Model overfits greatly when given partial yearly data

[bwassie]

Hi folks. I'm using NeuralProphet for time series forecasting at the daily level. I have around 3 years of data right now and if I feed the model 2.5 years of data as training, it overfits severely on the remaining set. However, if I fit it with only 2 years of complete data, the model is very accurate on the holdout set. Has anyone seen this behavior before? I'm not sure if it's because I have limited data or what. I'm tuning for most parameters (all regularization terms, n_lags, epochs, learning rate...etc) except for number of layers (fixed at 0). I'm attaching the time series below. I'm very perplexed by this and would appreciate some guidance.

[ourownstory]

Hi @bwassie this is interesting.
Can you please provide a minimal code snippet to reproduce this?

(on a first guess: be careful when using a large number of lags and forecasts. the number of parameters is proportional to their product.)