Stroke survivors usually experience paralysis in one half of the body, i.e., hemiparesis, and the upper limbs are severely affected. Continuous monitoring of hemiparesis progression hours after the stroke attack involves manual observation of upper limb movements by medical experts in the hospital. Hence it is resource and time intensive, in addition to being prone to human errors and inter-rater variability. Wearable devices have found significance in automated continuous monitoring of neurological disorders like stroke. In this paper, we use accelerometer signals acquired using wrist-worn devices to analyze upper limb movements and identify hemiparesis in acute stroke patients, while they perform a set of proposed spontaneous and instructed movements. We propose novel measures of time (and frequency) domain coherence between accelerometer data from two arms at different lags (and frequency bands). These measures correlate well with the clinical gold standard of measurement of hemiparetic severity in stroke, the National Institutes of Health Stroke Scale (NIHSS). The study, undertaken on 32 acute stroke patients with varying levels of hemiparesis and 15 healthy controls, validates the use of short length (< 10 minutes) accelerometry data to identify hemiparesis through leave-one-subject-out cross-validation based hierarchical discriminant analysis. The results indicate that the proposed approach can distinguish between controls, moderate and severe hemiparesis with an average accuracy of 91%.