An orientation predicting method, adapted to a virtual reality headset, comprises obtaining an orientation training data and an adjusted orientation data, wherein the adjusted orientation data is obtained by cutting a data segment off from the orientation training data, wherein the data segment corresponds to a time interval determined by an application latency; training an initial neural network model based on the orientation training data and the adjusted orientation data corresponding to the time interval; retrieving a real-time orientation data by an orientation sensor of the virtual reality headset; and inputting the real-time orientation data to the trained neural network model to output a predicted orientation data. The present disclosure further discloses a virtual reality headset and a non-transitory computer-readable medium.