Features define the information available to a model

The Python project constructs features in ordinary functions. The Compiled project declares indicators and features in YAML. Both projects retain recorded observations and explicit preparation boundaries.

Prerequisites

The matching wheel and recorded-input bundle supply the complete projects. The preparation diagnostics require the data and stats extras. The research population and its bar clock are established in historical data and data bars.

The hypothesis determines the feature family

The package provides 117 indicator exports and 120 feature-helper exports. Those counts include related calculations and utilities. The research decision concerns the information represented by a feature family and the observations required to compute it.

Bar-derived flow measures describe information recoverable from bars. An order-book hypothesis requires its own observed inputs. The indicator and feature references own the individual input and output contracts.

Python preparation states the calculation directly

The introductory Python experiment uses one decimal-valued momentum feature. Its denominator is the close window observations earlier. The same function prepares research observations and later inference history:

1def features(bars, window):2    return bars.with_columns(3        (pl.col("close") / pl.col("close").shift(window) - 1).alias("momentum")4    )

A window of four quarter-hour bars represents a different history from four activity bars. The resulting feature also differs in units from Limen's percent-valued roc. Feature names and units therefore remain part of the experiment definition.

Compiled preparation declares the feature dependencies

The composable Compiled experiment calculates its ROC indicator before these lagged and fractionally differentiated features. Group names allow preparation to admit the declared momentum and price families:

1    features:2      - func: limen.features.lag_range3        params:4          col: roc_15          start: 16          end: 37          group: momentum8      - func: limen.features.fractional_diff9        params:10          d: 0.411          cols: [close]12          threshold: 0.0513          group: price

The lag range consumes an existing roc_1 column. Fractional differentiation consumes close prices and retains its declared order and truncation threshold. Changing those choices changes the model input and its warm-up population. The fractional differentiation reference develops training-only order selection and the resulting columns.

Native preparation preserves feature order, excludes declared target-source columns, and retains fitted transformation state for inference. Training-fitted scaling and principal component analysis are developed in robustness research.

Stateful regimes retain a chronological policy

The seven-label market-regime classifier provides batch and streaming evaluation through the same implementation. Each series retains independent rolling history, prior-only ranks, confirmation counts, and breakout state. Its JSON snapshot preserves the bounded state required to continue a stream under the same configuration and implementation.

The Python helper campaign applies an explicit configuration version to the recorded source:

1def regime_labels(raw):2    from limen.features.classify_market_regime import RegimeConfig3    bars = raw.with_columns(pl.lit('recorded-spot').alias('series_id'))4    return features.classify_market_regime(bars, config=RegimeConfig(version='baseline-v1'), include_audit=True)

The returned table joins to source bars by series_id and UTC datetime. Initialization and unavailable observations remain explicit. Trend, range, and breakout labels describe the configured rule policy. The stress field measures ranked volatility conditions. Validation of those hypotheses belongs to the research comparison.

Compiled helper_regimes.yaml exercises the separate categorical regime-feature helpers declared in that file. The stateful regime contract defines the seven-label API, its streaming state, and its output fields.

Availability determines admission into inference

A current-bar close becomes available when that bar completes. Rolling features also require preceding completed observations. Training and inference must agree on both the numerical calculation and the history presented to it.

The line helpers distinguish completed-line information from retrospective active-span counts. include_research_only=False excludes those active-span columns from grouped line features. The crash-reversal helper has a separate next-row date dependency under its default structural rule. Those contracts determine the appropriate decision clock; splitting observations alone does not establish within-split causality.

The recorded helper campaign reports the calculations exercised and retains their output columns. These counts describe diagnostic coverage on the supplied observations:

The helper tables and retained files support inspection of warm-up, missing values, and added columns. A feature comparison additionally needs a common target, assessment population, and selection rule. Continue to target design to define the outcome learned from those inputs.