Most digital cameras perform color demosaicing and compression sequentially to yield a color output. Recent reports indicate that the alternative compression-then-demosaicing approach outperforms the demosaicing-then-compression approach in terms of image quality and complexity. This paper presents a fast reversible Bayer image compression algorithm for the alternative approach. A statistic-based prediction is proposed to de-correlate the wavelet subband coefficients. By learning from experiences, the proposed predictor can improve its prediction performance adaptively. A context-based Golomb Rice code is then proposed to compress the subband residues. Simulation results show that, as compared with the existing lossless CFA image coding methods, the proposed algorithm can achieve a low bitrate with lesser computation.