An adaptive reversible data hiding method through autoregression is presented in this paper. In the proposed algorithm, we focus on the image pixel value prediction, which plays a key role in the data embedding process. Unlike conventional data hiding techniques, a threshold is adjusted for each image to divide all pixels into two regions: the smooth region and the texture region. Then the proposed algorithm optimally estimates the coefficients of autoregression model for pixel value prediction through least-squares minimization. The prediction error is adaptively minimized to achieve high prediction accuracy so that more redundancy in the image is exploited to achieve very high data embedding capacity while keeping the distortion low. Experimental results show that the proposed algorithm outperforms typical state-of-the-art methods in general.