ATR-Adaptive Laguerre RSI

Non-anticipative volatility-adaptive momentum indicator for sequence-to-sequence forecasting.

Overview

This library implements the ATR-Adaptive Laguerre RSI indicator, designed for robust feature engineering in financial time series forecasting. The indicator combines:

• True Range (TR) - Volatility measurement including gaps
• ATR with Min/Max Tracking - Rolling volatility envelope
• Adaptive Coefficient - Volatility-normalized adaptation
• Laguerre 4-Stage Cascade - Low-lag smoothing filter
• Laguerre RSI - Momentum from filter stage differences

Key Features

• ✅ Non-anticipative: Guaranteed no lookahead bias
• ✅ O(1) Incremental: Efficient streaming updates with .update() method
• ✅ Multi-interval: Supports 1s-1d timeframes with 85-feature extraction (133 without filtering)
• ✅ Redundancy filtering: Optional 133→85 feature reduction (|ρ| > 0.9 removed)
• ✅ Flexible datetime: Works with DatetimeIndex, 'date' column, or custom column names
• ✅ Validated: Information coefficient > 0.03 on k-step-ahead returns

Installation

PyPI (Public)

# Core package (indicators, features, backtesting.py integration)
uv add atr-adaptive-laguerre

# With optional Binance data adapter (BinanceAdapter, Parquet caching)
uv add "atr-adaptive-laguerre[data]"

AWS CodeArtifact (EonLabs Private)

The package is also available in the EonLabs CodeArtifact repository for internal use.

For details on setting up CodeArtifact publishing and credentials, see CodeArtifact Setup Guide.

Install from CodeArtifact

# Configure CodeArtifact credentials (see CODEARTIFACT_SETUP.md)
uv add atr-adaptive-laguerre --index-url <codeartifact-url>

# With optional data adapter
uv add "atr-adaptive-laguerre[data]" --index-url <codeartifact-url>

Feature Modes: Choose Your Use Case

This package supports two operational modes with very different capabilities:

⚠️ Important: Multi-Interval Mode is Recommended

If you're building ML features, you want multi-interval mode (85 features), which includes:

• Base interval features (31)
• First multiplier interval features (31)
• Second multiplier interval features (31)
• Cross-interval analysis features (40) ← Unique to multi-interval mode!
• Redundancy filtered: 133 → 85 features (48 redundant features removed)

Cross-interval features detect multi-timeframe patterns like:

• Regime alignment across timeframes
• Divergence detection
• Momentum cascades
• Gradient analysis
• Statistical stability metrics

Quick Start

Multi-Interval Mode (Recommended - 85 Features)

from atr_adaptive_laguerre import ATRAdaptiveLaguerreRSI, ATRAdaptiveLaguerreRSIConfig

# RECOMMENDED: Use multi-interval mode for full feature set
config = ATRAdaptiveLaguerreRSIConfig.multi_interval(
    multiplier_1=4,   # 4x base interval (e.g., 2h → 8h)
    multiplier_2=12   # 12x base interval (e.g., 2h → 24h)
)
indicator = ATRAdaptiveLaguerreRSI(config)

# Extract 85 features across 3 timeframes (31 per interval + 40 cross-interval, filtered)
features_df = indicator.fit_transform_features(df)

print(f"Features extracted: {indicator.n_features}")  # 85
print(f"Min data required: {indicator.min_lookback_base_interval} bars")  # 360

Single-Interval Mode (Minimal Lookback - 31 Features)

Use this mode only if you have limited historical data or need single-timeframe analysis:

from atr_adaptive_laguerre import ATRAdaptiveLaguerreRSI, ATRAdaptiveLaguerreRSIConfig

# Single-interval mode (WARNING: only 31 features, missing cross-timeframe analysis)
config = ATRAdaptiveLaguerreRSIConfig.single_interval(
    atr_period=14,
    smoothing_period=5,
    date_column='date'  # Or use DatetimeIndex
)
indicator = ATRAdaptiveLaguerreRSI(config)

# Get single RSI value
rsi_series = indicator.fit_transform(df)  # Returns pd.Series (single RSI column)

# Or get 31 single-interval features
features_df = indicator.fit_transform_features(df)  # Returns DataFrame with 31 columns

print(f"Features extracted: {indicator.n_features}")  # 31
print(f"Min data required: {indicator.min_lookback} bars")  # 30

Advanced: Incremental Updates (O(1) Streaming)

Both modes support efficient incremental updates:

# After initial fit_transform
new_row = {'open': 100, 'high': 101, 'low': 99, 'close': 100.5, 'volume': 1000}
new_rsi = indicator.update(new_row)  # Returns float (O(1) complexity)

Disabling Redundancy Filtering (85 → 133 Features)

# Disable redundancy filtering to get all 133 features
config = ATRAdaptiveLaguerreRSIConfig.multi_interval(
    multiplier_1=4,
    multiplier_2=12,
    filter_redundancy=False  # Get all 133 features
)
feature = ATRAdaptiveLaguerreRSI(config)

# Returns DataFrame with 133 columns (all features, including 48 redundant ones)
features_df = feature.fit_transform_features(df)

# Verify feature count
print(f"Features: {feature.n_features}")  # 133 (85 by default)

# Redundancy filtering (enabled by default):
# - Data: 3 years of 2h OHLCV (BTCUSDT, ETHUSDT, SOLUSDT)
# - Threshold: |ρ| > 0.9 (perfect correlations and near-redundant features)
# - Removes: Base RSI values, redundant distance metrics, duplicate regime features
# - Retains: Rate-of-change, cross-interval, temporal features, and tail risk features
# - IC validation: Tail risk features validated on out-of-sample data (2025-10-08)

Documentation

• API Reference - Complete API documentation
• Examples - Runnable usage examples
• Changelog - Release notes and version history

License

MIT License - Eon Labs Ltd.