A variation of recurrent neural networks (RNN) called the GRU neural network is suggested in order to address the long-term reliance of RNN network structure. GRU overcomes the long-term reliance of time series on the basis of inheriting RNN’s strong memory ability of time series. This research proposes a financial time series prediction model based on difference operation and GRU neural network with a focus on the reliance of financial time series data. It also applies GRU extension to financial time series prediction. The model can handle the complicated nonlinearity, nonstationarity, and series correlation properties of financial time series data. The experimental results show that the suggested scheme can improve the GRU neural network’s generalization capability and prediction accuracy. Additionally, when compared to the traditional prediction model, this model has a better prediction effect and a relatively lower calculation cost for financial time series. The experiment makes a prediction about the S&P 500 stock index’s modified closing price.