JQGoh commented 4 months ago

Changes

Fixes https://github.com/Nixtla/neuralforecast/issues/866 issue
Add optimizer, optimizer_kwargs params to TimeLLM and NLinear models. The introduced models did not consider these parameters because the optimizer-related args were also introduced about the same time when these models were developed.

Rationale of fix

Instead of filtering the generated cutoffs by first assuming step_size = 1, we directly impute the series provided by the users internally when performing predict_insample().
As compared to the approach considerd in earlier work, we still pass effective step_size parameter to _insample_times() to validate np.divmod(sizes - h, step_size), ensuring modulo operation = 0.