Forecasting multiple time series can quickly become a complicated task; Traditional approaches require either a separate model per series (i.e., SARIMA) or for all series to be correlated (i.e., VARMA). Neural networks offer a flexible approach that allows forecasting of multiple series with a single model, regardless of the correlation of the series.
Furthermore, this approach allows for easy incorporation of exogenous variables and can forecast multiple time steps into the future, resulting in a powerful general solution that works well in a wide variety of cases.
In this article, we will show how to perform the data windows necessary to transform our data from a time series to a supervised learning format for univariate and multivariate time series. Once our data has been transformed, we will show how to train both a deep neural network and an LSTM to perform multivariate forecasting.
Examining our data
We will be working with a dataset that captures the daily mean temperature and humidity in Delhi, India, between 2013 and 2016. This data is available on Kaggle and is licensed for use under the CC0: Public domain making it ideal…
