I have worked with lstm's in the past, specifically for time series forecasting. However, the target feature in these time series were relatively "stable". With the loosely defined "stable" I refer to the case where there are no significant changes or anomalies. Take the figure below for example, the time series contains months where there are zero counts/values.

enter image description here

I'm actually doing multivariate, but for the sake of the example let's say it's univariate. If a training instance is in the form of $ (y_t,y_{t+1},y_{t+2}) $ with corresponding target $y_{t+3}$, then you could get a training instance along the line of, for example, $x_i = (100, 110, 130) $ with target of $y_i = (0)$ when preparing the training dataset. Basically, I'm worried the trained model will give me a forecast of a very small value, given a sequence of large values.

Will this mess with the model, or can I just train as usual? Other option is to aggregate to quarterly, but this will significantly reduce the data. Or remove those training instances where a zero value occurs in the target.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.