Dilated neural networks for time series forecasting

Dilated neural networks are a class of recently developed neural networks that achieve promising results in time series forecasting. Chenhui Hu discusses representative network architectures of dilated neural networks and demonstrates their advantages in terms of training efficiency and forecast accuracy by applying them to solve sales forecasting and financial time series forecasting problems.

Recently, deep neural networks (DNNs) have been used with great success to solve time series forecasting problems such as web traffic forecasts, product sales forecasts, and financial time series forecasts. The ability to capture long-range and nonlinear data dependencies is the key reason why deep learning methods can achieve better forecast accuracy than traditional machine learning approaches. However, DNNs need more time and data to train. Chenhui Hu offers an overview of dilated neural networks—a class of DNNs in the frontier of deep learning research that tackle the above challenges—and demonstrates their advantages with real use cases. This class of networks includes both dilated convolutional neural networks (dilated CNNs) and dilated recurrent neural networks (dilated RNNs). Based on dilated connections, dilated neural networks can capture data dependencies in a large spatial or temporal range while only requiring that the number of parameters grow logarithmically. As a result, they are more efficient to train when compared with the current state-of-the-art DNN models such as long short-term memory (LSTM) models. Chenhui demonstrates the advantages of dilated neural networks in terms of training efficiency and forecast accuracy by applying them to solve sales forecasting and financial time series forecasting problems and shows that they can obtain at least as good or better accuracy on such nonlinear, noisy forecasting tasks. Source code for the implementations will be available in GitHub.