Skip to content

Glossary

factrix names its dispatch axes after the structure of the data (Scope × FactorDensity) and the regression layout (DataStructure). The literature this borrows from uses overlapping but not identical vocabulary — this glossary maps the factrix terms to their nearest industry equivalents and flags collisions where the same word means different things in factrix vs Alphalens / Barra / standard panel-econometrics.

Scope axis — INDIVIDUAL vs COMMON

INDIVIDUAL

Per-asset factor values: each (date, asset_id) pair has its own factor reading. The cross-section is the unit of inference (per-date information coefficient (IC), FM λ, CAAR per event date). This is the regime most equity factor research operates in.

Industry equivalents:

  • Alphalens: "characteristic" / "factor". Alphalens panels are always per-asset signals, so the term is implicit.
  • Standard cross-sectional asset pricing: "stock characteristic" (size, book-to-market, momentum). Same regime; factrix routes these through Individual × Continuous.
  • MSCI / Axioma: "factor exposure" of asset i to factor k.

COMMON

A factor that does not vary across assets: on any given date every asset_id carries the same factor value. The panel layout is unchanged (still long-format (date, asset_id) rows) — scope is a factor attribute, not a data shape. Macro factors (VIX, USD index, term spread, monetary-policy shocks) and broadcast event dummies (FOMC announcement days) live here.

Industry equivalents and collisions:

  • Macro / global factor literature: "common factor" = a single series that drives multiple assets. Matches factrix COMMON.
  • Barra "common factor": a shared risk factor estimated from cross-sectional regressions on a panel of asset characteristics (Barra style / size / value / momentum factors). Estimation-driven, per-asset latent loadings. Collides with factrix COMMON, which is a given broadcast series with no estimation step on the factor itself.
  • Alphalens: no direct equivalent. Closest is "group-neutral" / benchmark series, but Alphalens does not natively dispatch to a COMMON cell.

When porting Barra terminology in: a Barra "common factor" with raw cross-sectional loadings is a factrix INDIVIDUAL × CONTINUOUS problem (the loadings are the per-asset characteristic). A Barra "common factor" treated as a single series shared across assets is a factrix COMMON × CONTINUOUS problem.

FactorDensity axis — CONTINUOUS vs SPARSE

CONTINUOUS

Factor takes a real-valued reading on every observation in the panel. Most equity characteristics; macro factors; ML predictor scores.

SPARSE

Factor is non-zero only on event dates and zero elsewhere; the panel encodes a discrete event arrival process. The general schema is {0, R} where R is any real value (positive, negative, or any magnitude); the simplest form is {0, 1} for a pure event flag. Null factor cells are missing values, not non-events; fill them to 0 only when that is the intended event contract. The automatic SPARSE detector is zero-ratio based, but an explicit sparse metric can still run below the routing threshold with a frequent_event_signal warning when zero non-event rows are present. See compute_caar's input-form table for the resulting estimator distinction.

Industry equivalents:

  • MacKinlay (1997) event-study vocabulary: factrix's SPARSE cell is the event-study cell; the factor column is the announcement indicator. factrix's factor != 0 filter is the event-window selector.
  • Alphalens: no direct equivalent — Alphalens assumes continuous characteristics. Event-study workflows live outside Alphalens.

DataStructure axis — PANEL vs TIMESERIES

DataStructure is derived from data at evaluate-time, not configured:

  • n_assets >= 2 → PANEL.
  • n_assets == 1 → TIMESERIES.

PANEL

Multi-asset panel; cross-section participates in inference (per-date aggregation, cross-asset SE).

Industry equivalents: panel data; longitudinal data; pooled cross-section / time-series. The econometrics literature usually says "panel" with no further qualification.

TIMESERIES

Single-asset series. Time axis is the only sample axis; cross-section is degenerate.

Industry equivalents: time-series regression; single-asset ordinary least squares (OLS). Note that factrix TIMESERIES mode does not mean Alphalens' "time-series of cross-sectional means" (that is per-date aggregation of a PANEL); the two collide on the word time-series and disagree on what is aggregated. See Timeseries-mode conventions for the specific contracts that apply when factrix routes here.

Other terms

mainstream metric

The headline mean-significance test conventionally used for a (scope, density, structure) cell — ic, fm_beta, caar, common_beta, or predictive_beta for the single-asset dense TIMESERIES cell. It is an authoring convention, not a code-enforced tier: evaluate() runs exactly the metrics the caller passes, and every metric (mainstream or supplementary) carries its own MetricResult.p_value. By convention the mainstream metric's p_value is the one the FDR screening verbs read for a cell. Other metrics applicable to the same cell are supplementary / diagnostic and are run alongside it via list_metrics. For practical selection guidance, see Choosing a metric: First-pass metrics vs diagnostics.

factor

The signal column. Same role as Alphalens "factor", Barra "exposure", or generic "alpha score" / "predictor". The procedure-level INPUT_SCHEMA consumes a column literally named factor; pass evaluate(..., factor_cols=["alpha"]) to use a differently-named column without renaming the panel first (Data schema).

forward_return

The forward return column consumed by every dispatch cell. factrix's compute_forward_return(df, forward_periods=forward_periods) produces a per-period return with t+1 entry:

forward_return[t] = (price[t+1+forward_periods] / price[t+1] − 1) / forward_periods

Two non-textbook choices to internalize:

  • t+1 entry, not t: signal at t is computed from data up to and including price[t], so trading at price[t] would assume same-bar execution. Entry at t+1 enforces a causal boundary and cleanly separates the return window from the BMP estimation window.
  • Divided by forward_periods: result is expressed per period rather than as the cumulative forward_periods-period return, so factor evaluations at different forward_periods are directly comparable. If you need the cumulative forward_periods-period number, multiply by forward_periods.

Only simple returns are implemented; no log-return option. factrix takes forward_return as input rather than computing it inside evaluate() — attach it with compute_forward_return before dispatch so the forward_periods horizon is explicit and aligned with the forward_periods passed to evaluate().

Distinct from "spot return" (contemporaneous one-period return) and "realised return" (ex-post return for risk attribution).

signal_horizon (datasets only)

A property of the synthetic signal embedded in make_cs_panel / make_event_panel — the horizon at which the generator's nominal IC is realised. Aligning evaluate()'s forward_periods with signal_horizon is the regime where the synthetic IC / drift is calibrated; mismatched horizons induce IC decay (synthetic) and bias (in TIMESERIES mode — see Timeseries-mode conventions § Non-overlap).

non-overlapping returns

Sampling every h-th date so consecutive observations are independent under the h-period forecasting null. factrix's Individual × Continuous cell defaults to this for ic and caar; Newey-West (NW) heteroskedasticity-and-autocorrelation-consistent (HAC) is the alternative on the full overlapping series. See Statistical methods § HAC SE.

estimation_window

Per-asset pre-event sample used to fit the abnormal-return baseline for bmp_z and corrado_rank. See Metric applicability § estimation_window.

Multiple testing

FDR — False Discovery Rate

The expected proportion of false positives among rejected nulls, in contrast to FWER (Family-Wise Error Rate, the probability of any false positive). For a screening rule that rejects R of m hypotheses and produces V false positives, FDR ≡ E[V / max(R, 1)]. factrix's primary screening primitive is BHY (Benjamini-Hochberg-Yekutieli), which controls FDR at a user-chosen q under arbitrary dependence. See Statistical methods and bhy.

BHY — Benjamini-Hochberg-Yekutieli

Step-up FDR-controlling procedure that allows arbitrary dependence between p-values — including violations of the positive regression dependence on a subset (PRDS) condition required by Benjamini-Hochberg 1995, at the cost of a log(m) factor. Mathematically implements Benjamini & Yekutieli (2001). factrix's default multiplicity-correction method because financial p-values are typically not independent. For the BH / BY / BHY naming convention, see Statistical methods § Multiple-testing under dependence.

family

The set of hypotheses jointly entered into a single FDR / FWER controlling procedure. Once a hypothesis joins a family, the procedure's FDR claim only holds within that family. Family choice is a contract decision: re-running BHY on a filtered subset of survivors does not preserve FDR ≤ q. See Cross-function reference § expand_over for the sample-restriction vs hypothesis-dimension split.

Screening result (BhyResult / PartialConjunctionResult / HierarchicalBhyResult)

Result type from the screening functions (bhy, partial_conjunction, bhy_hierarchical) — one procedure-paired dataclass each, returned keyed by metric name. Each carries the full tested family (entries) with the post-correction adjusted p-value (adj_p_all) aligned to it, exposes the surviving subset as survivors / adj_p, and preserves the survivor[i] ↔ adj_p_all[i] ≤ q duality so downstream functions (compare(result.survivors)) can render leaderboards without re-applying the threshold.