A　new　prediction　algorithm　based　on　Empirical　model　decomposition　(EMD)　and　Support　vector　machine　(SVM)　is　put　forward　in　this　paper,　and　this　algorithm　solves　the　problem　of　the　hydrogen　atomic　clock　differences　prediction,　which　is　affected　by　the　non-linearity　and　non-stability.　The　clock　differences　were　decomposed　into　Intrinsic　mode　functions　(IMF)　and　the　residual　series.　The　suitable　kernel　function　and　parameters　were　chosen　to　build　the　different　SVM　for　predicting　each　IMF　and　the　residual　series.　Each　prediction　result　was　summed　to　obtain　the　clock　differences　prediction.　Results　show　that　the　EMD-SVM　algorithm　is　effective　compared　with　the　linear　regression　and　single　SVM.　The　relative　prediction　error　is　reduced　from　0.4327%　to　0.2371%,　and　the　dispersion　is　less　than　other　methods.