Factor catalogue

This page lists every factor shipped in the curated mlquant.features library, with a one-line rationale and a pointer to its implementation.

The full feature set comprises 213 factors: 9 curated Alpha101 formulas (features.alpha101) plus 204 hand-crafted legacy factors (features.legacy_factors) organised into nine thematic families. All factors are mask-contract compliant and GPU-vectorised.

The legacy factors are ported from legacy/features/Feature.py into modern tensor-based modules, auto-registered into LEGACY_REGISTRY and accessible via compute_legacy_set().

Tensor primitives

All factors are built from these primitives in mlquant.features.tensor_factors. Every primitive is mask-aware, GPU-vectorised, and has a unit test cross-checking it against the pandas reference.

Primitive	Signature	Notes
cs_rank	(x, mask) -> (rank, mask)	percentile rank in (0, 1], ties averaged
cs_zscore	(x, mask) -> (z, mask)	per-date standardisation
ts_sum, ts_mean, ts_std	(x, mask, w) -> ...	trailing window
ts_min, ts_max, ts_rank	(x, mask, w) -> ...	trailing window
ts_corr, ts_cov	(x, y, mask, w) -> ...	rolling pairwise
ewma	(x, mask, alpha) -> ...	float64 recurrence
delay, delta	(x, mask, k) -> ...	cross-time arithmetic

Curated alpha set

Name	Idea	Module
alpha001	momentum (rank of ts_argmax(close, 5))	alpha101.alpha_001
alpha002	volume-vs-intraday-return reversal	alpha101.alpha_002
alpha003	open / volume rank divergence	alpha101.alpha_003
alpha004	low-quantile mean reversion	alpha101.alpha_004
alpha006	open / volume rolling correlation	alpha101.alpha_006
alpha007	20-day deviation from mean, z-scored	alpha101.alpha_007
alpha012	sign(Δvol) × −Δclose	alpha101.alpha_012
alpha053	9-day Δ(close-location)	alpha101.alpha_053
alpha101	intraday close location within range	alpha101.alpha_101

Legacy factor zoo (204 factors)

All factors from the original Feature.py are ported into modular files under src/mlquant/features/:

Module	Family	Count	Description
_factors_better.py	better_*	28	Momentum/reversal variants
_factors_best.py	best_*	21	Close-location momentum
_factors_old.py	old_*	50	Classic alpha signals
_factors_stock.py	stock_*	22	Per-stock derived series
_factors_extra.py	extra_*	14	Turnover + amount features
_factors_add.py	add_*	30	Additional factor variants
_factors_change.py	change_*	5	Short-window velocity changes
_factors_original.py	original_*	28	Direct close/volume stats
_factors_market.py	cs_rank_*	6	Market breadth signals

Usage:

from mlquant.features import compute_legacy_set, LEGACY_REGISTRY

# Compute all 204 factors
factors, mask, names = compute_legacy_set(panel)

# Compute a subset
factors, mask, names = compute_legacy_set(panel, names=("best_001", "old_027"))

Adding a new factor

1. Implement it on top of the tensor primitives — no Python loops over stocks or dates.
2. For Alpha101-style: decorate with @_register("alphaXXX") in mlquant.features.alpha101.
3. For legacy-style: decorate with @register_legacy_factor("name") in the appropriate _factors_*.py module.
4. Add a one-liner to this file describing the rationale.
5. Add a smoke test (both test_alpha101.py and test_tensor_factors.py iterate their registries, so registered factors are auto-tested for shape / finiteness; deeper tests welcome).