In　international　trade,　it　is　common　practice　for　multinational　companies　to　use　financial　market　instruments,　such　as　financial　derivatives　and　foreign　currency　debt,　to　hedge　exchange　rate　risks.　Making　accurate　predictions　and　decisions　on　the　direction　and　magnitude　of　exchange　rate　movements　is　a　more　direct　way　to　reduce　exchange　rate　risks.　However,　the　traditional　time　series　model　has　many　limitations　in　forecasting　exchange　rate,　which　is　nonlinear　and　nonstationary.　In　this　paper,　we　propose　a　new　hybrid　model　of　complete　ensemble　empirical　mode　decomposition　(CEEMDAN)　based　multilayer　long　short-term　memory　(MLSTM)　networks.　It　overcomes　the　shortcomings　of　the　classic　methods.　CEEMDAN　not　only　solves　the　mode　mixing　problem　of　empirical　mode　decomposition　(EMD),　but　also　solves　the　residue　noise　problem　which　is　included　in　the　reconstructed　data　of　ensemble　empirical　mode　decomposition　(EEMD)　with　less　computation　cost.　MLSTM　can　learning　more　complex　dependences　from　exchange　rate　data　than　the　classic　model　of　time　series.　A　lot　of　experiments　have　been　conducted　to　measure　the　performance　of　the　proposed　approach　among　the　exchange　rates　of　British　pound,　the　Australian　dollar,　and　the　US　dollar.　In　order　to　get　an　objective　evaluation,　we　compared　the　proposed　method　with　several　standard　approaches　or　other　hybrid　models.　The　experimental　results　show　that　the　CEEMDAN-based　MLSTM　(CEEMDAN-MLSTM)　goes　on　better　than　some　state-of-the-art　models　in　terms　of　several　evaluations.