Deep sequential models versus statistical models for web traffic forecasting

Web traffic forecasting is a challenging problem for most websites; however, accurate forecasting leads to improved customer experience. Most websites collect data on the number of visitors visiting their site each day. This is a time series problem and there are various techniques that can be used to make accurate forecasts using this data. Machine learning (ML) and deep learning (DL) techniques (also termed artificial intelligence (AI)) are used in various domains for time series analysis. However, before the advent of AI, many statistical techniques were utilized in solving forecasting problems. Due to the sophisticated structure of DL models, we hypothesized that deep sequential models would perform better on time series data for web traffic forecasting as a multi-horizon problem than traditional statistical models like Auto-regressive integrated moving average (ARIMA) and Seasonal auto-regressive integrated moving average (SARIMA). We analyzed Wikipedia web traffic data to compare the performance of statistical models with that of the popular deep sequential models like Long short-term memory (LSTM), gated recurrent units (GRU), convolution neural networks (CNN) and bi-directional LSTM (BiLSTM). Under the two-evaluation metrics used in this study, the GRU and BiLSTM models showed the best performance. In general, DL models outperformed statistical models for web-traffic forecasting. Therefore, our data supports the hypothesis that deep sequential models are a better choice compared to statistical models for web traffic forecasting. Our results can be used to further investigate the performance of deep sequential models for time series analysis under other domains.