Skip to content

v0.1.21

Choose a tag to compare

@AzulGarza AzulGarza released this 16 Nov 22:39
4f3c940

Features

  • CIProto: The CIProto wrapper supports Continuous Intelligence's ContinuousIntelligence/ci-proto, a ~167M-parameter zero-shot probabilistic forecaster with a very long context (16,384 steps on the standard checkpoint; longer histories switch to the long checkpoint automatically). It returns nine native quantiles between 0.1 and 0.9 and accepts known-future covariates through X_df. See #99.

    import pandas as pd
    
    from foundationforecast import FoundationForecast
    from foundationforecast.models import CIProto
    
    DATA = "https://timecopilot.s3.amazonaws.com/public/data/electricity_price"
    df = pd.read_parquet(f"{DATA}/train.parquet")
    X_df = pd.read_parquet(f"{DATA}/futr_exog.parquet")
    
    ff = FoundationForecast(models=[CIProto()])
    fcst_df = ff.forecast(df=df, h=24, X_df=X_df, level=[80])
  • Optional ci-proto extra: CIProto needs the continuousintelligence package, installed only with foundationforecast[ci-proto]. It is not a core dependency, foundationforecast.models imports without it, and constructing the model without the extra raises an ImportError with the install command. foundationforecast.models.ciproto.is_ciproto_available() reports whether the extra is installed.

  • Conformal finite-sample intervals: New opt-in method="conformal_error_finite_sample" for ConformalIntervals uses the exact ceil((n + 1) * level / 100)-th absolute-error order statistic, without interpolation. When the rank exceeds the calibration sample size the bounds are infinite. Existing methods and defaults are unchanged. See #92.

Fixes

  • Post-publish release smoke test: The After publishing test step now runs outside the repo checkout, pins foundationforecast==<released version>, and retries until the PyPI index serves it, so it checks the published wheel instead of the repo source. See #98.

Tests

  • CIProto: Mocked-model tests in tests/models/test_ciproto.py (loader, cache key, batching and left padding, timestamps, native future covariates, level and quantiles, install hint). CIProto joins the test_models.py parametrization whenever the ci-proto extra is installed. The example notebook was executed on a Modal A10G with the real checkpoint.
  • Conformal finite-sample: Contract test parametrized over every registered model and an executed M4 Hourly rolling-origin comparison of coverage, width, and finite intervals.

Documentation

  • CIProto family notebook: ci-proto-family example notebook forecasts with CIProto on M4 Hourly, with native quantiles and known-future covariates.
  • Model hub, README and installation: Document the CIProto checkpoint, add it to the supported-models table and the exogenous-variables table, and show the optional extras as pip/uv tabs.
  • Conformal prediction guide: Documents the finite-sample method and its exchangeability assumption.

New Contributors


Full Changelog: v0.1.20...v0.1.21