Timing recovery in partial-response-based magnetic recording systems customarily employs the “decision-directed” method wherein phase error is recovered from the differences between the noise-corrupted received signal samples and their estimated ideal (noise and phase error free) values. The filtered phase error drives a numerically-controlled oscillator which determines the instants at which the signal is resampled, attempting to place said instants at the ideal sampling times. The resampled signal contains errors due to mistiming as well as to the original corrupting noise, and these errors directly influence the success of subsequent detection. However, the noise can be reduced using adaptive linear prediction, having the effect of reducing the output error for a given noise input, or maintaining the same error for a larger noise input.