neural network to data from forced desynchrony studies. In the participants on the forced desynchrony protocol, classification accuracy drops to 55-65%. We find that this accuracy is highly dependent upon the phase angle between DLMO and sleep onset with highest accuracy at normal phase angles: ranging between 20 and 80% depending upon the phase angle. This suggests that the method is valid when circadian phase and sleep-wake with its associated behaviors are appropriately aligned. More work should be done in cases of circadian misalignment with rest-activity rhythms.