Deep learning for time series forecasting and nowcasting
This study is an exploration of where we can expect added value for forecasting and nowcasting time series in official statistics by using deep learning techniques, as an alternative to classic time series models.
We will compare several neural network algorithms, identify the key differences with classic time series methods and determine in what situations we can expect to benefit from these algorithms. In an empirical study, the methods are applied to several time series. We find that neural network algorithms can yield similar forecast and nowcast accuracy as classic methods for univariate time series, but it requires some effort to achieve this. When applied to a more challenging problem with several auxiliary variables and a more volatile series, in our case the Long Short-Term Memory model gave accurate results. This leads to the conclusion that deep learning can offer added value compared to classical methods for specific problems, but more research is needed.
