Abstract: Analyzing the hidden relationship between historical tax data and using mathematical models to predict future tax revenue is the focus of tax forecast research.A tax prediction model of long short-term memory (LSTM) recurrent neural network combined with wavelet transform is proposed in this paper.Combining wavelet transform on data preprocessing to remove noise from tax data and improve the generalization ability of the model.The LSTM neural network can better learn the correlation between historical tax data by adding hidden neural units and gated units,and further extract valid state innovations between input sequences,and overcome the long-term dependency problem of recurrent neural networks.Experimental results show that the encoder-decoder structure based on the LSTM neural network can enhance the time step of tax prediction.Compared with the single-step sliding window LSTM neural network model and the gray model based on difference differential equations in the long-term tax prediction,the model and the regression-based autoregressive moving average model (ARIMA) significantly improve the prediction accuracy.

Key words: Encoder-decoder, Long-short term memory network, Tax forecasting, Wavelet transform