Apologies for the noob question. I am attempting time series forecasting (with a combination of lag and categorical features) using tensorflow and struggling to find an optimal combination of RNN/LSTM cells and some attention, convolution components (Luong attention as well as multihead attention), as there seems just too many parameters to decide and an infinite number of possible architectures. For what it is worth, I did spend some time to understand the functionalities of attention, LSTM etc. and internal operations of the layers. But that does not help in deciding the architecture. So is it purely an art, where I mix some experience with experimentation to arrive at a model architecture, or is there some guidance, method, or even library for an architecture search? Any help will be gratefully appreciated.
For more specifics
- I am trying to predict the next seven days of combined sales of some items using past seven days of sell as lookback features, and some item details as categorical features. The resolutions of the target and lookback features are different. Lookback uses past seven individual daily sales of items, but the goal is to predict next seven days total sale. I did the feature engineering and got the training samples already.
- The loss function to use is mean-absolute-percentage-error
- I am mostly familiar with tensorflow APIs, subclassing the layers, defining the model functionally etc. But cannot decide on a specific architecture of the model with the components to fit the data.
