Prediction Deployment

timocall

Hi Forum, 

Using the brilliant Tom_Ott's tutorial on "Predicting Historical Volatility with RapidMiner" as a template, I've changed a couple of inputs and parameters to suit my requirements and have a model that is producing a pretty strong output (prediction trend accuracy = 0.800). The challenge I'm having is how do I use this to show the predicted trend on the horizon?

ie. The model tells me the prediction against known actuals but how do I translate that to t+1, t+2,t+3 etc view? In my example this would be Jan 15th, Jan 22nd etc etc

I'm sure I'm missing something basic here so apologies if this is a no brainer question. Any help would be greatly appreciated!


 
Cheers,
Tim

BalazsBaranyRM

Hi @timocall,

I see you have attributes like Date-0 in your data. You can only predict from data that you have at the time of prediction. This might also lead to your model being so good, as it can learn from very recent data.

If you want to predict e. g. 7 days in the future, make sure that you only learn from data at least this *old*. E.g. only use the -7, -8, ... attributes. 

Then if you take the data you have for the prediction (today, yesterday) and set up the same attributes (e. g. call today's data Date-7), you'll get a prediction that is 7 days into the future.

You can validate this method with a sliding window validation. Make sure that no more recent data than your prediction horizon leaks into the model, because if you do that, you'll lie to yourself.

Best regards,
Balázs