Abstract

This document describes the construction of FPUS1, the US equity risk model used across FactorPulse. Each factor has its own page, listed on the FPUS1 model page. The universe is US-listed common stocks and depositary receipts. FPUS1 has 25 factors: the market, 18 orthogonalized style factors (three of which, the dynamics factors, are built from positioning and flow data), and 6 ETF-regression-based theme factors. Each security carries a vector of standardized factor exposures (loadings), and each style and theme factor is realized as a net-flat (zero-delta), cap-weighted long/short factor portfolio whose daily return is observable ex post.

We give the full construction pipeline: the estimation universe and its survivorship/point-in-time controls (§3); the treatment and standardization of descriptors and loadings (§4); the style and thematic exposure construction including the Gram–Schmidt orthogonalization that makes the factors mutually near-orthogonal (§5–§6); the pure-factor portfolio weights (§7); and the resulting factor returns, factor-return covariance, and return attribution (§8). §9 records the realized behavior of the factor portfolios over the sample.

On disclosure. Formulas are presented in full structural form. Calibration constants (composite component weights, winsorization magnitudes, and clip bounds) are shown symbolically (e.g. \(w_{\bullet}\), \(a\), \(k\)). The construction schematic (Figure 1) is illustrative; the exposure-orthogonality, factor-return, and correlation statistics in §5, §8 and §9 are measured on the production model (daily sample, Jan 2020 – Nov 2026). History before 15 July 2026 was computed retroactively with the methodology adopted on that date; it is not a live record. Later methodology changes apply from their effective date forward.

1. Model at a Glance

FPUS1 is cross-sectional and point-in-time: on each trading day \(t\) we observe a universe of \(N_t\) securities, compute an exposure matrix \(\mathbf{X}_t \in \mathbb{R}^{N_t \times K}\) over \(K\) factors, and form factor portfolios from \(\mathbf{X}_t\). Realized factor returns are measured one day forward, against day \(t{+}1\) security returns.

FamilyCountBasisMembers
Market1every security has exposure 1; the return is the square-root-cap-weighted mean return of the estimation universe market
Style18fundamental + technical descriptors, cross-sectionally standardized then Gram–Schmidt orthogonalized international, beta, size, long-term momentum, residual volatility, liquidity, growth, value, leverage, profitability, dividend yield, short interest, hedge-fund ownership, medium- / short-term / one-day momentum, short-sale activity, morning activity (hierarchy order)
  of which dynamics3style factors built from positioning and flow data short interest, hedge-fund ownership, short-sale activity
Theme6rolling time-series betas to liquid theme ETFs, standardized then orthogonalized against market beta and the earlier themes semiconductors, oil, treasury (duration), gold, bitcoin, China (hierarchy order)

Theme exposures are measured net of market beta and of the themes ahead of them, but they are not neutralized against the style factors. The theme share of cross-sectional variance (§9.5) is measured after the style factors.

Each style factor is published in two variants: a univariate raw-normalized loading (suffix _u, e.g. beta_u) and a Gram–Schmidt orthogonalized loading published under the bare factor name (e.g. beta). The contrast between them is itself informative and is discussed in §5.5 and §9.

RawdescriptorWinsorize+ clipCross-sectionalstandardizeGram–SchmidtorthogonalizeCap-weightedfactor portfolioDaily factorreturn
Figure 1. End-to-end construction of a single factor, from raw descriptor to realized daily factor return. Every style and thematic factor follows this pipeline.

Historical Factor Performance

Results first: how the orthogonalized factor portfolios have behaved over the daily sample (January 2020 – November 2026; history before 15 July 2026 was computed retroactively, see the Abstract). Full construction and significance detail follow in §5–§9.

Factor returns by calendar year

Calendar-year return of each long/short factor portfolio (sum of its daily returns within the year), in %, newest year on the left. Green = high-loading names outperformed, amber = underperformed. Thematic factors show “—” before their benchmark ETF began trading; an asterisk (*) marks a theme's first partial year (ETF launched mid-year); 2026 is year-to-date.

Factor2026202520242023202220212020
Style
International-2.7%+5.1%-4.5%-0.5%+3.0%-8.0%-1.1%
Beta+1.5%+10.3%+5.6%+22.7%-16.0%+7.8%+3.3%
Size+1.2%+1.3%+3.1%+4.1%-4.1%+4.3%-2.6%
Long-Term Momentum+6.4%+8.9%+11.0%-0.9%+3.9%-2.0%+18.0%
Residual Volatility+2.9%+4.2%+1.2%+1.8%-3.6%-11.8%+12.1%
Liquidity-0.5%+1.3%-0.7%+2.1%-0.8%+5.3%+16.9%
Growth-3.4%+3.1%+4.1%-1.8%-2.2%+1.9%+4.3%
Value+4.2%-6.5%-10.4%-9.2%+0.3%+0.2%-20.1%
Leverage-3.4%-2.5%-0.7%-6.0%-2.7%-1.2%-11.6%
Profitability+2.3%-3.5%+3.0%+4.4%+7.5%+4.7%+1.2%
Dividend Yield+3.7%+0.9%+3.6%-2.8%+4.3%+3.7%-6.0%
Short Interest-0.6%-4.5%-3.7%-4.7%-5.4%-5.1%+7.9%
Hedge-Fund Ownership-3.8%+0.7%+0.6%+1.9%+0.8%-0.1%-0.5%
Medium-Term Momentum+1.7%-1.9%+4.8%-1.0%-1.6%-0.9%+6.1%
Short-Term Momentum-6.2%+0.7%+0.7%-3.2%+2.3%+1.7%-14.2%
One-Day Momentum+2.7%-3.1%-1.5%-4.4%+3.4%-2.6%-8.5%
Short-Sale Activity-0.1%-8.0%-5.5%-1.8%-4.4%-1.8%-4.9%
Morning Activity-1.6%+6.2%+2.2%-1.6%+3.2%+1.9%+4.9%
Thematic
Semiconductors+3.2%+6.8%+0.1%+4.8%-4.3%-6.2%+7.9%
Oil+7.4%-4.3%-3.2%-3.1%+17.8%+2.2%-8.9%
Treasury (Duration)-4.8%-5.0%-2.6%+0.6%+2.8%+2.5%+11.4%
Gold-3.7%+11.7%+1.0%+0.8%+0.5%-6.6%+5.5%
Bitcoin / Crypto-6.5%-1.3%+2.0%-1.4%-2.4%-1.3%*—
China-4.0%-0.5%-3.1%-2.8%-1.8%-18.6%+0.1%

2026 factor returns by month

The same portfolios at monthly resolution for 2026 (sum of daily returns within each month), in %, newest month on the left. The latest month is month-to-date (through November 3, 2026).

FactorNovOctSepAugJulJunMayAprMarFebJan
Style
International-0.4%-1.4%+0.3%-0.2%+1.5%-1.7%-1.0%-1.0%-0.0%+0.0%+1.0%
Beta+0.1%-12.1%+5.0%+0.9%-11.2%+1.5%+7.9%+8.5%+0.2%-2.0%+2.9%
Size-0.3%-0.9%+2.8%-0.2%+0.3%-1.3%+1.2%+1.7%+0.0%-1.2%-0.9%
Long-Term Momentum-1.8%-3.1%+4.1%-2.4%-5.8%+3.3%+1.6%+3.7%-0.9%+4.5%+3.2%
Residual Volatility-0.8%-0.5%+0.3%+3.7%-0.8%-4.3%+4.5%+0.2%+0.6%-0.1%+0.1%
Liquidity-0.6%-0.7%+0.9%+2.1%+0.0%-3.4%+1.8%+0.8%+0.4%-1.2%-0.6%
Growth-0.1%-0.4%+0.0%+0.6%+1.3%+0.8%+0.4%+0.2%-0.4%-2.8%-3.0%
Value+0.2%+0.2%+0.0%-0.6%+0.6%-1.1%-0.3%+0.6%+1.3%+1.0%+2.3%
Leverage-0.1%+0.2%-1.1%-1.8%-1.6%+0.0%+0.2%+0.7%-0.5%+0.3%+0.4%
Profitability+0.4%+1.1%+0.0%-1.5%-0.4%+0.7%-1.0%+0.2%-1.3%+2.1%+1.9%
Dividend Yield-0.5%-0.1%-0.2%+0.7%-0.7%-0.9%+1.1%+0.8%+1.1%+0.7%+1.6%
Short Interest-0.1%+0.3%-0.9%+0.0%-0.2%+0.9%+0.6%+0.1%+0.3%-1.0%-0.8%
Hedge-Fund Ownership+0.2%-0.5%-0.4%+0.4%+0.3%+0.4%-1.4%-0.2%-0.2%-0.6%-1.7%
Medium-Term Momentum-0.4%-1.0%-1.5%+1.2%-1.7%+2.5%+1.9%+0.3%-1.7%+0.9%+1.2%
Short-Term Momentum+0.1%+2.4%-1.1%-1.9%-5.5%-1.6%+0.5%-0.7%+2.1%+0.4%-1.0%
One-Day Momentum+0.3%+0.6%-1.9%-1.7%-0.0%+1.3%+2.4%+0.6%+1.1%+0.3%-0.3%
Short-Sale Activity-0.1%+0.4%+0.3%+0.5%-0.1%-0.0%+0.0%-1.0%-1.0%+0.3%+0.6%
Morning Activity-0.4%-0.6%-0.5%+1.9%+0.5%-1.1%-0.2%-1.2%-0.1%+0.7%-0.5%
Thematic
Semiconductors-0.7%-3.8%+2.7%-2.7%-5.6%+2.7%-0.0%+3.9%+0.3%+3.6%+2.9%
Oil-2.1%+0.0%+1.2%+2.9%+1.5%-4.2%+1.8%-0.3%+2.5%+2.4%+1.8%
Treasury (Duration)-0.1%+1.6%-1.7%-1.2%+0.8%-0.4%+0.0%-1.9%+0.2%-1.2%-0.9%
Gold-0.0%-1.8%-1.0%+4.7%+1.3%-3.6%-1.6%-2.4%-3.4%+1.3%+2.9%
Bitcoin / Crypto-0.6%-1.6%+0.1%+2.0%+0.1%-0.3%-0.3%-0.8%+1.0%-3.7%-2.4%
China-0.2%-1.3%+0.2%+0.3%+2.1%-3.6%-2.2%-1.6%-1.1%+0.1%+3.3%

Cross-sectional variance explained, by category and year

Variance-weighted share of daily cross-sectional return variance attributable to the market (level), the style block, the thematic block, and stock-specific (unexplained) return. Each column sums to 100%; the per-factor breakdown is in §9.5.

Component2026202520242023202220212020
Market8.6%19.8%13.1%17.6%30.4%14.7%40.2%
Style24.4%16.2%14.4%14.2%20.0%19.6%16.6%
Thematic2.9%2.0%2.1%2.2%1.9%2.0%1.0%
Unexplained (stock-specific)64.1%62.0%70.4%66.0%47.7%63.7%42.2%

2. Notation & Conventions

Fix a trading day \(t\) with universe of size \(N\) (the subscript \(t\) is suppressed where unambiguous). For security \(i\):

  • \(m_i\): market capitalization; \(r_{i}\): daily total return.
  • \(x_i\): a raw descriptor; \(z_i\): its standardized (loading) form.

Cap weights. A square-root market-cap weight damps mega-cap dominance while keeping the weighting economically meaningful:

\[ w_i \;=\; \frac{\sqrt{m_i}}{\sum_{j=1}^{N}\sqrt{m_j}}, \qquad \sum_i w_i = 1. \]

These weights define the inner product under which factors are orthogonalized and portfolios are formed. The cap-weighted mean and variance of a descriptor \(x\) are

\[ \bar{x}_w = \sum_i w_i x_i, \qquad \operatorname{Var}_w(x) = \sum_i w_i\,(x_i - \bar{x}_w)^2 . \]

Cross-sectional standardization. Unless stated otherwise, the standardization operator \(z(\cdot)\) uses the equal-weight cross-sectional moments,

\[ z(x)_i \;=\; \frac{x_i - \mu_x}{\sigma_x}, \qquad \mu_x = \tfrac{1}{N}\sum_i x_i,\quad \sigma_x = \sqrt{\tfrac{1}{N}\sum_i (x_i-\mu_x)^2}, \]

so that every standardized loading has mean \(0\) and standard deviation \(1\) before clipping. Orthogonalized loadings (§5.4) are instead re-standardized under the cap-weighted moments \(\bar{x}_w, \operatorname{Var}_w\).

3. Estimation Universe

The estimation universe is the set of securities eligible to enter the exposure matrix on a given day. It is constructed point-in-time from information available as of day \(t\); no forward-looking membership is ever used.

3.1 Eligibility screens

  1. Security type. Restricted to ordinary common stock and American Depositary Receipts. Funds, units, warrants, rights, and other non-operating-company instruments are excluded.
  2. Price floor. Names trading below a nominal price floor (a sub-$1 close) are removed, so that micro-price quotation noise does not contaminate cross-sectional returns and ratios.
  3. Price-history continuity (survivorship control). A security must be present in the daily cross-section on every business day of a trailing window \(\mathcal{W}\) of length \(L\). Writing \(S_d\) for the set of symbols present on day \(d\), the eligible set is the intersection \[ \mathcal{U}_t \;=\; \bigcap_{d \in \mathcal{W}} S_d . \] This enforces a stable panel with sufficient history for the trailing estimators (beta, volatility, momentum) and prevents newly listed or intermittently quoted names from destabilizing the cross-section.

The daily cross-section is seeded from the broad US equity market augmented by the constituents of a large-cap reference index, so that the universe spans well beyond the largest names while remaining investable.

3.2 Market capitalization

Market capitalization is the product of the closing price and contemporaneous shares outstanding, \( m_i = P_i \times \text{shares}_i \). Capitalization enters the model in two places: as the raw input to the size descriptor, and, via \(\sqrt{m_i}\), as the weight that defines factor portfolios and the orthogonalization inner product (§2).

3.3 Point-in-time discipline

All fundamental inputs are aligned to their as-reported availability (filing dates), and all technical inputs use only trailing windows ending at or before \(t\). Because factor returns are measured close-to-close, an input qualifies for day \(t\) only if it was public before the close that begins the measurement window: semi-monthly short interest enters after its public dissemination, and daily short-sale volume uses the most recent close-tradeable file (the same-evening file publishes after that close and is excluded). Survivorship is controlled by the continuity screen above rather than by any ex-post index membership, so delisted names remain in historical cross-sections for the days on which they traded.

4. Descriptor Hygiene: Winsorization & Standardization

Cross-sectional descriptors have heavy tails: a handful of extreme observations can dominate a standardized loading and, through it, a factor portfolio. The model applies a disciplined three-stage hygiene pipeline before any descriptor is used. The two basic operations are a percentile winsorizer and a symmetric hard cap:

\[ \operatorname{winsor}_p(x)_i = \min\!\big(\max(x_i,\,Q_{p}),\, Q_{100-p}\big), \qquad \operatorname{cap}_c(x)_i = \min\!\big(\max(x_i,-c),\,c\big), \]

where \(Q_q\) is the \(q\)-th cross-sectional percentile of \(x\). \(\operatorname{winsor}_p\) pulls the tails to the \(p\) / \(100{-}p\) percentiles; \(\operatorname{cap}_c\) imposes a hard symmetric bound.

4.1 Stage 1: raw-descriptor treatment

Each raw descriptor is first mapped to its economically admissible range by a hard domain clip (e.g. dividend yield \(\ge 0\); beta within a wide symmetric band), with missing values filled by a neutral constant or the cross-sectional median as appropriate to the descriptor. It is then percentile-winsorized. By convention fundamental ratios are winsorized at the \(5\)/\(95\) percentiles and horizon returns at the \(3\)/\(99\) (or \(3\)/\(97\)) percentiles; one-sided descriptors (short interest, turnover, dividend yield) are winsorized only on the populated tail. Valuation multiples receive special handling: non-positive multiples (negative earnings, etc.) are mapped to the expensive cap rather than discarded, so that an unprofitable company reads as expensive on value, not as missing.

4.2 Stage 2: standardization and loading clip

The treated descriptor is standardized to a loading via \(z(\cdot)\) and then hard-capped:

\[ z^{\text{clip}}_i = \operatorname{cap}_{a}\big(z(x)_i\big). \]

The cap \(a\) is symmetric for most factors; for right-skewed descriptors (dividend yield, short interest, hedge-fund ownership, short-sale activity, morning activity) and the international membership flag an asymmetric band \([-a^-,\,a^+]\) with \(a^- < a^+\) is used, reflecting that their information is concentrated in the populated upper tail.

4.3 Stage 3: orthogonalized-loading cap

After Gram–Schmidt orthogonalization (§5.4) the residual loading is re-standardized under the cap-weighted moments and capped at \(\pm k\) cap-weighted standard deviations. Because small-capitalization tails can inflate the cap-weighted variance, the cap-and-restandardize step is iterated to a fixed point: standardize, cap at \(\pm k\), re-standardize, until the loading simultaneously has cap-weighted mean \(\approx 0\), cap-weighted variance \(\approx 1\), and the orthogonalized loading satisfies \(|z| \le k\).

Why three stages. Stage 1 protects the cross-sectional moments themselves from outliers; Stage 2 protects the standardized loading; Stage 3 protects the orthogonalized loading, whose variance is taken under the cap weights and is therefore more sensitive to a few small, extreme names. The magnitudes \(a\) and \(k\) are proprietary calibration constants.

5. Style Factor Construction

5.1 Descriptors

The 18 style factors and their raw descriptors are listed below. The rank column is the orthogonalization priority (§5.4): a factor is residualized against all factors of lower rank.

RankFactorFamilyDescriptorSign convention
1International
international
risklisting-segment membership: depositary receipts (ADR/ADRC) and curated foreign-headquartered common stockHigher = international listing segment.
2Beta
beta
riskmarket beta \(\beta_i\) from a trailing daily OLS regression on the broad cap-weighted market returnHigher = more market-sensitive.
3Size
size
style\(\log_4(\text{market cap})\)Higher = larger capitalization.
4Long-Term Momentum
long_term_momentum
styletrailing \(\approx\!252\)-trading-day total returnHigher = stronger 1-year winner.
5Residual Volatility
residual_volatility
riska blend of short-horizon realized volatility and a longer-horizon, beta-residualized volatilityHigher = more volatile.
6Liquidity
liquidity
styletrailing average daily dollar volume divided by market capitalization (turnover)Higher = more turnover relative to size.
7Growth
growth
stylecomposite of revenue growth, multi-year revenue CAGR, and sustainable growthHigher = faster fundamental growth.
8Value
value
stylenegated composite of standardized P/E, P/B, P/CF, P/S and EV/EBITDA multiplesHigher = cheaper (high book/price, earnings yield).
9Leverage
leverage
stylecomposite of market leverage, debt-to-assets, and net-debt-to-EBITDAHigher = more levered.
10Profitability
profitability
stylecomposite of ROIC, ROA, ROE, gross profitability, and operating profitabilityHigher = more profitable.
11Dividend Yield
dividend_yield
styletrailing dividends per share divided by priceHigher = higher payout yield.
12Short Interest
short_interest
positioningreported short interest as a fraction of shares outstanding, standardized within listing segmentHigher = more crowded short (relative to its own listing segment).
13Hedge-Fund Ownership
hfown
positioning13F crowding composite: within-size-tercile percentile ranks of holder breadth and ownership magnitudeHigher = more hedge-fund-crowded.
14Medium-Term Momentum
medium_term_momentum
flow\(r_{20d} - r_{5d}\) (one-month less one-week return)Higher = stronger 1-week-to-1-month trend.
15Short-Term Momentum
short_term_momentum
flow\(r_{5d} - r_{1d}\) (one-week less one-day return)Higher = stronger multi-day trend ex-yesterday.
16One-Day Momentum
one_day_momentum
flowprior-session return \(r_{1d}\)Higher = stronger prior-day return.
17Short-Sale Activity
short_sale
flowa short-sale velocity score from threshold-list status and short volumeHigher = elevated short-sale flow.
18Morning Activity
morning_activity
flowtrailing average rank of morning trading intensity — morning-session volume scaled by the stock's recent average daily volume, relative to the rest of the sessionHigher = more morning-concentrated, volume-heavy mornings.

5.2 Composite descriptors

Six descriptors are themselves composites of standardized sub-signals. Each composite is a weighted average of \(z\)-scored components; component weights are proprietary and shown as \(w_{\bullet}\) with normalizing divisor \(\textstyle\sum w\).

Value (sign-flipped so cheap \(\to\) high):

\[ \text{value}_i \;=\; -\,\frac{ w_{PE}\,z(\text{P/E})_i + w_{PB}\,z(\text{P/B})_i + w_{PCF}\,z(\text{P/CF})_i + w_{PS}\,z(\text{P/S})_i + w_{EV}\,z(\text{EV/EBITDA})_i } {\max\!\big(\sum_{\text{present}} w,\; w_{\min}\big)}. \]

The sum runs over the multiples the company actually has: a missing multiple drops out and the denominator is the weight of the present terms, floored at \(w_{\min}\) so that a single multiple cannot carry a full-weight score. A company with no multiple at all receives a neutral value loading. A negative multiple (negative earnings, etc.) is not treated as missing; it is set to the expensive cap (§4.1). For names with negative book equity the book term is removed (\(w_{PB}\!\to\!0\)) and the cash-flow term up-weighted, since book/price is uninformative when equity is impaired. Price-to-sales is not used for banks and brokers, whose reported revenue is not comparable with other companies' sales.

Growth: revenue growth, multi-year revenue CAGR, and sustainable growth:

\[ \text{growth}_i = w_{g1}\,z(\text{rev growth})_i + w_{g2}\,z(\text{rev CAGR})_i + w_{g3}\,z(\text{sustainable growth})_i, \]

with the convention that if a component is unavailable its weight is redistributed to the populated components (the weights always sum to one over present terms).

Leverage: market leverage, debt-to-assets, net-debt-to-EBITDA:

\[ \text{leverage}_i = \frac{ w_{l1}\,z(\text{MLEV})_i + w_{l2}\,z(\text{D/A})_i + w_{l3}\,z(\text{NetDebt/EBITDA})_i }{\sum w_l}. \]

Profitability: ROIC, ROA, ROE, gross profitability, operating profitability:

\[ \text{profitability}_i = \frac{ w_{p1} z(\text{ROIC}) + w_{p2} z(\text{ROA}) + w_{p3} z(\text{ROE}) + w_{p4} z(\text{GP}) + w_{p5} z(\text{OP}) }{\sum w_p}, \]

with ROE down-weighted and ROIC up-weighted for negative-equity names.

Residual volatility is a volatility blend (a convex combination of a short-horizon realized volatility and a longer-horizon, beta-residualized volatility), hard-bounded to a positive range. Blending in the longer-horizon, beta-residualized estimate dampens transient single-window spikes and the portion of volatility already explained by market beta, leaving a cleaner idiosyncratic-volatility descriptor at this high-priority rank.

Hedge-fund ownership is a within-size-tercile composite of two percentile ranks: the breadth of material 13F holders and the magnitude of 13F ownership relative to capitalization:

\[ \text{hfown}_i = \tfrac12\Big( \operatorname{pctrank}_{\text{tercile}}(\text{breadth})_i + \operatorname{pctrank}_{\text{tercile}}(\text{ownership}/m)_i \Big). \]

5.3 Standardized loadings

Each (composite or raw) descriptor is standardized and clipped per §4.2 to produce the univariate loading (suffix _u), \( z^{u}_{i} = \operatorname{cap}_{a}(z(x)_i) \).

5.4 Gram–Schmidt orthogonalization

Style factors share variance: high-beta names tend to be volatile, small names tend to be illiquid, and so on. To obtain loadings that explain return independently, the univariate loadings are orthogonalized in the fixed priority order of §5.1 under the cap-weighted inner product. Let \(u^{(1)},\dots,u^{(K)}\) denote the orthogonalized loadings. Setting \(u^{(1)} = z^{u}\) for international, each subsequent factor \(i\) removes its cap-weighted projection onto every prior orthogonal loading:

\[ \beta_{ij} \;=\; \frac{\sum_n w_n\, z^{u}_{i,n}\, u^{(j)}_n} {\sum_n w_n\, \big(u^{(j)}_n\big)^2}, \qquad v_i \;=\; z^{u}_{i} \;-\; \sum_{j<i} \beta_{ij}\, u^{(j)}, \]

and \(u^{(i)}\) is the iterated standardize-and-cap of the residual \(v_i\) (§4.3). The resulting loadings satisfy \(\mathbf{U}^{\!\top} \mathbf{W}\, \mathbf{U} \approx \mathbf{I}\): the orthogonalized factors are mutually near-uncorrelated under the cap-weighted metric (on the production model the maximum off-diagonal correlation magnitude is \(\approx 0.02\), mean \(\approx 0.001\)).

Orthogonal exposures are not uncorrelated returns. Gram–Schmidt removes the cross-sectional correlation between factor loadings (the exposure correlation matrix is \(\approx\mathbf{I}\)) but it says nothing about how the factor portfolios go on to perform. The realized daily factor returns (defined in §8.1) retain substantial correlation structure (Figure 2): value moves opposite growth and residual volatility, the short-horizon momentum factors co-move, and so on. Decorrelated exposures do not produce decorrelated returns, which is exactly why the factor-return covariance matrix \(\mathbf{F}\) (§8.3) is essential and far from diagonal.
INTLBETASIZELTMRVOLLIQGRWVALLEVPROFDVDSIHFMTMSTMODMSSMORNINTLINTL × INTL = +1.00INTL × BETA = -0.01INTL × SIZE = +0.09INTL × LTM = +0.15INTL × RVOL = +0.27INTL × LIQ = +0.00INTL × GRW = +0.02INTL × VAL = -0.13INTL × LEV = -0.13INTL × PROF = +0.04INTL × DVD = -0.02INTL × SI = -0.02INTL × HF = +0.15INTL × MTM = -0.02INTL × STM = -0.06INTL × ODM = -0.05INTL × SS = -0.13INTL × MORN = +0.07BETABETA × INTL = -0.01BETA × BETA = +1.00BETA × SIZE = -0.27BETA × LTM = -0.03BETA × RVOL = +0.40BETA × LIQ = +0.32BETA × GRW = +0.07BETA × VAL = -0.11BETA × LEV = +0.02BETA × PROF = -0.14BETA × DVD = -0.07BETA × SI = -0.05BETA × HF = +0.00BETA × MTM = -0.16BETA × STM = -0.21BETA × ODM = -0.17BETA × SS = +0.09BETA × MORN = +0.08SIZESIZE × INTL = +0.09SIZE × BETA = -0.27SIZE × SIZE = +1.00SIZE × LTM = +0.30SIZE × RVOL = -0.37SIZE × LIQ = -0.26SIZE × GRW = +0.25SIZE × VAL = -0.18SIZE × LEV = -0.30SIZE × PROF = +0.23SIZE × DVD = -0.15SIZE × SI = +0.02SIZE × HF = +0.00SIZE × MTM = +0.16SIZE × STM = +0.17SIZE × ODM = +0.17SIZE × SS = -0.12SIZE × MORN = -0.05LTMLTM × INTL = +0.15LTM × BETA = -0.03LTM × SIZE = +0.30LTM × LTM = +1.00LTM × RVOL = +0.05LTM × LIQ = -0.08LTM × GRW = +0.25LTM × VAL = -0.22LTM × LEV = -0.22LTM × PROF = +0.01LTM × DVD = -0.04LTM × SI = -0.07LTM × HF = -0.14LTM × MTM = +0.14LTM × STM = +0.13LTM × ODM = +0.04LTM × SS = -0.10LTM × MORN = +0.02RVOLRVOL × INTL = +0.27RVOL × BETA = +0.40RVOL × SIZE = -0.37RVOL × LTM = +0.05RVOL × RVOL = +1.00RVOL × LIQ = +0.38RVOL × GRW = +0.08RVOL × VAL = -0.31RVOL × LEV = -0.10RVOL × PROF = -0.35RVOL × DVD = +0.00RVOL × SI = +0.15RVOL × HF = +0.01RVOL × MTM = +0.01RVOL × STM = -0.06RVOL × ODM = -0.07RVOL × SS = +0.00RVOL × MORN = +0.04LIQLIQ × INTL = +0.00LIQ × BETA = +0.32LIQ × SIZE = -0.26LIQ × LTM = -0.08LIQ × RVOL = +0.38LIQ × LIQ = +1.00LIQ × GRW = -0.01LIQ × VAL = -0.02LIQ × LEV = +0.06LIQ × PROF = -0.07LIQ × DVD = -0.01LIQ × SI = -0.14LIQ × HF = -0.06LIQ × MTM = -0.17LIQ × STM = -0.22LIQ × ODM = -0.17LIQ × SS = +0.10LIQ × MORN = +0.07GRWGRW × INTL = +0.02GRW × BETA = +0.07GRW × SIZE = +0.25GRW × LTM = +0.25GRW × RVOL = +0.08GRW × LIQ = -0.01GRW × GRW = +1.00GRW × VAL = -0.39GRW × LEV = -0.32GRW × PROF = -0.08GRW × DVD = -0.43GRW × SI = +0.15GRW × HF = +0.18GRW × MTM = -0.10GRW × STM = -0.16GRW × ODM = -0.05GRW × SS = -0.07GRW × MORN = +0.01VALVAL × INTL = -0.13VAL × BETA = -0.11VAL × SIZE = -0.18VAL × LTM = -0.22VAL × RVOL = -0.31VAL × LIQ = -0.02VAL × GRW = -0.39VAL × VAL = +1.00VAL × LEV = +0.20VAL × PROF = +0.46VAL × DVD = +0.34VAL × SI = -0.21VAL × HF = -0.18VAL × MTM = -0.05VAL × STM = +0.08VAL × ODM = +0.08VAL × SS = +0.07VAL × MORN = -0.13LEVLEV × INTL = -0.13LEV × BETA = +0.02LEV × SIZE = -0.30LEV × LTM = -0.22LEV × RVOL = -0.10LEV × LIQ = +0.06LEV × GRW = -0.32LEV × VAL = +0.20LEV × LEV = +1.00LEV × PROF = -0.02LEV × DVD = +0.34LEV × SI = -0.28LEV × HF = -0.07LEV × MTM = -0.10LEV × STM = -0.09LEV × ODM = -0.13LEV × SS = +0.01LEV × MORN = +0.02PROFPROF × INTL = +0.04PROF × BETA = -0.14PROF × SIZE = +0.23PROF × LTM = +0.01PROF × RVOL = -0.35PROF × LIQ = -0.07PROF × GRW = -0.08PROF × VAL = +0.46PROF × LEV = -0.02PROF × PROF = +1.00PROF × DVD = -0.03PROF × SI = -0.11PROF × HF = -0.11PROF × MTM = +0.00PROF × STM = +0.06PROF × ODM = +0.11PROF × SS = +0.08PROF × MORN = -0.04DVDDVD × INTL = -0.02DVD × BETA = -0.07DVD × SIZE = -0.15DVD × LTM = -0.04DVD × RVOL = +0.00DVD × LIQ = -0.01DVD × GRW = -0.43DVD × VAL = +0.34DVD × LEV = +0.34DVD × PROF = -0.03DVD × DVD = +1.00DVD × SI = -0.20DVD × HF = -0.29DVD × MTM = +0.05DVD × STM = +0.13DVD × ODM = +0.05DVD × SS = -0.02DVD × MORN = -0.11SISI × INTL = -0.02SI × BETA = -0.05SI × SIZE = +0.02SI × LTM = -0.07SI × RVOL = +0.15SI × LIQ = -0.14SI × GRW = +0.15SI × VAL = -0.21SI × LEV = -0.28SI × PROF = -0.11SI × DVD = -0.20SI × SI = +1.00SI × HF = +0.16SI × MTM = +0.04SI × STM = +0.19SI × ODM = +0.24SI × SS = +0.06SI × MORN = -0.06HFHF × INTL = +0.15HF × BETA = +0.00HF × SIZE = +0.00HF × LTM = -0.14HF × RVOL = +0.01HF × LIQ = -0.06HF × GRW = +0.18HF × VAL = -0.18HF × LEV = -0.07HF × PROF = -0.11HF × DVD = -0.29HF × SI = +0.16HF × HF = +1.00HF × MTM = -0.11HF × STM = -0.12HF × ODM = -0.10HF × SS = -0.14HF × MORN = +0.13MTMMTM × INTL = -0.02MTM × BETA = -0.16MTM × SIZE = +0.16MTM × LTM = +0.14MTM × RVOL = +0.01MTM × LIQ = -0.17MTM × GRW = -0.10MTM × VAL = -0.05MTM × LEV = -0.10MTM × PROF = +0.00MTM × DVD = +0.05MTM × SI = +0.04MTM × HF = -0.11MTM × MTM = +1.00MTM × STM = +0.25MTM × ODM = +0.11MTM × SS = -0.05MTM × MORN = -0.01STMSTM × INTL = -0.06STM × BETA = -0.21STM × SIZE = +0.17STM × LTM = +0.13STM × RVOL = -0.06STM × LIQ = -0.22STM × GRW = -0.16STM × VAL = +0.08STM × LEV = -0.09STM × PROF = +0.06STM × DVD = +0.13STM × SI = +0.19STM × HF = -0.12STM × MTM = +0.25STM × STM = +1.00STM × ODM = +0.31STM × SS = -0.05STM × MORN = -0.07ODMODM × INTL = -0.05ODM × BETA = -0.17ODM × SIZE = +0.17ODM × LTM = +0.04ODM × RVOL = -0.07ODM × LIQ = -0.17ODM × GRW = -0.05ODM × VAL = +0.08ODM × LEV = -0.13ODM × PROF = +0.11ODM × DVD = +0.05ODM × SI = +0.24ODM × HF = -0.10ODM × MTM = +0.11ODM × STM = +0.31ODM × ODM = +1.00ODM × SS = -0.02ODM × MORN = -0.03SSSS × INTL = -0.13SS × BETA = +0.09SS × SIZE = -0.12SS × LTM = -0.10SS × RVOL = +0.00SS × LIQ = +0.10SS × GRW = -0.07SS × VAL = +0.07SS × LEV = +0.01SS × PROF = +0.08SS × DVD = -0.02SS × SI = +0.06SS × HF = -0.14SS × MTM = -0.05SS × STM = -0.05SS × ODM = -0.02SS × SS = +1.00SS × MORN = +0.08MORNMORN × INTL = +0.07MORN × BETA = +0.08MORN × SIZE = -0.05MORN × LTM = +0.02MORN × RVOL = +0.04MORN × LIQ = +0.07MORN × GRW = +0.01MORN × VAL = -0.13MORN × LEV = +0.02MORN × PROF = -0.04MORN × DVD = -0.11MORN × SI = -0.06MORN × HF = +0.13MORN × MTM = -0.01MORN × STM = -0.07MORN × ODM = -0.03MORN × SS = +0.08MORN × MORN = +1.00
Figure 2. Realized correlation matrix of the orthogonalized style-factor returns (daily, Jan 2020 – Nov 2026). Although the factor exposures are near-orthogonal by construction (§5.4), the factor returns show clear structure: a co-moving short-horizon momentum block (STM–ODM ≈ 0.31), value opposed to growth and residual volatility (≈ −0.39, −0.31) but aligned with profitability (≈ 0.46), and a beta–residual-volatility–liquidity risk cluster (≈ 0.32–0.40). Short interest and short-sale activity are nearly uncorrelated (≈ 0.06).

5.5 Univariate (_u) versus orthogonalized

Both variants are published. The univariate loading (_u) is a security's raw tilt to a characteristic; the orthogonalized loading, published under the bare factor name, is the part of that tilt not explained by higher-priority factors. Risk decomposition and attribution lead with the orthogonalized form, since the univariate form conflates a factor's own contribution with shared variance. The divergence between the two is diagnostic: a name with high \(\text{short interest}_u\) but near-zero orthogonalized \(\text{short interest}\) may be crowded-short largely because it is already small, volatile, or a momentum loser, in which case the short interest carries little information independent of those factors.

6. Thematic Factor Construction

Thematic factors capture exposure to liquid macro / sector themes that the fundamental style factors do not span. Each theme is proxied by a liquid benchmark ETF, and a security's thematic exposure is its rolling time-series beta to that ETF's daily return.

ThemeProxyNotes
Semiconductors
semis
a liquid semiconductor sector ETFHighest-priority theme after market beta; semis carry a large, distinct factor footprint.
Oil
oil
a liquid crude-oil commodity ETFEnergy / commodity beta.
Treasury (Duration)
treasury
a long-dated Treasury bond ETFRates / duration sensitivity.
Gold
gold
a physical-gold ETFPrecious-metal / real-asset beta.
Bitcoin / Crypto
bitcoin
a liquid bitcoin-linked ETFCrypto beta; relevant for miners, exchanges, and treasury-holders.
China
china
a China large-cap equity ETFChina large-cap equity beta.

6.1 Rolling theme beta

Over a trailing window of \(W\) business days (requiring at least \(W_{\min}\) valid observations), the exposure of security \(i\) to theme ETF \(e\) is the univariate regression beta of the security's daily return on the ETF's daily return:

\[ \beta_{i,e} \;=\; \frac{\operatorname{cov}_W\!\big(r_i,\, r_e\big)} {\operatorname{var}_W\!\big(r_e\big)} \;=\; \frac{\sum_{\tau \in W} (r_{i,\tau}-\bar r_i)(r_{e,\tau}-\bar r_e)} {\sum_{\tau \in W} (r_{e,\tau}-\bar r_e)^2}. \]

Securities with fewer than \(W_{\min}\) observations in the window receive no theme beta (treated as neutral). Raw betas are hard-bounded to a sensible range before standardization.

6.2 Standardization and orthogonalization

Theme betas are standardized and clipped per §4.2; because negative theme loadings carry little economic meaning for most commodity / sector themes, the loading clip is asymmetric on the negative side. The themes are then Gram–Schmidt orthogonalized (§5.4) in a fixed priority order led by market beta (market beta, then semiconductors, oil, treasury, gold, bitcoin, China), so that each thematic loading measures exposure to its theme net of the broad market and of the earlier themes: e.g. semiconductor exposure beyond what a stock's market beta already implies. Theme loadings are not orthogonalized against the style factors, so a theme exposure can be correlated with style exposures such as size or momentum. When variance is attributed (§9.5), the theme factors are fitted to what remains after the market and style factors, so the theme share is incremental to style.

7. Pure-Factor Portfolios

Each factor is realized as a net-flat (zero-delta) long/short portfolio: security weights that net to zero and are unit-scaled. Two constructions are used.

7.1 Cap-weighted loading portfolio

For a single loading \(z\) (univariate _u or orthogonalized), the simple factor portfolio multiplies the loading by the cap weight, demeans, and rescales so absolute weights sum to one:

\[ \tilde p_i = z_i\, w_i, \qquad p_i = \frac{\tilde p_i - \overline{\tilde p}}{\sum_j |\tilde p_j - \overline{\tilde p}|}, \qquad \sum_i p_i = 0,\quad \sum_i |p_i| = 1. \]

The result has \(\sum_i p_i = 0\) and unit gross exposure \(\sum_i|p_i|=1\): a long/short tilt toward high-loading names funded by low-loading names.

7.2 Weighted-least-squares pure-factor portfolio

For the orthogonalized exposure matrix \(\mathbf{X}\) (columns \(=\) orthogonalized loadings) and diagonal weight matrix \(\mathbf{W}=\operatorname{diag}(w)\), the cap-weighted cross-sectional regression yields the pure-factor portfolio matrix

\[ \mathbf{B} \;=\; \big(\mathbf{X}^{\!\top}\mathbf{W}\mathbf{X}\big)^{-1} \mathbf{X}^{\!\top}\mathbf{W} \;\in\; \mathbb{R}^{K \times N}. \]

Row \(f\) of \(\mathbf{B}\) is the portfolio with unit exposure to factor \(f\) and zero exposure to every other factor: \(\mathbf{B}\mathbf{X}=\mathbf{I}_K\). Each row is then demeaned and abs-sum-normalized as in §7.1. Because the columns of \(\mathbf{X}\) are already near-orthogonal, \(\mathbf{X}^{\!\top}\mathbf{W}\mathbf{X}\) is well-conditioned; a least-squares solve is used as a fallback if it is singular.

8. Factor Returns, Covariance & Attribution

8.1 Daily factor returns

The realized return of factor \(f\) on day \(t\) is the inner product of its portfolio weights (formed on day \(t\)) with next-day security returns:

\[ f_{f,t} \;=\; \mathbf{p}_{f,t}^{\!\top}\, \mathbf{r}_{t+1} \;=\; \sum_i p_{f,i,t}\, r_{i,t+1}. \]

Accumulated over time, \(\{f_{f,t}\}\) is the factor's return time series, the object used for regression-based attribution (§8.4). The risk model's covariance is estimated on the closely related per-unit-exposure returns (§8.3).

8.2 Cross-sectional attribution

On any given day each security's realized forward return decomposes into a factor part and an idiosyncratic residual:

\[ r_{i,t+1} \;=\; \sum_{f=1}^{K} \beta_{i,f,t}\, f_{f,t} \;+\; \varepsilon_{i,t+1}, \]

where \(\beta_{i,f,t}\) is security \(i\)'s exposure to factor \(f\). The per-factor contribution \(\beta_{i,f,t}\, f_{f,t}\) is the portion of the security's return attributable to that factor; summing the contributions plus the residual reconstructs the total return. Leading with the orthogonalized exposures makes these contributions additive and non-overlapping.

8.3 Factor-return covariance and the risk model

Let \(\mathbf{f}_t = (f_{1,t},\dots,f_{K,t})^{\!\top}\). The factor-return covariance matrix is estimated from the time series,

\[ \mathbf{F} \;=\; \operatorname{cov}(\mathbf{f}) \;=\; \mathbf{D}^{1/2}\, \mathbf{C}\, \mathbf{D}^{1/2}, \]

where \(\mathbf{D}=\operatorname{diag}(\operatorname{var}(f_f))\) and \(\mathbf{C}\) is the factor-return correlation matrix (Figure 2, §5.4). The total covariance of security returns then follows the variance decomposition

\[ \boldsymbol{\Sigma} \;=\; \mathbf{X}\,\mathbf{F}\,\mathbf{X}^{\!\top} \;+\; \boldsymbol{\Delta}, \]

with \(\mathbf{X}\) the exposure matrix and \(\boldsymbol{\Delta}\) the diagonal matrix of idiosyncratic (specific) variances. Portfolio risk decomposes into a common-factor term \(\mathbf{w}^{\!\top}\mathbf{X}\mathbf{F}\mathbf{X}^{\!\top}\mathbf{w}\) and a specific term \(\mathbf{w}^{\!\top}\boldsymbol{\Delta}\mathbf{w}\). As Figure 2 shows, the orthogonalized exposures are near-uncorrelated in the cross-section but the factor returns are not: value, for example, is negatively correlated with growth and residual volatility (\(\approx-0.39\) and \(-0.31\)) and positively aligned with profitability (\(\approx+0.46\)), so \(\mathbf{F}\) is far from diagonal and is essential to honest risk aggregation.

In the production risk model, \(K\) includes the market factor (exposure 1 for every security) alongside the 18 style and 6 theme factors. \(\mathbf{F}\) is estimated on the per-unit-exposure (cross-sectional regression) factor returns of §9.5 rather than on the abs-sum-normalized portfolio returns, with exponentially weighted variances and correlations and the correlations shrunk toward a structured target. Specific variances are estimated from each security's own residual history, with a peer-group fallback for names with too little history, and the forecast is scaled by a volatility-regime adjustment so that its level tracks recently realized risk.

8.4 Time-series alpha attribution

A portfolio's realized excess return can be regressed on the factor-return series, \( R^{\text{port}}_t = \alpha + \sum_f b_f\, f_{f,t} + u_t \). The fitted \(b_f\) are the portfolio's factor exposures, \(\sum_f b_f f_{f,t}\) is its factor-explained return, and the intercept/residual \(\alpha + u_t\) is the portfolio's residual (idiosyncratic) alpha net of all modeled factors.

9. Empirical Factor Behavior

Factor exposures are clean and orthogonal by construction (§5.4); factor returns are an empirical matter. The table reports the realized performance of the orthogonalized style-factor portfolios over the daily sample (January 2020 – November 2026; 1,719 trading days): annualized return and the \(t\)-statistic of the daily series. These figures are in-sample, gross of transaction and borrow costs, and strongly regime-dependent; they describe history, not a forecast.

FactorAnn. returnt-statRealized read
International-1.3%-0.7US names outperformed (segment spread)
Beta+5.2%+0.8High-beta outperformed — but entirely since 2023
Size+1.1%+0.5Flat
Long-Term Momentum+6.7%+1.7Momentum premium; crash-prone (skew ≈ −1.0)
Residual Volatility+1.0%+0.3Flat
Liquidity+3.4%+1.9High-turnover outperformed; gains in 2020–21
Growth+0.9%+0.6Flat
Value-6.1%-4.2Cheap names underperformed
Leverage-4.1%-3.6More-levered names underperformed
Profitability+2.9%+2.5Profitable names outperformed
Dividend Yield+1.1%+1.0Slightly positive
Short Interest-2.4%-2.1Modestly negative (within-segment crowding)
Hedge-Fund Ownership-0.1%-0.1Flat
Medium-Term Momentum+1.1%+0.6Slightly positive
Short-Term Momentum-2.7%-1.3Mild reversal (insignificant)
One-Day Momentum-2.1%-1.1Mild reversal (insignificant)
Short-Sale Activity-3.9%-5.6High short-sale flow underperformed
Morning Activity+2.2%+2.7Positive
Reading the table. Each factor portfolio is long high-loading names and short low-loading names (§7), so a positive annualized return means high-loading names outperformed. \(|t|\gtrsim 2\) is conventionally significant; values are rounded. The calendar-year and thematic breakdowns are in Historical Factor Performance above.

9.1 Short-sale flow, morning activity and short interest

The most statistically significant factor in the model is short-sale activity: long the highest short-sale-flow names, it realized \(\approx-4\%\) annualized (\(t\approx-6\)), negative in both halves of the sample. The construction follows a strict timing discipline: the daily short-sale file publishes after the close from which day-over-day returns are measured, so only the most recent close-tradeable file enters the descriptor. (Using the same-evening file instead would roughly double the factor's apparent return on information a close-priced portfolio could not act on; it is deliberately excluded.) The other microstructure flow descriptor, morning activity, is also robustly priced (\(\approx+2.2\%\), \(t\approx 2.7\), positive in both halves). The short-interest factor (the stock of crowded shorts, standardized within listing segment and built strictly point-in-time on FINRA's semi-monthly publication schedule) realized a modest negative premium (\(\approx-2.4\%\) annualized, \(t\approx-2.1\)): within-segment crowding carries some information, but far less than the flow. The two shorting factors are nearly uncorrelated (\(\approx 0.06\)): it is the flow of short-selling, not the level of short interest, that carries the pricing-relevant information.

9.2 Momentum

Long-term (12-month) momentum earned a positive premium of borderline significance (\(\approx +7\%\) annualized, \(t\approx 1.7\)), positive in both sub-periods, but with a strongly left-skewed distribution (skew \(\approx-1.0\); worst single day near \(-7\%\)): a sharp left tail in which prior losers rally hardest during broad rebounds. The shorter-horizon momentum factors are weaker: medium-term is mildly positive, while short-term and one-day momentum show mild, statistically insignificant reversal. The short-horizon momentum factors co-move moderately (\(\rho \approx 0.25\!-\!0.31\)); long-term momentum is only weakly linked to them (\(\approx 0.04\!-\!0.14\); Figure 2).

9.3 Value, profitability, leverage

Over this sample value (long cheap) realized a significantly negative return (\(\approx-6\%\) annualized, \(t\approx-4.2\), negative in both sub-periods): cheap names underperformed, with the losses concentrated in 2020 and 2023–2025. Value's return correlations are specific, not blanket: strongly negative with growth (\(\approx-0.39\)) and residual volatility (\(\approx-0.31\)), and only mildly negative with long-term momentum (\(\approx-0.22\)), but roughly uncorrelated with medium-term momentum and slightly positively correlated with short-term and one-day momentum (\(\approx+0.08\)). Value is most positively aligned with profitability (\(\approx+0.46\)), which itself earned a positive premium (\(\approx+2.9\%\), \(t\approx 2.5\)). Value is measured net of listing-segment structure: international names load “expensive” on the composite, and the international intercept (§9.4) keeps the segment spread from entering the factor; what remains is the within-segment value premium. Leverage was negative (\(\approx-4\%\), \(t\approx-3.6\)): more-levered names underperformed.

9.4 Liquidity, beta, size, international: regime dependence

The liquidity (turnover) factor realized a modestly positive return over the sample (\(\approx +3.4\%\), \(t\approx 1.9\)): high-turnover names outperformed, but almost entirely in 2020–2021, and roughly flat since. This is the opposite sign to the common long-horizon prior that high-turnover, high-attention names are over-priced, and is a clean reminder that realized factor premia are sample- and regime-specific. Beta was positive overall (\(\approx +5\%\), insignificant) but earned entirely from 2023 onward; over 2020–2022 a large 2022 loss offset earlier gains. Size, residual volatility, growth, and hedge-fund ownership were close to flat with insignificant \(t\)-statistics over the window. International, the listing-segment intercept, is the purest regime factor in the model: a small negative drift overall (\(\approx-1\%\) annualized, insignificant) built from persistently negative US-outperformance years punctuated by clean sign flips (positive in 2022 and 2025). It is carried as a neutralizing dimension, not an alpha: its role is to keep the US-vs-international segment spread out of every other style factor.

9.5 Cross-sectional variance explained, by factor and year

A second, complementary measure of factor significance is how much of the day-to-day cross-sectional dispersion in returns each factor accounts for. On each day the cap-weighted cross-sectional variance of returns decomposes exactly into a market (level) term, a per-factor term, and a stock-specific residual:

\[ \underbrace{\sum_i w_i R_i^2}_{\text{total}} \;=\; \underbrace{f_{\text{mkt}}^2}_{\text{market}} \;+\; \underbrace{\sum_f f_f^2}_{\text{factors}} \;+\; \underbrace{\sum_i w_i u_i^2}_{\text{unexplained}}, \qquad f_f = \sum_i w_i\, \beta_{i,f}\,(R_i - f_{\text{mkt}}), \]

where \(f_{\text{mkt}} = \sum_i w_i R_i\) is the cap-weighted average (market) return, \(f_f\) is the per-unit-exposure (cross-sectional regression) return of factor \(f\), the regression counterpart of the portfolio return of §8.1, and \(u_i\) is the stock-specific residual. Each factor's incremental share of variance is \(f_f^2 / \sum_i w_i R_i^2\); the thematic factors are added on the style residual, so their share is incremental to the style block (no double-counting). The table reports each component's variance-weighted share by calendar year, \(\sum_{\text{days}} (\text{component}) / \sum_{\text{days}}(\text{total})\), so the highest-volatility days dominate a year's figures. Every column sums to 100%.

Component2026202520242023202220212020
Market8.6%19.8%13.1%17.6%30.4%14.7%40.2%
Style factors(24.4)(16.2)(14.4)(14.2)(20.0)(19.6)(16.6)
International0.3%0.5%0.7%0.6%0.8%0.9%0.5%
Beta12.8%8.4%5.7%5.6%11.0%4.3%4.9%
Size0.7%0.9%1.5%1.5%0.7%1.4%1.4%
Long-Term Momentum3.7%1.9%1.7%1.4%1.9%2.5%3.5%
Residual Volatility1.6%1.1%1.0%1.0%1.6%4.3%1.0%
Liquidity0.5%0.4%0.3%0.3%0.3%0.5%0.5%
Growth0.3%0.3%0.2%0.2%0.2%1.0%0.5%
Value0.4%0.3%0.4%0.6%0.4%0.7%0.4%
Leverage0.2%0.2%0.3%0.3%0.2%0.3%0.4%
Profitability0.3%0.3%0.3%0.3%0.3%0.3%0.2%
Dividend Yield0.3%0.2%0.3%0.3%0.2%0.2%0.3%
Short Interest0.3%0.2%0.2%0.2%0.2%0.3%0.2%
Hedge-Fund Ownership0.2%0.1%0.1%0.1%0.1%0.1%0.1%
Medium-Term Momentum0.8%0.4%0.6%0.5%0.6%0.8%0.5%
Short-Term Momentum1.0%0.4%0.4%0.6%0.7%1.0%1.2%
One-Day Momentum0.8%0.4%0.5%0.5%0.6%0.8%0.8%
Short-Sale Activity0.1%0.1%0.1%0.1%0.1%0.1%0.1%
Morning Activity0.1%0.1%0.1%0.1%0.1%0.1%0.1%
Thematic factors(2.9)(2.0)(2.1)(2.2)(1.9)(2.0)(1.0)
Semiconductors0.9%0.5%0.5%0.5%0.3%0.5%0.3%
Oil0.7%0.5%0.5%0.7%0.5%0.6%0.2%
Treasury (Duration)0.2%0.2%0.3%0.2%0.3%0.3%0.2%
Gold0.6%0.3%0.2%0.3%0.2%0.2%0.2%
Bitcoin / Crypto0.3%0.2%0.3%0.2%0.2%0.0%0.0%
China0.2%0.3%0.3%0.3%0.4%0.4%0.1%
Unexplained (stock-specific)64.1%62.0%70.4%66.0%47.7%63.7%42.2%
Reading the table. Variance-weighted share of daily cross-sectional return variance (cells shaded by magnitude). Market is the systematic level move (the one component with the same loading on every name), so its share is large in high-correlation regimes, when stocks move together and the common move dominates the cross-section (e.g. 2020 and 2022), and shrinks when return dispersion is high. Unexplained is stock-specific (idiosyncratic) return. Within the style block, beta and momentum dominate while most individual factors are small: orthogonal factors each carve out only their incremental slice. Group sub-totals are shown in parentheses (the indented factor rows are what sum). 2026 is partial (through Nov 2026).

Appendix A: Factor Glossary

Quick reference for every FPUS1 factor besides the market (18 style + 6 thematic). Style factors list their orthogonalization rank; thematic factors are marked T; the three dynamics factors are tagged dynamics. Each card has a stable link target, #factor-<slug>, matching the address of that factor's page on the FPUS1 model page.

Internationalinternationalrisk
listing-segment membership: depositary receipts (ADR/ADRC) and curated foreign-headquartered common stock
Higher = international listing segment.
The listing-segment intercept: a two-point membership exposure placed first in the hierarchy, so every other style is measured net of the US-vs-international segment (the market intercept itself remains implicit in the demeaning). Its return is the cap-weighted international-vs-US spread — a regime factor (negative in US-outperformance years, positive in 2022 and 2025), neutralized rather than harvested.
Betabetarisk
market beta \(\beta_i\) from a trailing daily OLS regression on the broad cap-weighted market return
Higher = more market-sensitive.
Estimated against the broad cap-weighted market, so beta captures co-movement with the aggregate tape; large-capitalization moves are a meaningful part of that benchmark by construction. A name enters with a beta once a minimum count of paired daily observations exists; until it has a full window, the partial-window estimate is shrunk toward the market, \(\beta_i \leftarrow w\,\hat\beta_i + (1-w)\cdot 1\), damping the noisy, often extreme betas of recent listings. Second in the hierarchy, behind only the international flag, so its raw and orthogonalized forms nearly coincide.
Sizesizestyle
\(\log_4(\text{market cap})\)
Higher = larger capitalization.
Log base is a monotone rescaling; the choice only affects the spread of the raw descriptor before standardization.
Long-Term Momentumlong_term_momentumstyle
trailing \(\approx\!252\)-trading-day total return
Higher = stronger 1-year winner.
The 12-month trend descriptor. Orthogonalized early so that the shorter-horizon momentum factors capture only incremental, faster signals.
Residual Volatilityresidual_volatilityrisk
a blend of short-horizon realized volatility and a longer-horizon, beta-residualized volatility
Higher = more volatile.
Residualized against the factors ranked above it (international, beta, size, long-term momentum), isolating idiosyncratic volatility from market and size effects.
Liquidityliquiditystyle
trailing average daily dollar volume divided by market capitalization (turnover)
Higher = more turnover relative to size.
A turnover proxy. Realized a modestly positive, regime-dependent return over the sample, almost all of it in 2020–2021 — see §9.
Growthgrowthstyle
composite of revenue growth, multi-year revenue CAGR, and sustainable growth
Higher = faster fundamental growth.
Component weights re-balance when an input is unavailable so that present components always carry the full weight.
Valuevaluestyle
negated composite of standardized P/E, P/B, P/CF, P/S and EV/EBITDA multiples
Higher = cheaper (high book/price, earnings yield).
Sign is flipped so that high loadings denote cheap names. A missing multiple drops out and the remaining weights are rescaled; a name with no multiple at all is neutral. Negative multiples count as expensive. Negative-equity names drop the book term and up-weight cash flow; price-to-sales is not used for banks and brokers.
Leverageleveragestyle
composite of market leverage, debt-to-assets, and net-debt-to-EBITDA
Higher = more levered.
Blends a market-based and two book-based leverage measures.
Profitabilityprofitabilitystyle
composite of ROIC, ROA, ROE, gross profitability, and operating profitability
Higher = more profitable.
Negative-equity names shift weight from ROE toward ROIC, which is well-defined when equity is impaired.
Dividend Yielddividend_yieldstyle
trailing dividends per share divided by price
Higher = higher payout yield.
Hard-floored at zero (no negative yields); standardized loading is clipped asymmetrically given the long right tail.
Short Interestshort_interestpositioningdynamics
reported short interest as a fraction of shares outstanding, standardized within listing segment
Higher = more crowded short (relative to its own listing segment).
A positioning descriptor, built strictly point-in-time: each semi-monthly FINRA report enters the model only after its public dissemination (≈ 8 business days after the settlement date). Its realized premium is modestly negative (t ≈ −2), far weaker than the short-sale flow factor — see §9.
Hedge-Fund Ownershiphfownpositioningdynamics
13F crowding composite: within-size-tercile percentile ranks of holder breadth and ownership magnitude
Higher = more hedge-fund-crowded.
Built from quarterly regulatory 13F institutional-holdings disclosures; absence from the panel is treated as genuinely uncrowded (zero), not missing.
Medium-Term Momentummedium_term_momentumflow
\(r_{20d} - r_{5d}\) (one-month less one-week return)
Higher = stronger 1-week-to-1-month trend.
A return spread, recomputed from the winsorized horizon returns for internal consistency.
Short-Term Momentumshort_term_momentumflow
\(r_{5d} - r_{1d}\) (one-week less one-day return)
Higher = stronger multi-day trend ex-yesterday.
Isolates the multi-day move after stripping the most recent session.
One-Day Momentumone_day_momentumflow
prior-session return \(r_{1d}\)
Higher = stronger prior-day return.
Captures short-horizon reversal/continuation; deliberately last among the momentum block.
Short-Sale Activityshort_saleflowdynamics
a short-sale velocity score from threshold-list status and short volume
Higher = elevated short-sale flow.
A flow descriptor, built close-safe: the daily short-sale file publishes after the close that starts the return-measurement window, so only the most recent close-tradeable file is used. Nearly uncorrelated with short interest (≈ 0.06); it is the flow, not the stock of short interest, that carries the model's strongest negative premium (§9).
Morning Activitymorning_activityflow
trailing average rank of morning trading intensity — morning-session volume scaled by the stock's recent average daily volume, relative to the rest of the session
Higher = more morning-concentrated, volume-heavy mornings.
A microstructure flow descriptor — names whose morning-session dollar volume runs high relative to their own recent norm and to the rest of the session (an intensity measure, not a simple morning share of daily volume).
Semiconductorssemisthematic
rolling time-series beta to a liquid semiconductor sector ETF
Higher = more exposed.
Highest-priority theme after market beta; semis carry a large, distinct factor footprint.
Oiloilthematic
rolling time-series beta to a liquid crude-oil commodity ETF
Higher = more exposed.
Energy / commodity beta.
Treasury (Duration)treasurythematic
rolling time-series beta to a long-dated Treasury bond ETF
Higher = more exposed.
Rates / duration sensitivity.
Goldgoldthematic
rolling time-series beta to a physical-gold ETF
Higher = more exposed.
Precious-metal / real-asset beta.
Bitcoin / Cryptobitcointhematic
rolling time-series beta to a liquid bitcoin-linked ETF
Higher = more exposed.
Crypto beta; relevant for miners, exchanges, and treasury-holders.
Chinachinathematic
rolling time-series beta to a China large-cap equity ETF
Higher = more exposed.
China large-cap equity beta.

Appendix B: Notation

SymbolMeaning
\(N,\;K\)number of securities in the daily universe; number of factors
\(m_i,\;w_i\)market cap; cap weight \(w_i=\sqrt{m_i}/\sum_j\sqrt{m_j}\)
\(r_{i,t}\)daily total return of security \(i\)
\(z(\cdot)\)cross-sectional standardization to mean 0, std 1
\(\operatorname{winsor}_p,\;\operatorname{cap}_c\)percentile winsorizer; symmetric hard cap
\(a,\;k\)loading clip (Stage 2); orthogonal-loading clip in cap-weighted \(\sigma\) (Stage 3)
\(\mathbf{X},\;\mathbf{W}\)exposure matrix; diagonal cap-weight matrix
\(u^{(j)},\;\beta_{ij}\)\(j\)-th orthogonalized loading; Gram–Schmidt projection coefficient
\(\mathbf{B}\)pure-factor portfolio matrix \((\mathbf{X}^{\!\top}\mathbf{W}\mathbf{X})^{-1}\mathbf{X}^{\!\top}\mathbf{W}\)
\(f_{f,t},\;\mathbf{F}\)factor return; factor-return covariance matrix
\(\boldsymbol{\Sigma},\;\boldsymbol{\Delta}\)security return covariance; diagonal specific-variance matrix
This document describes methodology only. Calibration constants (\(w_{\bullet}, a, k\), winsorization magnitudes, window lengths) are proprietary. Figure 1 is schematic; Figure 2 and the §9 statistics are measured on the production model (daily sample, Jan 2020 – Nov 2026, gross of costs).

FactorPulse is free during launch · Sign up free (no credit card)