Roadmap 2026

[fkiraly]

From roadmap planning meeting on 2025-11-10

Present: @agobbifbk, @ankurankan, @fkiraly, @julian-fong, @phoeenniixx, @PranavBhatP, @RecreationalMath, @XAheli

Topics & Prioritization

pytorch-forecasting v2 API rework 👍 👍 👍👍 ✔️

• Addition of predict function and full integration of pkg class with rest of D1,D2, model layer pipeline
• Addition of preprocessing steps - scaling, transformations etc in D1/D2 layers
• Improve the test coverage.
• Add temporal splitting
• Migrate the models from v1 to v2 and deprecate v1
• Add extension Templates for new contributors.
• Add adpaters for nn losses.

Extending model zoo 👍👍👍✔️

• Addition of new models to DSIPTS.

Maintenance items, CI, tests, avoiding tech debt 👍👍 ✔️

• initial review of pytorch-tabular API, comparison to pytorch-forecasting
• Improve the test coverage in pytorch-forecasting
• negative tests and validation tests to be added for PyTorch based models.
• Add python-3.14 support for ptf (scheduled for the next release).
• urgent maintenance items in pytorch-tabular - pyproject, python 3.14
• adding torch to all extras in sktime, or replace tensorflow by torch

backend - sktime GPU support for torch based models 👍

• Support for GPU utilisation to be added

API homogenization of torch models within sktime 👍👍👍✔️

• Unify scattered torch dataset classes for classification, regression etc.
• multiple _pytorch.py exist in sktime repo, they should all be unified into a single class in a single file with a _pytorch_util.py containing all the helper functions.
• activation, criterion, optimizer, callbacks initializations currently done crudely. Improvements to be made.

API homogenization between sktime, pytorch-forecasting, pytorch-tabular, DSIPTS 👍 👍 👍 👍✔️✔️

• uniform way of network creation and definition. API design that can be re-used everywhere.
• Create an interface for pytorch-tabular and pytroch-forecasting much like prophetverse and sktime
• Having a uniform way of implementing DL based models across sktime, Ptf & ptt. Since we are implementing PyTorch based models in sktime from scratch, it will be great if we can borrow best practices from sister libraries and implement them in sktime.
• uniform way of network creation and definition. API design that can be re-used everywhere (sktime, ptf, ptt). This would also help in maintaining all three repos and allow cross collaboration.
• Handling training and prediction pipelines in a more modular way than currently in sktime. Learnings from ptf & ptt to be incorporated.
• Synchronise on ideas/approaches on Data layer & Modelling; ensuring benefit to both the repos.
• Enhancement related to common workflows in D1-D2 layers.

New Features

Global model support in PTF and sktime

• adding groups (categroical variable) and distinction between global and no global forecast

Foundation model support in PTF 👍👍 ✔️

• Addition of foundation models to PTF v2 (in parallel with the rework) - helps confirm the feasibility of the design for all the proposed models for v2. Re: open issue

Benchmarking support

• Benchmarking support for PTF v2. Very open ended right now.

Workstreams

one big workstream due to interconnection

• design / API makes most sense jointly
  • have design for pytorch-forecasting
  • no consilated design for sktime / torch
  • "inherited" design for pytorch-tabular
• short-term, "big projects"
  • maintenance items are independent
  • pytorch-forecasting v2
  • pytorch-forecasting adding models - contingent on API work
  • sktime API refaector
  • sktime adding models (could happen right away)
  • pytorch-tabular (anything) - need to onboard maintainers to pytorch-tabular
  • cross-cutting design work
  • cross-cutting API changes (contingent on design)

Prioritization:

• progress work that is currently ongoing

  • pytorch-forecasting v2 and model add
  • sktime API rework
  • pytorch-tabular basic maintenance
    • perhaps onboard new mentees
  • DSIPTS rework

• design review

  • create shared design document on the topic of torch related API in 4 packages
    • action Nirbhai - new hackmd doc to fill - collect current designs and notes
    • action Aryan - send link to existing ptf document
  • presentations where necessary (tbd)
  • aim is for all to gain working knowledge of the three packages
  • based on that, joint design work