Exchange　rate　is　considered　as　a　highly　nonlinear　and　non-stationary　time　series　which　can　hardly　be　properly　modeled　and　accurately　predicted　by　traditional　linear　econometric　models.　This　study　attempts　to　propose　an　exchange　rate　ensemble　learning　paradigm　called　EMD-SVR.　This　methodology　decomposes　the　original　non-stationary　and　irregular　exchange　rate　series　into　a　finite　and　often　small　number　of　sub-signals　by　empirical　mode　decomposition　(EMD).　Then　each　sub-signal　is　modeled　and　forecasted　by　a　Support　Vector　Regression　(SVR).　Finally　the　forecast　of　exchange　rate　is　obtained　by　aggregating　all　prediction　results　of　sub-signals.　We　verify　the　effectiveness　and　predictability　of　EMD-SVR　using　EUR/RMB　time　series　as　sample.　The　result　shows　that　EMD-SVR　has　a　strong　forecasting　ability　and　is　remarkably　superior　to　normal　SVR.　©　2010　IEEE.