feat: add support for T0 (theforecastingcompany/t0-alpha)#348
Draft
GeoffNN wants to merge 1 commit into
Draft
Conversation
Adds The Forecasting Company's open-weights T0 foundation model as a foundation forecaster, via the tfc-t0 package (Python 3.11–3.13). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Adds T0, The Forecasting Company's open-weights time series foundation model, as a new foundation forecaster. T0 (
t0-alpha, ~102M params, Apache-2.0) is a decoder-style patch transformer producing probabilistic multi-horizon quantile forecasts — currently #1 on the fev-bench leaderboard (skill score 42.2) and CRPS 0.4941 on GIFT-Eval.(Disclosure: I work at The Forecasting Company.)
Changes
timecopilot/models/foundation/t0.py—T0forecaster wrapping thetfc-t0package, following the existing foundation-model pattern (_get_modelcontext manager,TimeSeriesDatasetbatching,QuantileConverterfor levels/quantiles). The model predicts 5 quantile knots and interpolates arbitrary requested levels in a single forward pass; the median is the point forecast. Ragged batches are left-padded with NaN, which T0 treats as missing.pyproject.toml—tfc-t0>=0.1.2for Python 3.11–3.13 (the package's supported range), mirroring the TiRex/uni2ts marker pattern.tests/models/conftest.py—T0(context_length=256, batch_size=2)added to the model matrix under a version guard, plus atest_t0_import_failsguard test mirroring TiRex/Sundial.Status
Note
Draft until
tfc-t00.1.2 lands on PyPI (in flight): it's a metadata-only release relaxing the einops/jaxtyping floors so tfc-t0 coexists withtimecopilot-uni2ts(einops==0.7.*,jaxtyping<0.3). Once published I'll regenerateuv.lockand mark this ready. With those floors,uv lockresolves cleanly (verified locally).Validation
Run locally on macOS / Python 3.12 (with the 0.1.2 floors):
forecastwith defaults,quantiles=[0.1, 0.5, 0.9], andlevel=[80]on multi-series daily data: correct columns, finite values, monotone quantiles, median == point forecast.cross_validation(df, h=6): correct shape and columns.uv run pytest tests/models/test_models.py::test_t0_import_fails(skips on 3.11–3.13, asserts ImportError outside).ruff check+ruff format --checkpass on the new/changed files.🤖 Generated with Claude Code