I have obtained 15 years worth of temporal data that I am using to build a neural network model. I am currently attempting to determine the best network architecture and hyperparameters, so I am using forward chaining which I think the below image explains pretty well:
From what I understand this is one of the best ways to go about training with temporal data. An alternative would be to use a fixed size training set at all stages, but this would limit the amount of data being used, which I would rather avoid.
My question is then:
- Will a certain architecture/hyperparameter setting produce different results for different sized datasets? E.g. Could my network work well for a dataset with 10,000 points but badly for one with only 1,000 points?
Through basic testing that I have performed myself I would assume that it does.
- If so, what is the best way to combat this so that I can get a good idea of the model's accuracy at each stage of the forward chaining process?