Hi All,
My question is about modeling time series using LSTM (Long-Short-Term-Memory).
I have 18 response variables for which all of them are monthly time series for about 15 years, and I would like to try to model it with LSTM and forecast into the next year. And I also have a independent/predictor variable (exogenous signals) pool about 400 macro-economic signals.
For starters, about LSTM modeling for Time Series:
Do I need to difference the response sequences?
Do I need to difference/lag the independent/predictor sequences?
Do I need to de-trend/de-season the response sequences first?
Do I still need stationarity?
Do I need to do a structural break before modeling?
And since I'm new to LSTM model, are there any rule-of-thumb in terms of the setup of the LSTM? Like how many layers should I have and how should I set-up each layer?
And finally, in terms of evaluating the LSTM model, should I do a rolling-window back-testing? Like row the training window back one month at a time and hold-out 1 year data to do back testing? Is there anything LSTM-specific that I should evaluate/test?
Thank you so much for your help!
All the best,
Kathy G.
2. use a two layered LSTM architecture coupled with a dense output layer to make a prediction.
3. look at couple of approaches to predict the output — a.) Forecasting step by step on the test data set, b.) Feed the previous prediction back into the input window by moving it one step forward and then predict at the current time step.
- Badges
- Science topic