Elman neural networks can predict stock market closing prices in the short term with good trend but low accuracy. Using the Elman neural network, we propose a new Elman combination model based on empirical mode decomposition (EMD). Firstly, the EMD is applied to forecast the closing price whose sequence is decomposed into intrinsic mode function（IMF）component and the residual component on different time scales. Secondly, we use the partial autocorrelation function (PACF) to calculate the lag time of each component, and further determine the input and output variables of each component in the Elman neural network. Thirdly, the prediction value of each component is obtained and summed up to achieve the final prediction results. Experimental results show that compared with the single Elman network, EMD-Elman model, BP neural network and EMD-BP model, the short-term prediction model can achieve prediction results with better MSE, MAE, and MAPE. The new combination model can effectively achieve the short-term prediction of the closing price of SSE and SZSE and reduce the effect of stock's closing price volatility on prediction results. This study provides an effective basis for further prediction of stock market trends, and also provides a better reference for investors to make decisions.