An earthquake process includes pre-seismic stress accumulation, co-seismic rock rupture and post-seismic elastic and/or viscoelastic rebound. Although co-seismic and post-seismic deformations have been readily observed using the global positioning system (GPS), detecting pre-seismic stress accumulation hidden in time-series data remains challenging. This study applies the Hilbert-Huang transform to extract non-linear and non-stationary pre-earthquake deformation data from GPS records for central Taiwan. By converting the derived surface deformation into horizontal azimuths, the randomly oriented GPS-azimuths are reoriented in a similar direction several days before and after earthquakes due to loading and rebound stress, respectively. Analytical results demonstrate that the stress accumulation and release along the entire course of an earthquake process provide significant evidence supporting the seismic rebound theory. This finding would be applicable to areas with dense GPS networks and active plate interactions. Surface deformations detected by the proposed analytical technique have encouraging potential for mitigating future seismic hazards.