limen.scalers

Canonical docs

In the research projects

The Python project fits a StandardScaler/Ridge pipeline and reuses those evaluated objects. The Compiled project fits its declared LogRegScaler on training features and reconstructs it with the selected round. The scaler reference distinguishes these fitted paths from stateless transforms.

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 case	Notes
`LinearScaler`	Rule-based scaling over mixed market-data columns	Exported at the package root
`LogRegScaler`	Standard scaler used by logistic-regression SFDs	Exported at the package root
`RobustScaler`	Median and IQR scaling for outlier-heavy data	Exported at the package root
`CausalRollingRobustScaler`	Robust scaling that adapts to drift, with no look-ahead	Exported at the package root
`RankGaussScaler`	Rank-based Gaussianization	Exported at the package root
`SCALER_REGISTRY`	Resolve scalers by manifest parameter name	Used by `set_scaler_from_params()`
`build_rules`, `inverse_transform`	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

Module or family	Responsibility
`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

text · README.md

1scalers/2├── README.md3├── __init__.py4├── causal_rolling_robust_scaler.py5├── linear_scaler.py6├── logreg_scaler.py7├── rank_gauss_scaler.py8├── registry.py9└── robust_scaler.py

Merged source PR · 0102f669Package layout · Limen 6.0.1 · current package validation · package source

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.
build_rules places overrides before base rules, including existing-key replacements. Each override entry becomes an unescaped ^entry$ regex; remaining base rules keep their order.
Unrecognized columns fall through to the catch-all none rule and stay unchanged.
Columns assigned standard or log_standard scaling must have a finite, nonzero training standard deviation. LinearScaler and LogRegScaler do not guard zero variance; division by zero produces non-finite output.