Accurate　forecasts　of　the　stock　market　have　important　implications　for　both　investors　and　regulators.　The　decomposition-integration　framework　is　widely　used　in　forecasting　research　of　financial　time　series.　However,　most　of　previous　studies　only　use　single　historical　data　to　predict　the　components,　which　ignoring　the　influence　of　other　low-frequency　heterogeneous　data　on　the　components.　This　study　proposes　a　novel　decomposition-integration　model　EEMD-Mixed　Frequency　CNN-BiLSTM-Attention/LSTM-LSTM　(EE-MFCBA/L-L),　which　take　the　advantages　of　the　decomposition-integration　and　mixing　data　sampling.　Firstlly　the　stock　index　is　decomposed　into　several　different　frequency　components　by　EEMD.　Meanwhile　the　fuzzy　entropy　algorithm　is　used　to　identify　the　frequency　characteristic　of　the　components.　Then　the　components　are　predicted　by　MFCBA/L　model　where　the　low　frequency　data　will　be　considered　according　to　the　frequency　characteristics　of　the　components.　Finally,　the　LSTM　model　is　used　to　nonlinearly　integrate　the　predictor　of　each　component.　The　empirical　results　show　that　the　proposed　model　can　better　adapt　to　the　characteristics　of　returns.　Compared　with　the　traditional　model,　the　proposed　model　has　significant　advantages　in　predicting　non-stationary　and　nonlinear　return　series,　with　the　lowest　prediction　error　and　the　highest　directional　prediction　accuracy.　©　2022　Systems　Engineering　Society　of　China.　All　rights　reserved.