In week 3 where we used RNN and LSTM, the mae was found to be ~5 to 6. But in week 2 using a simple 3 layer DNN we got around mae=4.5
If DNN gave a better performance why do we have to use a heavier model like LSTM?
Or, is it that for this data it turned out to be like this, but generally LSTM would perform better for time-series data due to carrying forward of state?
Yeah in general an LSTM would logically outperform the DNN because it has whats called “memory”. For this particular case it seems that for this data the DNN is doing a bit better but there are many parameters to consider and tune further and also this is synthetic sinplistic data as far as I remember.
However one can not say for sure which method could turn out to be more predictive for time seties because time series are rather unpredictable and complex in their nature