Features

Features sit one layer above indicators. They combine, re-express, lag, or contextualize price and volume information into model-ready signals or regime flags.

This page covers feature-helper selection, input requirements, and added columns. For lower-level signal primitives, see Indicators.

Conventions

• Feature helpers append columns and return the original frame with the extra outputs attached.
• Helpers either operate on plain kline data or require trade-derived columns such as maker_ratio, no_of_trades, price, or quantity.
• Not every helper is intended as a final predictor. Some are utilities used to build targets or wider feature families.
• Regime helpers output compact categorical-style columns such as regime_ma_slope or regime_price_band.
• find_min_d returns a scalar value instead of a frame; it searches for a stationarity-preserving fractional-differentiation order.

Quick example

from limen.data import HistoricalData
from limen.experiment import MLManifest
from limen.features import kline_imbalance, vwap
from limen.indicators import atr, roc
from limen.sfd.reference_architecture import logreg_binary


def manifest():
    return (
        MLManifest()
        .set_data_source(
            method=HistoricalData.get_spot_klines,
            params={'kline_size': 3600, 'start_date_limit': '2025-01-01'},
        )
        .set_test_data_source(
            method=HistoricalData.get_spot_klines,
            params={'kline_size': 7200, 'row_count_limit': 5000},
        )
        .set_split_config(8, 1, 2)
        .add_indicator(roc, period='roc_period')
        .add_indicator(atr, period=14)
        .add_feature(vwap)
        .add_feature(kline_imbalance, window='imbalance_window')
        .with_reference_architecture(logreg_binary)
    )

Kline-derived position and volatility features

These helpers work directly on bar data and require only kline-style inputs.

Function	Adds by default	Notes
atr_percent_sma	atr_percent_sma	ATR scaled by close price using SMA smoothing.
atr_sma	atr_sma	SMA-smoothed ATR variant.
close_position	close_position	Close location inside the current bar's high-low range. Accepts window; 1 preserves single-bar behavior.
close_position_rolling	close_position_rolling	Rolling mean of close location inside the high-low range.
close_ma_distance_atr	close_ma_distance_atr	Close-to-SMA distance normalized by SMA-smoothed true range.
distance_from_ma	distance_from_ma	Close distance from its rolling moving average.
distance_from_high	distance_from_high	Distance from a rolling high.
distance_from_low	distance_from_low	Distance from a rolling low.
gap_high	gap_high	Current high relative to the previous close.
kaufman_efficiency_ratio	kaufman_efficiency_ratio	Directional displacement divided by the rolling path length.
narrow_range	narrow_range	Current range divided by trailing maximum range.
price_range_position	price_range_position	Rolling range position over a wider window.
range_pct	range_pct	Current bar range as a percentage.
stochastic_k_abs	stochastic_k_abs	Absolute distance between stochastic %K and 0.5.
trend_strength	trend_strength	Fast-versus-slow trend strength summary.
volume_to_range	volume_to_range	Rolling mean volume per unit of high-low range.
volume_regime	volume_regime	Volume context over a lookback window.
vwap	vwap	Requires a datetime-like datetime column because VWAP resets by trading day.
wick_proportion	wick_proportion	Rolling mean wick share of full candle range.

Calendar and cyclical time features

These helpers derive time-of-bar context from datetime without depending on earlier indicator columns.

Function	Adds by default	Notes
calendar_time_features	hour, minute, weekday, day_of_month, day_of_year, week_of_year, month, quarter, half_of_year, is_weekend	Adds discrete calendar fields for downstream splits, filters, and rules. weekday uses ISO numbering (Monday=1 to Sunday=7), and week_of_year uses ISO week numbering.
cyclical_time_features	hour_sin, hour_cos, minute_sin, minute_cos, weekday_sin, weekday_cos, day_of_month_sin, day_of_month_cos, day_of_year_sin, day_of_year_cos, week_of_year_sin, week_of_year_cos, month_sin, month_cos, quarter_sin, quarter_cos	Encodes cyclical calendar fields without introducing artificial ordinal jumps. Uses the same ISO conventions as calendar_time_features; weekday cycles are phase-aligned with weekday - 1 before applying sine/cosine.
is_funding_hour	is_funding_hour	Parameterized funding-cadence hour indicator; default hours are 0, 8, and 16.
time_to_funding	hours_to_funding	Continuous hours until the next funding settlement; default cadence is every 8 hours from 0 UTC, and zero at a settlement bar.
is_us_open_hour	is_us_open_hour	Parameterized US open-hour indicator; default hour is 14.

Range-based volatility features

These helpers estimate volatility directly from OHLC structure instead of relying only on close-to-close returns.

Function	Adds by default	Notes
parkinson_volatility	parkinson_volatility	High-low range estimator that ignores close-to-close drift.
parkinson_vol_of_vol	parkinson_vol_of_vol	Rolling standard deviation of Parkinson variance.
garman_klass_volatility	garman_klass_volatility	OHLC estimator using open, high, low, and close information.
rogers_satchell_volatility	rogers_satchell_volatility	Drift-robust OHLC volatility estimator.
volatility_ratio	volatility_ratio	Short rolling Parkinson variance divided by long rolling Parkinson variance.
volatility_spike	volatility_spike	Current Parkinson variance divided by its fixed-lag value.
yang_zhang_volatility	yang_zhang_volatility	Combines overnight, open-close, and range information into a higher-fidelity volatility estimate. Requires window > 1.

Liquidity and impact features

These helpers translate ordinary OHLCV bars into liquidity, impact, and slippage proxies.

Function	Adds by default	Notes
dollar_volume	dollar_volume	Price-times-volume activity proxy using close and volume.
amihud_illiquidity	amihud_illiquidity	Absolute return per dollar of volume, a compact price-impact proxy.
volume_ratio	volume_ratio	Volume relative to its simple-moving-average baseline.
return_per_dollar_volume	return_per_dollar_volume	Signed return per dollar of volume for directional impact analysis.
range_per_dollar_volume	range_per_dollar_volume	Bar range scaled by dollar volume.
illiquidity_shock	illiquidity_shock	Current Amihud-style illiquidity relative to its own trailing mean.
liquidity_drop	liquidity_drop	Current LOB liquidity divided by fixed-lag LOB liquidity.
liquidity_range	liquidity_range	Rolling mean high-liquidity to low-liquidity ratio.
maker_liquidity_share	maker_liquidity_share	Maker liquidity divided by total liquidity.
maker_volume_share	maker_volume_share	Maker volume divided by total volume.
maker_volume_ratio	maker_volume_ratio	Rolling mean maker-volume share.
taker_imbalance_ratio	taker_imbalance_ratio	Rolling mean absolute taker imbalance as a share of volume.
trade_density	trade_density	Rolling mean number of trades per unit of volume.
trade_imbalance	trade_imbalance	Rolling maker volume divided by rolling total volume.
trade_size_ratio	trade_size_ratio	Short average trade size divided by long average trade size.
bulk_volume_classification	bvc_buy_volume, bvc_sell_volume	Splits bar volume into buy and sell by bulk-volume classification (standardized return through the normal CDF). Needs only close and volume.
order_flow_imbalance	order_flow_imbalance plus bvc_buy_volume, bvc_sell_volume	Rolling net BVC-classified signed flow as a share of volume; a bar-level order-flow imbalance proxy, not level-2 OFI.
vpin	vpin plus bvc_buy_volume, bvc_sell_volume	Volume-synchronized Probability of Informed Trading from the BVC buy/sell imbalance; a flow-toxicity gauge. Feed volume bars for canonical equal-volume buckets.

Dollar-bar crash reversal

dollar_bar_crash_reversal is the primary signal used by the bundled rule-based SFD of the same name. It requires a UTC-sorted datetime plus open, close, liquidity_sum, and maker_liquidity, and appends the Int8 column dollar_bar_crash_reversal_position.

For row t, the transform:

1. finds the latest open at or before t - 4h by backward as-of join and computes log(close_t / reference_open) * 10_000
2. computes maker flow as 1 - 2 * maker_liquidity / liquidity_sum only where liquidity is finite and positive
3. standardizes flow against causal 30-day rolling medians of flow and absolute deviation, both closed on the left with at least 100 observations
4. triggers when momentum is at or below momentum_threshold_bps and the robust flow score is above flow_z_threshold
5. holds the trigger active for hold_minutes of wall-clock time

The structural core permits a new trigger only when the next row belongs to the same UTC date. Therefore the last row of each UTC day, including the final row in the input, cannot initiate a trigger. A position initiated earlier may remain active there until its wall-clock hold expires. This is a one-row availability boundary: the exact research trigger is not same-row causal. The built-in backtest's one-bar execution lag is an execution adaptation, not a claim that the raw trigger was knowable on row t.

The hold is time-based rather than bar-count-based. Dollar bars arrive irregularly, so the physical span represented by a 60-minute hold can exceed 60 minutes between observed execution rows.

Realized risk and tail features

These helpers describe the quality of recent movement, not just its level.

Function	Adds by default	Notes
realized_semivariance	upside_semivariance, downside_semivariance	Splits rolling squared returns into upside and downside components.
downside_volatility_ratio	downside_volatility_ratio	Rolling downside squared-return share of total squared returns.
realized_skewness	realized_skewness	Rolling skewness of close-to-close returns.
realized_kurtosis	realized_kurtosis	Rolling kurtosis of close-to-close returns.
jump_variation_proxy	jump_variation_proxy	Positive gap between realized variance and bipower variation proxy.
tail_event_intensity	tail_event_intensity	Share of recent bars whose absolute return exceeds a configurable threshold.
volatility_of_volatility	volatility_of_volatility	Rolling variability of rolling close-to-close return volatility.
return_autocorrelation	return_autocorrelation	Rolling correlation between returns and one-bar lagged returns.
return_volatility_correlation	return_volatility_correlation	Rolling correlation between returns and Parkinson variance.
volume_volatility_correlation	volume_volatility_correlation	Rolling correlation between volume and Parkinson variance.

Seasonality-normalized features

These helpers compare current bar behavior to the trailing mean for the same hour of the week.

Custom SFD authors must keep seasonality baselines causal. Split-wide normalization such as .mean().over(['season_weekday', 'season_hour']), or any full-split hour-of-week baseline computed in one pass, leaks future information within the split. Use trailing-only or train-fitted seasonality baselines instead. The built-in relative_volume_seasonality, relative_range_seasonality, and relative_volatility_seasonality helpers use trailing hour-of-week baselines.

Function	Adds by default	Notes
relative_volume_seasonality	relative_volume_seasonality	Current volume relative to the trailing baseline for the same hour-of-week bucket.
relative_range_seasonality	relative_range_seasonality	Current range percentage relative to the trailing hour-of-week baseline.
relative_volatility_seasonality	relative_volatility_seasonality	Current absolute return magnitude relative to the trailing hour-of-week baseline.

Candle structure and auction features

These helpers focus on how a bar moved internally, not just where it finished.

Function	Adds by default	Notes
body_to_range	body_to_range	Absolute candle body size divided by the full bar range.
wick_imbalance	wick_imbalance	Upper-wick minus lower-wick imbalance as a share of full range.
range_overlap	range_overlap	Overlap share between the current bar range and the previous bar range.
rejection_intensity	rejection_intensity	Wick-heavy rejection proxy based on total wick share and directional close location.
absorption_intensity	absorption_intensity	High-volume, small-body absorption proxy using a trailing shifted volume baseline.

Cross-timescale context features

These helpers summarize cross-horizon agreement or disagreement on market state.

Function	Adds by default	Notes
trend_coherence	trend_coherence	Average sign agreement across short, medium, and long return horizons.
volatility_term_structure	volatility_term_structure	Average ratio between short, medium, and long rolling volatility estimates.
sma_ratios	<price>_sma_<period> and sma_<period>_ratio per configured period	Price-to-SMA ratios across multiple horizons, keeping the SMA columns.

Breakout and regime features

These helpers provide state or structure rather than only a continuous numeric series.

Function	Adds by default	Notes
breakout_features	lagged breakout columns plus long_roll_mean, long_roll_std, short_roll_mean, short_roll_std, roc_long_12_1, roc_short_12_1	Enriches pre-existing breakout flags.
breakout_percentile_regime	price_range_position, regime_breakout_pct	Uses percentile thresholds over price-range position.
hh_hl_structure_regime	regime_hh_hl	Captures higher-high and higher-low style structure.
ichimoku_cloud	tenkan, kijun, senkou_a, senkou_b, chikou	Full Ichimoku feature set.
ma_slope_regime	regime_ma_slope	Regime label based on moving-average slope.
price_vs_band_regime	regime_price_band	Uses price distance relative to a band definition.
sma_crossover	crossover, signal	Compact crossover-state helper.
window_return_regime	ret_24, regime_window_return	Return plus regime thresholding over a window.

Line-based context features

These helpers summarize how recently and how densely price interacted with detected price lines — pairs of bars at most max_duration_hours apart whose close-to-close change is at least min_height_pct (positive lines are long, negative are short; quantile lines are those at or above the quantile_threshold height quantile per direction).

The two grouped transforms below are the YAML-facing surface. Each detects lines internally from scalar params on the frame it receives — per split under the manifest pipeline, so detection cannot observe other splits — and adds its full column family in one detection pass. Line detection lives in limen.utils.find_price_lines / limen.utils.filter_lines_by_quantile.

Function	Adds by default	Notes
price_lines	active_lines, hours_since_big_move, line_momentum_<m>h, trending_score, reversal_potential	All-line family: span count, end recency (capped at big_move_lookback_hours), and long-minus-short end counts over the trailing [t-m, t) window with their balance (trending_score, in [-1, 1]) and min/max ratio (reversal_potential, in [0, 1]).
quantile_price_lines	hours_since_quantile_line, active_quantile_count, quantile_line_density_<d>h, quantile_momentum_<m>h, avg_quantile_height_<h>h, quantile_direction_bias	Quantile-line family: end recency and span count, end density over density_lookback_hours, signed height sum over ends in [t-m, t], and mean height plus height-weighted direction (in [-1, 1]) over ends in [t-h, t].

active_lines and active_quantile_count count lines that span the current bar before the line's end — the event that defines it — is knowable. That within-line lookahead is inherited from the tradeline research design: treat both columns as research-only, not live-computable. The grouped transforms expose include_research_only; set it to false to omit the active-span columns. Bundled live-safe templates set include_research_only: false. The end-event columns are causal.

The per-column building blocks below take pre-computed line structures (list[dict] with start_idx/end_idx) and remain available for programmatic composition; the grouped transforms above compose them.

Function	Adds by default	Notes
active_lines	active_lines	Count of long and short lines active at each bar.
active_quantile_count	active_quantile_count	Active-line count restricted to quantile-filtered lines.
quantile_line_density	quantile_line_density_<lookback>h	Count of quantile-line endings within a trailing lookback_hours window.
hours_since_big_move	hours_since_big_move	Bars since the most recent line end, capped at lookback_hours.
hours_since_quantile_line	hours_since_quantile_line	Bars since the most recent quantile-line end, capped at lookback_hours.

Lag helpers and threshold utilities

These helpers expand existing columns or define cutoffs for target construction.

Function	Adds or returns	Notes
lag_column	one lagged column such as close_lag_2	Requires col and lag.
lag_columns	one lag per listed column	Requires cols and lag.
lag_range	a lag range such as close_lag_1 through close_lag_3	Requires col, start, and end.
lag_range_cols	a lag range for each listed column	Requires cols, start, and end.
rolling_zscore	configurable *_zscore_* column	Applies identity, log1p, or abs before rolling z-score standardization.
cusum_filter	cusum_event	Int8 flag of symmetric CUSUM events on the close log-return path (1 up, -1 down, 0 none); gates which moves are worth sampling.

Stationarity and long-memory helpers

These helpers reduce non-stationarity while preserving more long-memory structure than a first difference would.

Function	Adds or returns	Notes
fractional_diff	one *_fracdiff column per selected input column	Applies fixed-width fractional differentiation. Original columns are preserved.
find_min_d	scalar d value	Iterates over candidate orders and uses the Augmented Dickey-Fuller test to find the smallest stationary order.

Two practical details matter:

• fractional_diff needs cols=['close'] and writes new columns such as close_fracdiff.
• if one split is too short to produce the same fractional-diff column as another split, Manifest now drops that extra column during split alignment so the final data_dict stays consistent.

Trade-shape and microstructure features

These helpers need richer data than ordinary OHLCV bars.

Function	Adds by default	Notes
kline_imbalance	imbalance	Requires maker_ratio and no_of_trades from data retrieval or bar formation.
conserved_flux_renormalization	synthetic OHLCV plus value_sum, vwap, flux_rel_std_mean, flux_rel_std_var, entropy_mean, entropy_var, Δflux_rms, Δentropy_rms	Works on trade-level datetime, price, and quantity, then rolls those into kline-aligned diagnostics.

Dynamic-target and entry-score features

This family builds volatility-conditioned targets, stops, and regime weights, plus the microstructure entry score they combine with. The helpers compose: volatility_measure and regime_multiplier feed dynamic_target and dynamic_stop_loss, the momentum and candle-position helpers feed entry_score_microstructure, and feature_aliases snapshots the family into *_feature columns with nulls filled.

Function	Adds by default	Notes
close_to_extremes	close_to_high, close_to_low	Close position relative to bar high and low extremes.
dynamic_stop_loss	dynamic_stop_loss	Volatility- and regime-conditioned stop-loss level.
dynamic_target	dynamic_target	Volatility- and regime-conditioned target level.
ema_alignment	ema, ema_alignment	EMA alignment score with power transformation.
entry_score_microstructure	entry_score, entry_score_base	Microstructure timing score from momentum, spread, candle position, and volume spikes.
feature_aliases	dynamic_target_feature, entry_score_feature, momentum_score_feature, regime_high_feature, regime_low_feature, regime_normal_feature, vol_60h_feature, vol_percentile_feature	Null-filled aliases snapshotting the family for model consumption.
log_returns	log_returns	Logarithmic returns of the close series.
market_regime	sma_20, sma_50, trend_strength, volatility_ratio, volume_sma, volume_regime, market_favorable	Trend-strength and volume-regime favorability score.
micro_momentum	micro_momentum	Short-horizon price momentum.
momentum_confirmation	momentum_score	Momentum confirmation score from recent price changes.
momentum_periods	momentum_<period> per configured period	Momentum over multiple horizons.
momentum_weight	momentum_weight	Momentum-direction weighting factor.
position_in_candle	position_in_candle	Close position within the bar high-low range.
position_in_range	position_in_range	Close position within the bar high-low range over a rolling window.
regime_multiplier	regime_multiplier	Volatility-regime multiplier for dynamic parameter adjustment.
returns_lags	returns_lag_<lag> per configured lag	Lagged simple returns.
spread	spread	High-low range normalized by close (same formula as range_pct, but stored in a spread column).
spread_percent	spread_percent	High-low range normalized by close, stored as spread_percent for microstructure scoring.
volatility_1h	volatility_1h	Alias of an existing volatility column at the one-hour horizon.
volatility_measure	volatility_measure	Combined rolling-volatility and ATR-percentage measure.
volatility_weight	volatility, volatility_weight	Inverse-volatility weighting factor.
volume_spike	volume_spike	Volume relative to a rolling-statistics baseline.
volume_trend	volume_trend	Short-term versus long-term volume average trend.

Choosing between indicators and features

• Use an indicator for a direct market calculation such as RSI, ATR, or MACD.
• Use a feature for structure around those signals, such as lags, regimes, relative position, or multi-step aggregation.
• Use the lag helpers when the main value is temporal context rather than a new market calculation.
• Use fractional_diff when stationarity itself is part of the design problem, not just a preprocessing afterthought.

Read next

• Indicators for lower-level signal primitives
• Transforms for stateless target shaping and cleanup helpers
• Calibration for post-model probability calibration and threshold optimization
• Experiment Manifest for how features plug into the split-first manifest pipeline
• Utilities for the exported adf_test() helper that find_min_d builds on