Forecast instability

[fbonaita]

Hi,

I've recently tested both Prophet and NeuralProphet on a public dataset (bike rentals in London) to see how their performance compares and I've run in some interesting behaviours.

They seem to forecast very similarly with the default setup but once I add extra regressors (weather data), Prophet consistently improves its forecasts while NP seems to struggle to converge towards stable predictions across different cutoff dates. Sometimes the forecast follows the pattern you would expect, other times it's just a flat-ish line.

Below is an example of such behaviour (Prophet on the left, NP on the right, compared on two cutoffs).

In all of this AR-Net is always left disabled, the time series is quite regular and daily data is provided for several years so
I guess I would have expected for NP to behaves similarly to Prophet?

I'm mostly familiar with traditional statistical modelling and I approached neural nets only very recently but it'd be great if you had any insights into this!

Thanks