0step ahead prediction

Hi, I’m japanese student. I’m interested in time series data analysis with NuPIC.
So, I have installed NuPIC and tried running hotgym example. It’s amazing!

So far,I’m running hotgym example with various"predictionSteps"(2stepahead,4stepahead,…).
In my results, as predictionSteps is far as the result is bad.(at this point “bad” meaning that error of the predicted value and true value is big. ) I can understand this. I think far ahead prediction is difficult.

But when I tried 0step ahead prediction, the result is bad. Why is this?
Is’nt HTM good at predicting 0step ahead?

Sorry for my poor English. Thanks.

1 Like


I’m kind of interested in this question also, it seems that predicting 0 steps ahead is meaningless and probably shouldn’t be allowed? (Predicting what already happened - i.e. The present??)

@masajirv 1 step is normal operation and is the default (predicting what will happen next).

David is right. Since we are almost always dealing with time snapshots, a “0-step ahead” prediction means “the data point just seen”. So getting a prediction for now is only possible by using a previous inference from a timestep in the past that looked ahead to this point in time. But I’m not sure why that would be useful.

@rhyolight I’m wondering if an exception should be thrown if the classifier receives “0” as a configuration?