Comparative study on the time series forecasting of web traffic based on statistical model and Generative Adversarial model

We evaluated the accuracy of several classical statistical methods of Time series forecasting with ground truth dataset which was obtained from Kaggle web traffic forecasting competition hosted by Google. A novel way of seasonal, trend and cycle pattern decomposing method was used for the specific time series daily data. We proposed using the combination of four traditional methods to reduce the RMSE and thus achieved better forecasting accuracy. Results showed error rate was lowered down 10 to 20 percentage points. After studying the characteristics of the web traffic time series, we presented the Generative Adversarial Model (GAN) with Long-Short Term Memory (LSTM) as generator and deep Multi-Layer Perceptron (MLP) as discriminator to forecast the web traffic time series. The forecasting performances was compared among the traditional statistical methods and the deep generative adversarial network. We concluded from experiments there was no remarkable difference for this specific times series forecasting accuracy using these two kinds of methods. (C) 2020 Published by Elsevier B.V.

Automated summary

The study evaluated the accuracy of several classical statistical methods in time series forecasting and proposed a novel decomposition method to reduce RMSE and improve forecasting accuracy. Results showed a decrease in error rate and compared traditional statistical methods with a deep generative adversarial network, finding no significant difference in forecasting accuracy for this specific time series.