Ml4t Feature Families

Feature Families

A model trained on six momentum variants learns one signal six ways. Diversifying across feature families - each driven by a different economic mechanism - produces more robust predictions.

The Problem

Feature sets dominated by a single family (e.g., all momentum) are highly correlated internally. The model wastes capacity learning redundant information and becomes fragile when that one mechanism stops working. A momentum crash wipes out all signal simultaneously.

The Pattern

WRONG

import polars as pl

# All momentum variants - same family, correlated, fragile
features = df.with_columns(
    mom_5d=pl.col("close").pct_change(5).over("symbol"),
    mom_21d=pl.col("close").pct_change(21).over("symbol"),
    mom_63d=pl.col("close").pct_change(63).over("symbol"),
    mom_126d=pl.col("close").pct_change(126).over("symbol"),
    mom_252d=pl.col("close").pct_change(252).over("symbol"),
)

CORRECT

import polars as pl
import numpy as np

# One representative from each family - diverse signals
features = df.sort("symbol", "timestamp").with_columns(
    # Momentum: trend-following
    momentum_63d=pl.col("close").pct_change(63).over("symbol"),
    # Mean-reversion: deviation from moving average
    mean_rev_z=(pl.col("close") - pl.col("close").rolling_mean(20).over("symbol"))
    / pl.col("close").rolling_std(20).over("symbol"),
    # Volatility: risk regime
    realized_vol=pl.col("returns").rolling_std(21).over("symbol") * np.sqrt(252),
    # Carry: yield/cost signal (example: dividend yield or funding rate)
    carry_proxy=pl.col("dividend_yield"),
    # Value: fundamental anchor
    pe_ratio=pl.col("pe_ratio"),
)

The Five Families

Diversity Diagnostic

# Check inter-family correlation - should be low
corr = features.select(feature_cols).to_pandas().corr()
avg_cross_family = corr.abs().mean().mean()  # Target: < 0.3

Guardrails

• Max 2-3 features per family in initial models - add more only if IC justifies it
• Cross-family correlation < 0.3 on average - higher means redundancy
• Each feature needs an economic hypothesis - if you cannot explain why it predicts, it may be noise
• Not all families apply to all assets: carry is irrelevant for assets without yield

Production Implementation

ml4t-engineer provides a catalog of 120+ features organized by registry category. The mapping to economic families is approximate, so carry and value signals often remain external features in research code.

from ml4t.engineer import compute_features, feature_catalog

# Browse registry categories
feature_catalog.list(category="momentum")
feature_catalog.list(category="volatility")

# Compute a diversified set from current registry names
features = compute_features(data, [
    "mom", "rsi", "realized_volatility", "garman_klass_volatility",
])

Checklist

• Features span at least 3 of the 5 families
• No single family contributes more than 40% of total features
• Cross-family correlation checked (target < 0.3)
• Each feature has a stated economic hypothesis
• Family coverage documented in feature config