Skip to content

ML Feature Engineering

QuantWave ships production-ready ML feature extractors on top of the Ehlers DSP and regime-detection stack. Features follow the same Universal Indicator pattern as classic TA: batch (Polars) and streaming (Next<T>) paths are bit-identical.

Quick start (Polars)

import polars as pl

df = pl.read_parquet("ohlcv.parquet").lazy()

# Ehlers feature columns (struct-friendly outputs)
df = df.with_columns(
    pl.col("close").ta.cybercycle(30).alias("cybercycle"),
    pl.col("close").ta.trendflex(30).alias("trendflex"),
    pl.col("close").ta.instantaneous_trendline(30).alias("itl"),
    pl.col("close").ta.hurst_exponent(100).alias("hurst"),
)

See the Indicator Gallery for Ehlers DSP and regime entries, and per-indicator guides under ML Features in the native index.

Discovery & metadata

import quantwave as qw

# Browse by category
for cat in qw.categories():
    if "ehlers" in cat.lower() or "regime" in cat.lower() or "ml" in cat.lower():
        print(cat, qw.category(cat)[:5], "...")

meta = qw.metadata("hurst_exponent")
print(meta.data_inputs, meta.warmup_bars, meta.category)

info = qw.boundary_info("cybercycle")
print(info.warmup_behavior)

Streaming (live features)

import quantwave as qw

cls = qw.streaming_class("rsi")  # classic example; same pattern for DSP features
wrapped = qw.wrap_streaming(cls(14), name="rsi")

for price in live_closes:
    val = wrapped.next(price)
    if wrapped.is_ready:
        ...  # feed to model / strategy

Use qw.track_streaming(inst, name="hurst_exponent") for warmup-aware readiness.

Feature matrix for ML pipelines

import quantwave as qw

df = qw.build_feature_matrix(ohlcv_df, features="recommended")
# Or drop warmup rows before training:
df = qw.build_feature_matrix(ohlcv_df, drop_warmup=True, warmup_bars=100)

Polars (Rust)

# lf.ta().features().recommended_matrix()  # via quantwave-polars

Build a wide matrix without lookahead:

  1. Use qw.build_feature_matrix() or .ta.features.recommended_matrix().
  2. Call qw.warmup_bars(name, params) per column and mask or drop leading rows.
  3. Join regime labels (market_state, HMM outputs) on the same bar index.
  4. Assert batch/streaming parity in CI with qw.assert_parity(...).

End-to-end: features → backtest

The canonical cross-epic notebook wires Hurst, CyberCycle, dominant cycle, and regime features into a realistic strategy with rich trade metadata and verified batch/streaming parity:

For backtest engine basics, see Backtest Quickstart.

Sources

  • Ehlers DSP papers in references/Ehlers Papers/implemented/
  • Regime modules: quantwave-core/src/regimes/
  • Feature Polars surface: quantwave-polars/src/features.rs
  • Stability validation: docs/examples/notebooks/ml_feature_stability.md