Information coefficient vs Fama-MacBeth

This page is task-oriented: assume you already know factrix exists and you need to pick a factory. For the canonical 5-scenario table (research question → factory → procedure → literature) see Concepts § Five analysis scenarios.

Factory methods are type-safe constructors. Unsupported combinations (e.g. metric=IC on a sparse signal) are caught by the IDE before runtime — no need to memorise the legal axis triples.

Information coefficient (IC) vs FM

Both apply to (INDIVIDUAL, CONTINUOUS). Choose by research question:

	IC	FM
Question	Predictive rank ordering?	Unit-exposure return premium?
Method	Spearman ρ per date → Newey-West (NW) heteroskedasticity-and-autocorrelation-consistent (HAC) t on E[information coefficient (IC)]	ordinary least squares (OLS) slope λ per date → NW HAC t on E[λ]
Robust to	Outliers (rank-based)	Proportional exposure differences
Economic interpretation	Directional signal quality	Premium per unit of factor exposure
n_assets sensitivity	Drops dates with < 10 assets	Runs at N ≥ 3 but unstable at low N

Use IC when you care about rank ordering (stock selection). Use FM when you need an economically interpretable premium estimate (risk premia, factor pricing).

For the lookup table — which metrics are supported under which (scope, signal) cell, with sample-size floors and warning codes — see Reference § Metric applicability.

Standalone metrics vs evaluate()

evaluate() runs the canonical inference procedure for a cell. Standalone metrics in factrix.metrics provide supplementary diagnostics without a formal PASS/FAIL outcome:

When to use evaluate()	When to use standalone metrics
Canonical signal validity inference	Diagnose shape, asymmetry, regime splits
Benjamini-Hochberg-Yekutieli (BHY) family input (needs FactorProfile)	Multi-statistic decomposition
Primary screening gate	out-of-sample (OOS) decay, tradability, concentration

See Metric pipelines for the full module list.