Hi everyone. I want to use Neuralforecast to predict the offset produced by a ground-based telescope in different time series, but these time series are unevenly spaced time series. Is it possible to use models (like AutoNBEATSx or AutoNHITS) with these kind of time series?

Having a variable sampling rate is an interesting issue. My approach in the past has been resampling the signal at a fixed rate. But I would be very interested to know if there is a forecasting technique that can natively handle signals with time-varying frequencies.

Hi @Angel Berihuete Macías! The current methods do not natively incorporate the time delta between timestamps to make predictions. They fundamentally assume the time delta between samples is constant. Having unevenly spaced time series will lead to unexpected behaviour, probably producing inaccurate forecasts. I suggest, as Farzad mentioned, to resample the signal.

Many thanks!

I don't have an example handy. I used to do it in MATLAB years ago when I was doing signal processing. You need to either upsample or downsample to get to a uniform frequency. I haven't done it in Python but I'm sure there are libraries for it. If you were going to write your own code you would take the largest time step you have and interpolate your signal at those intervals for downsampling. Numpy does the interpolation if you give it your timestamps. For upsampling you can take the smallest time interval you have and repeat the same process. You can also choose any time interval between your smallest and largest.

Thank you for your advice @Farzad E!