Skip to content

Design notes

A record of choices factrix deliberately did not implement, with the trade-off and the literature anchor that grounds each decision. Reading this page is the fastest way to understand what factrix is not: it explains the negative space around the scope so contributors do not re-litigate settled boundaries. Positive design lives in Architecture; the literature targets are in the Bibliography. For the user-facing positioning of these decisions against peer libraries, see Where factrix fits — §1 and §3.1 anchor the no-composite and IVX-flag-only choices in the §1/§7 rationale below.

The seven sections below correspond to the seven structural "considered but not done" decisions that recur in design discussions.

1. No composite factor score

factrix evaluates each factor through a list of separately reported metrics — information coefficient (IC), FM-λ, quantile spread, monotonicity, out-of-sample (OOS) decay, turnover, etc. — and renders the cell primary p-value. It deliberately does not collapse these into a single weighted score.

  • A composite score becomes a target to be optimised against. Once factor generators tune to the composite, the score stops measuring what it nominally measures — Goodhart's law in its original monetary-policy form (Goodhart 1984) and as rediscovered in the asset-pricing context.
  • Different metrics test different null hypotheses (mean, monotonicity, persistence, capacity). Folding them into one weighted scalar implicitly assigns a price to each null, which is not a problem the data can answer.
  • Naive equal weighting beats most "optimal" weight schemes when estimation error swamps optimisation gain — the same finding that motivates equal-weight portfolios over mean-variance under realistic parameter uncertainty (DeMiguel-Garlappi-Uppal 2009).
  • Per-metric pass/fail keeps the user in the inference loop. A score can hide which dimension failed; per-metric pass/fail makes it explicit.

2. No factor selection or Stage-2 ranking

factrix does not decide which factors should make it into a final multi-factor portfolio. Outputs are per-factor diagnostics; the selection step is left to the user (or to a separate stage built on factrix's outputs).

  • Single-factor quality and incremental contribution to a model are separable problems. Conflating them is what the shrinkage and zoo-taming literatures explicitly warn against.
  • Greedy forward selection is implemented as a diagnostic (greedy_forward_selection) but documented as not for inference — t-stats from greedy paths are inflated by selection, and post-selection inference requires PoSI machinery (Berk-Brown-Buja-Zhang-Zhao 2013, Leeb-Pötscher 2005) that factrix does not ship.
  • Spanning tests (Barillas-Shanken 2017) give the honest incremental-α answer for a fixed pair of factor sets. They are exposed as spanning_alpha; building a selection pipeline on top is a downstream concern.

3. No capacity / size-of-trade scoring

factrix reports notional_turnover, breakeven_cost, net_spread, and top_concentration as profile diagnostics — values the user inspects, not gates that affect the inference.

  • Capacity is a property of the portfolio implementation, not of the factor signal. A factor with concentrated long-leg α can be excellent for a $10M book and unusable for a $10B book; factrix cannot know which the user is sizing for.
  • Linear cost models (the breakeven formulation) hold only at small scale. Above a few percent of average daily volume, market impact follows a square-root law (Almgren & Chriss, "Optimal Execution of Portfolio Transactions," Journal of Risk 2001 — cited as background; not in bibliography) and the linear approximation understates cost.
  • The trading-cost-aware factor selection in DeMiguel-Martin-Utrera-Nogales-Uppal 2020 internalises cost into the selection objective. factrix exposes the inputs (net_spread) but stops short of taking the selection decision — same reason as section 2.

4. No cross-market aggregation

factrix evaluates a factor on the panel the user supplies. It does not provide a primitive for "is this factor globally pervasive across N markets" beyond what the user can run by calling evaluate() per market and comparing the returned FactorProfiles.

  • The global-vs-local pricing debate is unsettled empirically (Asness-Moskowitz-Pedersen 2013 argues for substantial commonality; the Fama-French international tests (Fama-French 1998 / 2012 / 2017 — not in bibliography) show local models price local cross-sections better). Either aggregation choice would foreclose a research path.
  • Standardisation across markets has industry conventions but no canonical academic answer. Picking one inside the library would impose a prior the user did not opt into.
  • A separate cross-market layer can be built on factrix outputs without library support: each FactorProfile.metrics is serialisable and easy to combine externally.

5. BHY rather than Bayesian multiple-testing

factrix's multi_factor.bhy runs Benjamini-Yekutieli (Benjamini-Yekutieli 2001) false discovery rate (FDR) control with the c(m) = Σ 1/i dependence correction. It does not implement Bayesian alternatives such as Harvey-Liu 2020 "Lucky Factors" or Bryzgalova-Huang-Julliard 2023 BMA stochastic discount factor (SDF) search (not in factrix's bibliography catalog).

  • BHY is a frequentist recipe with a deterministic threshold given (p, m). It composes cleanly with downstream business logic that needs a yes/no answer — the user reports bhy_passed to a governance step without negotiating priors.
  • Bayesian methods integrate dependence structure more naturally and can yield posterior probabilities of being a true factor. The cost is an explicit prior over factor returns and a sampler / variational fit per evaluation. factrix's lean-dependency policy (numpy + polars only in the hot path) does not pay this cost.
  • The case that BHY is empirically close enough is supported by Chordia, Goyal & Saretto (2020) "Anomalies and False Rejections," Review of Financial Studies 33(5) (cited as background; not in bibliography) and Jensen-Kelly-Pedersen 2023: under realistic factor pools, BHY rejects the marginally significant exactly where Bayesian methods do, and the disagreement is concentrated in the corners of the decision space where neither is decisive.
  • Users who want Bayesian inference can run it externally on the per-factor MetricOutput.stat series; factrix does not block that path.

6. Panel-aware FM rather than pooled OLS

The Individual × Continuous cell with Metric.FM runs Fama-MacBeth (Fama-MacBeth 1973) with Newey-West (NW) heteroskedasticity-and-autocorrelation-consistent (HAC) at stage 2 (Newey-West 1987), not pooled OLS with clustered SE.

  • Pooled OLS without clustering understates SE in the time-effect-dominant panels where factrix factor evaluations typically run (Petersen 2009). FM and pooled OLS with date-clustered SE are asymptotically equivalent under a balanced panel; the choice between them is about small-sample behaviour, not the basic correction. factrix defaults to FM because the cross-section-first aggregation order — per-date OLS, then NW HAC t on the λ series — is exactly identifiable in finite samples and degrades transparently when an entire date is dropped.
  • The future extension to two-way clustering (Cameron-Gelbach-Miller 2011, Thompson 2011) is acknowledged in the architecture; it is not the default because the typical factrix panel exhibits stronger date-correlation than asset-correlation, and the two-way correction adds complexity without changing the point estimate.
  • pooled_ols is exposed as an explicit comparison metric so a user can see whether FM and pooled disagree on sign or magnitude. The comparison stays in the FactorProfile.metrics payload — divergence is information for the user to interpret, not a diagnose()-level signal that triggers automatic action.

7. Per-metric registered procedures rather than a unified test

factrix dispatches each Scope × Signal × Metric cell to a registered procedure that runs a fixed pipeline (sample guards → cross-section step → significance test → diagnostics). It does not combine cell-canonical scalars into a single procedure-level F-test or χ² of the form "is anything in this profile significant?".

  • A unified composite test forces a hypothesis structure that the underlying data does not have. IC and FM-λ test the same cell with different aggregation orders; combining them double-counts the same evidence under positive correlation.
  • Event-study tests (Brown-Warner 1985, MacKinlay 1997) and cross-sectional tests have fundamentally different null hypotheses. A unified statistic that blends them would obscure which null actually rejected.
  • Registered procedures keep the per-metric audit trail readable. A reviewer can see exactly which test fired, on what subsample, with which lag rule — the same auditability principle that motivates pre-registered analysis plans more broadly (Harvey 2017).
  • The cost is that factrix produces a vector of evidence rather than one number. The user can reduce the vector to a binary at the threshold they choose, while the vector itself remains accessible for inspection.