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

Introduction

Time series forecasting refers to using historical data in a period to predict the future. The application scenarios of time series forecasting are relatively rich, including quantitative trading, macroeconomic trend forecast, merchandise sales forecast in e-commerce, store sales forecast in the retail industry, and electricity forecast of power supply. The input and output format of time series forecasting is similar to regression problems in supervised learning, but there is one fundamental difference: time-series data usually has time dependence (Robinson & Sims, 1994). For a batch of input data of the same size, if the order of the data is changed, all data patterns and relations are likely to undergo huge changes, and the prediction results of models may be completely different as well. At the same time, other regression problems may not have this characteristic.

Although there are many successful deep learning applications in time series forecasting in academia, many industrial and commercial applications still use traditional statistical models or machine learning models to deal with time series forecasting problems. The industrial and commercial sectors are still in the exploratory stage of applying deep learning technology to time series forecasting. The main reasons are:

• 1.

  Most time-series forecasting in academia use a very small dataset for testing complex deep learning models, which is of little significance for evaluating the actual performance of models. Few studies used industrial-grade datasets to test these models’ performances. Thus, conclusions drawn from such experiments are not quite convincing;

• 2.

  Compared with mature fields, such as natural language processing, computer vision, and recommendation systems, the research field of time series forecasting is relatively niche. This means that the industry lacks related expertise; and

• 3.

  At present, the industries that can generate the greatest benefits from time series forecasting are mainly the financial industry, especially quantitative trading. However, due to the unique nature of the industry, the results and details of such applications are rarely disclosed.

Thus, the objectives of this chapter are:

• 1.

  Comparing the difference in model performance between the statistical model, machine learning model, and different deep learning models on the same data set; and

• 2.

  Comparing the performance differences of different deep learning models on the same data set in detail, and summarize the advantages and disadvantages of different deep learning algorithms.

Time series forecasting has many important applications in commercial and industrial sectors. However, compared to popular fields, such as natural language processing, recommendation systems, and computer vision, time series forecasting is a relatively niche research field. Few articles have summarized the most valuable method, i.e., deep learning time series forecasting algorithms, so far. Therefore, to make up for this gap, this chapter will compare in detail the effects of different deep learning algorithms on the same real-life industrial data set and evaluate the performance of different deep learning algorithms. These algorithms are representative and popular and used as the solutions of the time series forecasting in the industry and competitions. They have potential great commercial value. This chapter provides references for various industries to solve practical time series forecasting problems in the real world.