explore residuals as target

[SarahAlidoost]

in #2, the implementation uses monthly sst as target. But. we can use residuals as target. Lets explore it:

target = residuals = monthly_data_full - daily_data_with_gaps.mean()

[SarahAlidoost]

In this notebook, I compared two models 1. trained on target = monthly_data_full, 2. trained on target = residuals = monthly_data_full - daily_data_with_gaps.mean(). I also compared the results on a test subset. The results shows that predicting residuals instead of the full monthly field is better.

But there are some considerations in this example: the var(residuals) << var(monthly_data_full). This makes training easier. There were not many gaps in time for this specific example. This needs further investigations when we have the proper training workflow including dataloader in place. For now, I suggest considering both scenarios in our examples.