limen.scalers

Fit train-only feature scaling and apply it consistently across validation and test data.

Canonical docs

• Scalers
• Experiment Manifest

What this package owns

Owns fitted scaler classes, the scaler registry, and the rule-based logic that decides how columns should be transformed. Does not own raw feature creation or the experiment loop itself.

Key entry points

Entry point	Use it when	Notes
LinearScaler	You want rule-based scaling over mixed market-data columns	Exported at the package root
LogRegScaler	You want the standard scaler used by logistic-regression SFDs	Exported at the package root
RobustScaler	You want median and IQR scaling for outlier-heavy data	Exported at the package root
CausalRollingRobustScaler	You want robust scaling that adapts to drift, with no look-ahead	Exported at the package root
RankGaussScaler	You want rank-based Gaussianization	Exported at the package root
SCALER_REGISTRY	You want to resolve scalers by manifest parameter name	Used by set_scaler_from_params()
build_rules, inverse_transform	You need to customize or interpret LinearScaler behavior	Available from the module-level implementations

Adjacent modules

• limen.experiment.Manifest is the main consumer of this package.
• limen.transforms handles stateless transforms, which is a different stage from fitted scaling.
• limen.features and limen.indicators produce the columns that scalers later operate on.

Quick orientation

scalers/
├── linear_scaler.py                 # LinearScaler, rule helpers, inverse transform
├── logreg_scaler.py                 # Logistic-regression tuned scaling
├── robust_scaler.py                 # Median and IQR scaling
├── causal_rolling_robust_scaler.py  # Causal trailing-window median and IQR scaling
├── rank_gauss_scaler.py             # Rank-based Gaussianization
└── registry.py                      # SCALER_REGISTRY

Things to know

• Scalers are fit on the training split and then reused on validation and test splits. That fit/apply discipline is part of why this package stays separate from limen.transforms.
• LinearScaler uses ordered regex rules. The first matching rule wins.
• Unrecognized columns fall through to the catch-all none rule and stay unchanged.
• Zero-variance columns are passed through rather than exploding the fit step.

Read next

• Scalers
• Transforms
• Experiment Manifest