Web Traffic Prediction Using Autoregressive, LSTM, and XGBoost Time Series Models

Abstract

Web traffic is vital to the success of any online company or website in the current era of digital technology. Insightful marketing, web development, and resource allocation choices may be made with the support of reliable online traffic forecasts. In this study, we investigate the effectiveness of the Autoregressive (AR), Long Short-Term Memory (LSTM), and eXtreme Gradient Boosting (XGBoost) time series modeling strategies for forecasting website traffic. We evaluate the accuracy of these models in forecasting future online traffic by comparing their results on a real-world dataset. The performance of four different models for predicting a target variable was evaluated based on the provided information. The AR model had the highest test error, indicating poor performance, while the ARIMA model had a lower test error than the AR model, but its high SMAPE value on the training dataset suggested overfitting. The LSTM model had the lowest test error, but its high SMAPE value on the training dataset indicated that it may not have captured underlying patterns in the data well. The XGBoost model had a relatively low test error, suggesting good performance, and performed slightly better on the testing dataset than the ARIMA model. The study did not consider external factors that may impact website traffic, such as changes in search engine algorithms or other external shocks. These external factors can significantly impact website traffic, and not considering them may limit the generalizability of our study's findings.