I'm currently working on timeseries forecasting in pairs where one timeseries is suspected to cause the other timeseries. I also have the forecast of the causing timeseries and I use them to predict the other timeseries. Note that extrapolation capabilities are required since the forecasts of the first timeseries are not always in the range of the training data.

I'm trying to use a simple LSTM model for that task and I tried two ways: training a model using past values of both timeseries and the current value of the first timeseries, and training using only the past and current values of the other timeseries.

To my surprise, the forecasts using only the first timeseries values without the values of the timeseries we forecast are better. Does this make sense or am I doing 100% something wrong? When scaling the data do I need to scale values of each timeseries separately or scale them together?

Also, what other ways you suggest for forecasting using another timeseries data and existing forecast?



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.