TrackRecordWeighting

TrackRecordWeighting(
    gap_column='consensus_gap',
    surprise_column='surprise',
    skill_column=RELATIVE_SKILL_COLUMN,
    skill_quantiles=defaults.SKILL_BUCKET_QUANTILES,
    winsorize_fraction=defaults.TRACK_RECORD_WINSORIZE_FRACTION,
    maximum_absolute_gap=defaults.TRACK_RECORD_MAXIMUM_ABSOLUTE_GAP,
)

Weight the gap by how well the forecaster’s skill tercile has predicted surprise.

Attributes

Name	Type	Description
gap_column	str	Column holding the forecast-consensus gap.
surprise_column	str	Column holding the realised surprise against consensus.
skill_column	str	Column holding the trailing skill.
skill_quantiles	tuple[float, float]	Training quantiles separating the three buckets.
winsorize_fraction	float | None	Share of each tail of surprise and gap clipped before fitting; None clips nothing.
maximum_absolute_gap	float	Training rows with a wider gap are left out.

Methods

Name	Description
fit	Fit the cut-offs and the coefficients on training rows alone.

fit

fit(subjects)

Fit the cut-offs and the coefficients on training rows alone.

Parameters

Name	Type	Description	Default
subjects	pd.DataFrame	Training rows only, carrying surprise, gap and skill.	required

Returns

Name	Type	Description
	FittedTrackRecordWeighting	The fitted weighting.

Raises

Name	Type	Description
	PanelError	A column is absent, or no training row survives.