I am trying to build neural network (NN) to forecast the probability of an event (e.g. thunderstorm will occur). As a base, I have a panel with weather data per state over 10 years. I saw some posts with similar questions (e.g. Keras Recurrent Neural Networks For Multivariate Time Series) and they all seem to use a RNN for this problem.
I would like to understand why a RNN seems the go-to solution and not e.g. a simple fully connected NN. Conventionally, I would use a logit model with fixed-effects for this problem.
Maybe someone can point me towards a paper or two which discusses this?
