limen.transforms

Canonical docs

In the research projects

The Python project shifts observations while constructing its next-bar label. The Compiled project declares a shifted fitted target. The transform reference explains why a stateless operation does not own the train-fitted state used on later splits.

What this package owns

Owns stateless DataFrame transforms. Does not own fitted feature scaling, higher-level feature engineering, experiment orchestration, or post-model helpers such as probability calibration and threshold optimization (see limen.calibration).

Key entry points

Entry point	Use case	Notes
`shift_column_transform`	Forward-looking target column	Available inside caller-owned preparation
`zscore_transform`	Frame-wide normalization step	Stateless per call
`winsorize_transform`, `mad_transform`, `quantile_trim_transform`	Clipping or row filtering before modeling	These do not carry their own train-only state

Adjacent modules

limen.features and limen.indicators produce the columns these transforms often operate on.
limen.scalers handles fitted train-only scaling, which is a separate concern from this package.
limen.calibration handles probability calibration and threshold optimization after model training.
limen.experiment.Manifest provides the most common place to attach target transforms.

Quick orientation

Module or family	Responsibility
`shift_column_transform.py`	Forward-looking target shifts
`zscore_transform.py`	Stateless DataFrame normalization
`winsorize_transform.py`	Percentile clipping
`mad_transform.py`	Median absolute deviation scaling
`quantile_trim_transform.py`	Quantile-based row trimming

text · README.md

1transforms/2├── README.md3├── __init__.py4├── mad_transform.py5├── quantile_trim_transform.py6├── shift_column_transform.py7├── winsorize_transform.py8└── zscore_transform.py

Merged source PR · 87894eb8Package layout · Limen 6.0.1 · current package validation · package source

Things to know

All helpers in this package are intentionally stateless per call.
For fit-on-train and apply-to-all semantics, use a manifest fit parameter or a scaler rather than assuming a transform will remember state.
shift_column_transform with shift=-1 creates a future-looking label and therefore introduces a trailing null row that later prep stages drop.