Abstract: The fitting effect of neural network model on large sample time series is often better than that of traditional time series models, but it may not be advantageous for the prediction of low-frequency time series such as annual, monthly and daily data. In view of this, we propose HoltwintersBP combined model to forecast the advertisement income and expand the multi-dimensional explanatory variables, using daily advertisement income data from a social news type app. By comparing the prediction results with those of RNN and LSTM, it is found that Holtwinters-BP combined model has higher prediction accuracy and stability; explanatory variables have a significant impact on advertisement income, the prediction accuracy of the model with explanatory variables is higher than that of single variable model; it also verifies the validity and applicability of Holtwinters-BP combined model for low-frequency data prediction.

Key words: Advertisement Income Forecasting, Neural Network Model, Holtwinters Model, Combined Model