Two complementary bounds of the predicted pixel are defined and used to compute two estimates of the prediction error. The more suitable value for reversible watermarking among the two estimates is selected for data insertion. A reversible watermarking scheme based on context embedding ensures the detection of the selected value, without the need for any additional information. The scheme is general and works regardless the particular predictor. The proposed scheme is of interest for embedding bit-rates of less than 0.5 bpp. Interesting results are reported for the case of pairwise embedding reversible watermarking. The proposed scheme compares very well with the most efficient schemes published so far.