limen.data.utils

Canonical docs

In the research projects

The Python project owns its walk-forward loop and selection inside each fitting window. The Compiled project records the fold campaign and reconstructs qualified members. The shared splitting helpers supply geometry; the project controls what is fitted and assessed.

What this package owns

Owns the manifest-facing data helpers: the compute_data_bars() bar-formation entry point, the train/validation/test split functions, and random window slicing. Does not own raw data access (owned by limen.data.HistoricalData) or the low-level bar builders (owned by limen.data.bars).

Key entry points

Entry point	Use case	Notes
`compute_data_bars`	Public bar-formation entry point	Dispatches to `limen.data.bars`; used by `set_bar_formation()`
`split_sequential`	Ordered train/validation/test windows	The standard manifest split
`split_walk_forward`	Expanding or rolling chronological folds	Returns bounded folds; fitting remains a per-fold responsibility
`split_random`	Randomised split	Non-sequential alternative
`split_by_dates`	Explicit date-boundary split	For fixed calendar windows
`split_data_to_prep_output`	Build the `data_dict` (`x_train`, `y_test`, …)	Does not mutate the passed split list or column list
`split_data_to_rule_based_prep_output`	`data_dict` variant for rule-based SFDs	Preserves columns rule strategies read
`random_slice`	Random contiguous window	Sampling helper
`lookback_window`	Trailing elapsed-day window	Ends at the last supplied observation
`sample_regime_blocks`	One contiguous causal-regime block	Explicit regime version, seed and minimum length

Adjacent modules

limen.data.HistoricalData produces the raw frames these helpers split.
limen.data.bars provides the bar builders compute_data_bars() dispatches to.
limen.experiment consumes the resulting data_dict through manifests and the Universal Experiment Loop.

Quick orientation

Module or family	Responsibility
`compute_data_bars.py`	Public bar-formation entry point
`splits.py`	Train/validation/test split helpers
`split_walk_forward.py`	Walk-forward fold geometry
`random_slice.py`	Random window slicing helper

text · README.md

1utils/2├── README.md3├── __init__.py4├── compute_data_bars.py5├── lookback_window.py6├── random_slice.py7├── sample_regime_blocks.py8├── split_walk_forward.py9└── splits.py

Merged source PR · 0eb799c4Package layout · Limen 6.0.1 · current package validation · package source

Things to know

split_data_to_prep_output() is non-mutating: it copies the split list and column list before building data_dict.
The data_dict keys (x_train, y_train, x_val, x_test, y_test, …) are the schema the rest of Limen's model code expects.
Rule-based SFDs need split_data_to_rule_based_prep_output() because they read raw columns the standard prep drops.