limen.features

Build higher-level model inputs and target helpers on top of raw bars and indicator columns.

Canonical docs

• Features
• Conserved Flux Renormalization
• Experiment Manifest

What this package owns

Owns the public feature-engineering layer, including regime features, lag helpers, breakout features, target helpers, and trade-shape features like CFR. Does not own raw technical indicators, feature scaling, or model fitting.

Key entry points

Entry point	Use it when	Notes
limen.features.* exports	You want feature functions inside a manifest or custom prep pipeline	The package root re-exports the public feature surface
lag_column, lag_columns, lag_range, lag_range_cols	You want lagged versions of existing columns	Useful for both raw and derived features
calendar_time_features, cyclical_time_features	You want calendar context or cyclical encodings from datetime	Time-of-bar features for schedules, regimes, and model inputs
parkinson_volatility, garman_klass_volatility, rogers_satchell_volatility, yang_zhang_volatility	You want range-based volatility estimates from OHLC bars	Useful when close-to-close volatility is too coarse
dollar_volume, amihud_illiquidity, return_per_dollar_volume, range_per_dollar_volume, illiquidity_shock	You want simple liquidity and impact proxies from OHLCV data	Keeps bar-derived liquidity features separate from trade-level microstructure
maker_*, trade_*, liquidity_*, taker_imbalance_ratio	You want maker/taker and LOB-liquidity features emitted by native klines	Uses columns from get_spot_klines and get_spot_dollar_klines
rolling_zscore	You want one primitive for rolling z-score features	Supports identity, log1p, and abs transforms
wick_proportion, stochastic_k_abs, distance_from_ma, close_ma_distance_atr, kaufman_efficiency_ratio	You want structural rolling OHLCV features	Adds bar-structure, distance, and path-efficiency context
realized_semivariance, realized_skewness, realized_kurtosis, jump_variation_proxy, tail_event_intensity, volatility_of_volatility	You want asymmetry, jump, and tail context around recent returns	These are useful risk-state features, not just volatility levels
relative_volume_seasonality, relative_range_seasonality, relative_volatility_seasonality	You want current bar behavior normalized against hour-of-week baselines	Helps detect unusually active or quiet bars in a 24/7 market
body_to_range, wick_imbalance, range_overlap, rejection_intensity, absorption_intensity	You want candle anatomy and auction-style context from plain bars	Focuses on rejection, overlap, and body-versus-wick structure
trend_coherence, volatility_term_structure	You want short/medium/long horizon agreement features	Summarizes cross-timescale alignment without a large feature bundle
conserved_flux_renormalization	You want trade-derived multi-scale flux diagnostics	Requires trade-level input rather than plain OHLCV bars

Adjacent modules

• limen.indicators usually runs first and produces columns that many features depend on.
• limen.transforms is commonly used alongside feature generation when building targets.
• limen.experiment.Manifest wires features into the prep pipeline through .add_feature(...).

Quick orientation

features/
├── lagged_features.py               # Lag helpers for arbitrary columns
├── breakout_*.py                    # Breakout and threshold features
├── *_volatility.py                  # Range-based and higher-order volatility features
├── *liquidity*.py, dollar_volume.py # Bar-derived liquidity and impact features
├── *seasonality.py                  # Hour-of-week normalized bar behavior
├── *intensity.py, *_range.py        # Candle anatomy and auction-style structure
├── volume_*.py, ma_slope_regime.py  # Regime and structure features
├── *_time_features.py               # Calendar and cyclical time context
├── distance_from_*.py, range_pct.py # Position and range features
├── vwap.py, ichimoku_cloud.py       # Higher-level technical composites
└── conserved_flux_renormalization.py

Things to know

• The package root exports the public feature surface, but not every helper file in the directory is part of that surface.
• Most manifest-driven uses run features lazily on pl.LazyFrame, so lazy-friendly expressions are the safe default.
• Some features assume earlier indicator columns already exist. The public reference calls these dependencies out.
• OHLCV-native feature helpers assume bars are already in chronological order before rolling or trailing calculations are applied.

Read next

• Features
• Conserved Flux Renormalization
• Experiment Manifest