Add SOFTSSharp model

[hljubic]

This PR adds SOFTSSharp, an extension of SOFTS that introduces stochastic variable-position encoding inside the STAD aggregation-redistribution component.

SOFTSSharp is intended as an incremental improvement over SOFTS while preserving linear computational complexity.

Reference implementation and ongoing research:
https://github.com/hljubic/SOFTSsharp

Changes:

• Adds the SOFTSSharp model implementation.
• Exposes SOFTSSharp through neuralforecast.models.
• Registers softssharp in the model filename mapping.
• Adds a minimal model test using the existing check_model AirPassengers workflow.

[CLAassistant]

All committers have signed the CLA.

[marcopeix]

Thanks for the contribution. Please, refer to these PRs to see how to contribute a new model to the repository:

• [FEAT] Add XLinear #1445
• fix: add XLinear to MODEL_FILENAME_DICT for save/load support #1469

In its current state, we cannot accept this PR. Thanks!

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[hljubic]

@marcopeix Thanks for the guidance. I updated the PR following the patterns in #1445 and #1469.

The PR now adds:

• AutoSOFTSSharp
• save/load support through MODEL_FILENAME_DICT
• model documentation and docs navigation entry
• capabilities overview entry
• tests for SOFTSSharp and AutoSOFTSSharp

Local checks passed.

[marcopeix]

Hi @hljubic. Looking good! A couple of things:

• Can you add an image as in [FEAT] Add XLinear #1445. A .png file that is included in the documentation.
• Can you also add a small comment to this PR to compare the results achieved between the paper and the implementation in neuralforecast, like in [FEAT] Add XLinear #1445. It helps validate the implementation. Small differences in the results are okay.
• Update the branch with the latest changes so we can merge.

Thanks a lot!

[marcopeix]

See comments above.

[hljubic]

@marcopeix Sorry for late response.

I added the requested PNG figure to the SOFTSSharp documentation and ran a paper-like ETTh1 96/96 reproducibility check for SOFTSSharp.

For ETTh1, I used h=96, input_size=96, the standard ETTh1 train/validation/test split, five random seeds [1, 2, 3, 4, 5], and the same main SOFTSSharp configuration across all runs.

Seed	MSE	MAE
1	0.393	0.407
2	0.391	0.406
3	0.391	0.407
4	0.385	0.404
5	0.387	0.407

Average across five seeds:

Model	MSE	MAE
SOFTSSharp	0.389	0.406

The results are consistent with the paper-level results, with slightly lower average errors in this implementation.

[marcopeix]

@hljubic sorry, some work was merged and the branch has a conflict in the auto version. Should be a simple fix, let me know if you need help! Otherwise, looks good!

[hljubic]

@marcopeix done, it should be okay now.

[marcopeix]

@hljubic
The linter check didn't pass. Make sure to install and run pre-commit hooks. Refer to our CONTRIBUTING.md.

[hljubic]

@marcopeix sorry, I added a new commit and double-checked it. I hope it’s okay now.