two tests red on ci (py 3.10/3.11/3.12) after sktime 1.2.0 + pandas 3.0.6 released:
test_evaluate_per_fold_error_surfaces_as_failure — relies on ThetaForecaster.predict erroring on MS-freq DatetimeIndex (<MonthBegin> is not supported as period frequency). sktime 1.2.0 added _to_offset_compat which fixes that, so evaluate now returns success: true instead of a per-fold failuretest_failed_update_rolls_back_fitted_state — relies on update rejecting PeriodIndex y on a DatetimeIndex-fitted forecaster. sktime 1.2.0 relaxed the datatype checks so the update now succeeds
both tests need a deterministic failure to exercise the executor's error paths (error_score="raise" propagation and the update rollback). they should use inputs guaranteed to fail on any sktime version, not index quirks that upstream can fix.
fix: y with a NaN for the evaluate test (ThetaForecaster has capability:missing_values=False), categorical y for the update test (forecasters reject categorical endogenous data). verified passing on sktime 1.1.0/pandas 2.3.3 and 1.2.0/pandas 3.0.6.
two tests red on ci (py 3.10/3.11/3.12) after sktime 1.2.0 + pandas 3.0.6 released:
test_evaluate_per_fold_error_surfaces_as_failure— relies onThetaForecaster.predicterroring onMS-freqDatetimeIndex(<MonthBegin> is not supported as period frequency). sktime 1.2.0 added_to_offset_compatwhich fixes that, so evaluate now returnssuccess: trueinstead of a per-fold failuretest_failed_update_rolls_back_fitted_state— relies onupdaterejectingPeriodIndexy on aDatetimeIndex-fitted forecaster. sktime 1.2.0 relaxed the datatype checks so the update now succeedsboth tests need a deterministic failure to exercise the executor's error paths (
error_score="raise"propagation and the update rollback). they should use inputs guaranteed to fail on any sktime version, not index quirks that upstream can fix.fix:
ywith a NaN for the evaluate test (ThetaForecaster hascapability:missing_values=False), categoricalyfor the update test (forecasters reject categorical endogenous data). verified passing on sktime 1.1.0/pandas 2.3.3 and 1.2.0/pandas 3.0.6.