A Comparison of Deep Learning Models in Time Series Forecasting of Web Traffic Data From Kaggle

Abstract

In recent years, time series forecasting has attracted more attention from academia and industry. This research used raw data from the “Web Traffic Forecasting” competition on the Kaggle platform to test the prediction accuracy of different time series models, especially the generalization performance of various deep learning models. The experiments used historical traffic data from 145,063 web pages from Wikipedia from 2015-07-01 to 2017-11-13. Traffic data from 2015-07-01 to 2017-09-10 was used to forecast traffic from 2017-09-13 to 2017-11-13, a total of 62 days. The experimental results showed that almost all deep learning models predicted far more effectively than statistical and machine learning models, showing that deep learning models have great potential for time series forecasting problems.