Among the elderly, falls are a well-known safety hazard, often resulting in major injury, hospitalization and death. To reduce the injuries caused by falls, it is first necessary to predict a fall as early as possible and then to provide protection for the person who is falling. This paper proposes a fall-prediction algorithm (FPA) that can predict whether the person will fall within one-walking-step. The fall prediction is different from the fall detection, and it is intended to predict a fall before it occurs and provide sufficient time to enable a safety mechanism. The proposed FPA adopts a neural network to perform prediction in which the inputs are accelerations and angular rates of upper trunk and the output presents fall or no fall. A wearable inertial sensor package with a triple axis accelerometer and a triple axis gyroscope is developed to measure the required motion data. Five subjects were asked to wear the inertial sensor package and perform a number of simulated falls. The experimental results show that the FPA could predict a fall 0.4 seconds prior to the beginning of the fall. The time interval is sufficient to inflate an airbag covering the head, trunk, and hip, an intervention that would reduce fall-related injuries among older people.