Forecasting multiple time series can quickly develop into an advanced task; traditional approaches either require a separate model per series (i.e. SARIMA) or that every one series are correlated (i.e. VARMA). Neural Networks offer a versatile approach that allows multi-series forecasts with a single model no matter series correlation.
Moreover, this approach allows exogenous variables to be easily incorporated and might forecast multiple timesteps into the long run leading to a strong general solution that performs well in a wide range of cases.
In this text, we’ll show the best way to perform the information windowing required to rework our data from a time series to supervised learning format for each a univariate and multivariate time series. Once our data has been transformed we’ll show the best way to train each a Deep Neural Network and LSTM to make multivariate forecasts.
Examining Our Data
We’ll be working with a dataset capturing day by day mean temperature and humidity in Delhi India between 2013 and 2016. This data is out there on Kaggle and is licensed for usage under the CC0: Public Domain making it ideal…
