We know that Neural Network are usually designed for independent observations. This assumptions comes from the loss function where it is more easy to compute it when we have independent observations. But it is quite different when we have times series, for example the evolution of a stock index. If each observation is given by the date and the value of the stock index at this date, we know that each observation would be highly correlated. My questions are :
If we use a RNN how the network will understand that the observation at time t depends on the observations on the previous time t-1, because when we use RNN for sentence in NLP for example, the utility of the RNN is we keep a track of the previous word but in the same observation, and here it is not the same observation so is the RNN capable to keep a track of an another observation ?
Which model can I use to make prediction on time series data as the observations are not independent ? In other project for example bird classification on kaggle, we do know that all the Xi are independent (i represents the i-th image), but here I cannot find any good model that achieve a good score.
