This　paper　integrates　the　maximal　overlap　discrete　wavelet　transform　(MODWT),　long　and　short-term　memory　neural　network　(EA-LSTM)　of　evolutionary　attention　mechanism　and　reference　vector　based　clustering　algorithm　(RVMOC)　and　proposes　a　new　prediction　method　of　the　stock　market　return　rate,　which　is　referred　to　as　the　stock　market　return　rate　prediction　method　based　on　MODWT-EA-LSTM-LSTM-RVMOC.　This　method　uses　a　clustering　strategy　based　on　a　reference　vector　to　extend　decomposition-integrated　learning　to　nonlinear　integrated　weighted　learning　based　on　local　data　feature　weighting,　overcomes　the　deficiency　of　the　integrated　learning　stage　in　the　decomposition-integration　method,　and　effectively　solves　the　problem　of　artificial　experience　setting　of　the　objective　function　weight　coefficient　and　clustering　accuracy　in　existing　cluster-integrated　learning.　The　empirical　results　show　that　compared　with　the　single　model　and　decomposition-integration　learning　model,　the　MODWT-EA-LSTM-RVMOC　algorithm　is　better　than　other　models　in　both　prediction　error　and　prediction　hit　rate.　The　results　also　indicate　that　the　RVMOC　clustering　algorithm　can　effectively　improve　the　prediction　performance　of　the　decomposition-integration　model.