I have a time series forecasting project, there are over 10, 000 time steps of data, so the data amount is not a problem. At first, I thought I've to create a time-based data pipeline that forms the data architecture like in shape:
[batch size, time step, number of features]
Which is the typical data input shape of LSTM. However, I've seen others use an single day's feature as input and the predict target in the same day as output.
This means that they apply the features to predict the target value of same day. This makes some sense though, because the input features contains the 'date' feature like: 2015-01-09, so I believe in the single day feature it also include the timing.
However, if we have only one variable that forms the time series sequence itself, I couldn't train in the same way.
Will it be better to make prediction of future values by several time steps of past data? Will it make the model more robust?
Or it's although differ by case, that I should test when encounter a different data?