Obvious, the better this prediction, the smaller the residue. The “compression” level of FLAC is the amount of resources FLAC is allowed to spend on finding the best possible linear prediction. Store the coefficients + residue in a file and the decoder can reconstruct the original signal. The bad news is that there is nothing linear about audio. They can be stored using very little space. The advantage is that the prediction can be written as a couple of coefficients.
