Any Axes Are Allowed:
A Characteristic–Axis Integral Diagnosis of Factor Models

Let $x$ be a stock-level characteristic observed at formation date $\tau(t)$, and let $\mathcal{I}^{x}_{\tau(t)}$ denote the set of stocks for which $x$ is valid. This set need not be the full CRSP universe; for accounting-based characteristics it is the valid-characteristic subuniverse with observed book equity, operating profitability, or investment. The key requirement is that $\mathcal{I}^{x}_{\tau(t)}$ is defined ex ante and uses no return information after the formation date.

Sort stocks in descending order of $x_{i,\tau(t)}$, and assign each valid stock a strictly positive formation-date weight $\pi^{x}_{i,\tau(t)}$ that uses no post-formation return information. A natural family is the power-of-cap weight

$$\pi^{x}_{i,\tau(t)}\;\propto\;ME_{i,\tau(t)}^{\rho},\qquad\rho\geq 0,$$

(1)

which nests value weighting ($\rho=1$) and equal weighting ($\rho=0$) and admits intermediate choices such as $\rho=\tfrac{1}{2}$ or $\rho=2$. Normalize the weights to sum to one and traverse the stocks in the fixed characteristic order. In the continuous limit, the sorted axis is $u\in[0,1]$ and $d\mu(u)$ is the induced formation-date measure along that order, so $du$ denotes cumulative $\mu$-mass rather than wealth share alone. The order is fixed during the holding period, while the traversal weights are carried forward each trading day as start-of-day buy-and-hold shares. In the cap axis, both the ordering variable and the measure are market equity ($\rho=1$); in a general axis, $x$ fixes the order while $\mu$ fixes the measure. Cutoff $p$ is therefore not a characteristic quantile but cumulative $\mu$-mass along the characteristic order.

Let $r^{x}_{t}(u)$ be the return of the infinitesimal portfolio at rank $u$. The characteristic-valid aggregate return is

$$R^{A,x}_{t}=\int_{0}^{1}r^{x}_{t}(u)\,du.$$

(2)

For the cap measure ($\rho=1$), $R^{A,x}_{t}$ is close to the investible market return; under a general measure it is the $\mu$-weighted aggregate return of the valid subuniverse—still an implementable buy-and-hold portfolio, but no longer the market portfolio. The pricing statement below is thus conditional on the selected characteristic, its valid universe, and the chosen measure.

For cutoff $p\in[0,1]$, define the prefix contribution as

$$C^{x}_{t}(p)=\int_{0}^{p}r^{x}_{t}(u)\,du,$$

(3)

and subtract an equal aggregate exposure $pR^{A,x}_{t}$ to obtain the characteristic-axis bridge:

$$D^{x}_{t}(p)=C^{x}_{t}(p)-pR^{A,x}_{t}.$$

(4)

The bridge is closed at both endpoints,

$$D^{x}_{t}(0)=D^{x}_{t}(1)=0,$$

(5)

and the tail bridge is the negative of the prefix bridge,

$$\int_{p}^{1}r^{x}_{t}(u)\,du-(1-p)R^{A,x}_{t}=-D^{x}_{t}(p),$$

(6)

so a single prefix-bridge curve records the body–tail offset at every cutoff.

For factor model $m$ with factor vector $f_{m,t}$, estimate

$$D^{x}_{t}(p)=\alpha^{x}_{m}(p)+\beta^{x}_{m}(p)^{\prime}f_{m,t}+\varepsilon^{x}_{m,t}(p).$$

(7)

The function $p\mapsto\alpha^{x}_{m}(p)$ is the characteristic-axis bridge-alpha curve, and the zero-curve null is

$$H^{x}_{0}:\alpha^{x}_{m}(p)=0\quad\text{for all }p\in[0,1].$$

(8)

The bridge-alpha curve is a function, not a scalar. I summarize it by four functionals:

	$\displaystyle SA^{x}_{m}$	$\displaystyle=\int_{0}^{1}\alpha^{x}_{m}(p)\,dp,$		(9)
	$\displaystyle IAE^{x}_{m}$	$\displaystyle=\int_{0}^{1}\left|\alpha^{x}_{m}(p)\right|\,dp,$		(10)
	$\displaystyle ISE^{x}_{m}$	$\displaystyle=\int_{0}^{1}\alpha^{x}_{m}(p)^{2}\,dp,$		(11)
	$\displaystyle SUP^{x}_{m}$	$\displaystyle=\sup_{p\in[0,1]}\left|\alpha^{x}_{m}(p)\right|.$		(12)

$SA$ measures directional pricing error. $IAE$ measures total distortion without sign cancellation. $ISE$ gives more weight to large local distortions, and $SUP$ records the largest absolute pricing error on the axis. Reversing the sort direction changes only the sign of $SA$; $IAE$, $ISE$, and $SUP$ are unchanged. The descending-order convention therefore affects the economic interpretation of the signed area, not rejection.

The signed area can be implemented as the alpha of a single zero-investment portfolio. Integrating (4) over $p$ gives

	$\displaystyle\int_{0}^{1}D^{x}_{t}(p)\,dp$	$\displaystyle=\int_{0}^{1}\left\{\int_{0}^{p}r^{x}_{t}(u)\,du-pR^{A,x}_{t}\right\}dp$
		$\displaystyle=\int_{0}^{1}\left(\frac{1}{2}-u\right)r^{x}_{t}(u)\,du.$		(13)

Define the rank-area return as

$$Y^{x}_{t}=\int_{0}^{1}\left(\frac{1}{2}-u\right)r^{x}_{t}(u)\,du.$$

(14)

Then

$$\alpha_{m}(Y^{x})=\int_{0}^{1}\alpha^{x}_{m}(p)\,dp=SA^{x}_{m}.$$

(15)

Thus $SA$ is the alpha of a single long–short portfolio with linear rank weights $\tfrac{1}{2}-u$. This portfolio tilts long toward the top of the axis and short toward the bottom, with intensity proportional to rank.

Because $u$ is cumulative $\mu$-mass, the same measure governs both the aggregate weighting and the rank coordinate. The rank-area portfolio therefore remains a single tradeable long–short position, but its weights depend on $\mu$: changing the measure (for example from $\rho=1$ to $\rho=0$) reweights the position without altering the construction. Equal- and value-weighted implementations are thus two instances of the same object rather than separate tests.

The key interpretation is restricted pricing. For a fixed admissible measure, define the closed linear subspace generated by axis $x$ as

$$\mathcal{V}_{x}=\overline{\mathrm{span}\left\{R^{A,x},D^{x}(p):p\in[0,1]\right\}}.$$

(16)

Proposition 1 fixes the scope of the diagnostic. If the aggregate gate and the zero curve both hold, the model prices every prefix, tail, interval, and step-function portfolio generated by the axis. In this sense, the test is complete within the selected characteristic axis. It is not a test of full SDF validity. It does not price industry, other-anomaly, or idiosyncratic long–short directions outside the axis. If the valid universe is smaller than the market, the conclusion is restricted to that universe.

The empirical implementation replaces the continuous axis with a finite stock universe and a finite grid. At the formation date I sort stocks with valid characteristic $x$. Under the canonical cap measure ($\rho=1$) the formation-date weights coincide with buy-and-hold wealth shares $w^{x}_{i,t}$; a general admissible measure replaces $w^{x}_{i,t}$ throughout by the normalized weights $\pi^{x}_{i,t}$. Let $w^{x}_{i,t}$ be the start-of-day buy-and-hold wealth share of stock $i$ on holding day $t$:

$$\sum_{i=1}^{N^{x}_{t}}w^{x}_{i,t}=1,\qquad R^{A,x}_{t}=\sum_{i=1}^{N^{x}_{t}}w^{x}_{i,t}R_{i,t}.$$

(19)

Here $R^{A,x}_{t}$ is the aggregate return of the valid subuniverse for axis $x$—the investible market return in the cap axis, and the aggregate return of stocks with valid accounting or momentum information in the other axes.

Define cumulative wealth shares and midpoint ranks by

$$c^{x}_{i,t}=\sum_{j=1}^{i}w^{x}_{j,t},\qquad\mu^{x}_{i,t}=c^{x}_{i,t}-\frac{1}{2}w^{x}_{i,t}.$$

(20)

For target cutoff $p$, let $k_{t}(p)=\min\{i:c^{x}_{i,t}\geq p\}$ be the last stock in the whole-stock prefix and $s_{t}(p)=c^{x}_{k_{t}(p),t}$ the realized prefix exposure. The finite-sample bridge return is

$$D^{x}_{t}(p)=\sum_{i=1}^{k_{t}(p)}w^{x}_{i,t}R_{i,t}-s_{t}(p)R^{A,x}_{t},$$

(21)

which longs the realized prefix and shorts the same realized aggregate exposure. Since the boundary stock is not split, $s_{t}(p)$ need not equal $p$, but endpoint closure at $p=0$ and $p=1$ remains exact.

The signed-area statistic is implemented by the midpoint rank-area return

$$Y^{x}_{t}=\sum_{i=1}^{N^{x}_{t}}w^{x}_{i,t}R_{i,t}\left(\frac{1}{2}-\mu^{x}_{i,t}\right),$$

(22)

the finite-stock approximation to $\int_{0}^{1}D^{x}_{t}(p)\,dp$. Because the grid bridge uses whole-stock prefixes while the rank-area return uses midpoint weights, the two need not coincide exactly in finite samples; I test $SA$ through the rank-area alpha and record the gap between the grid-integrated signed area and the rank-area alpha as boundary approximation error.

On a grid $P=\{p_{1},\ldots,p_{L}\}$, I construct bridge returns and estimate

$$D^{x}_{t}(p_{l})=\alpha^{x}_{m}(p_{l})+\beta^{x}_{m}(p_{l})^{\prime}f_{m,t}+\varepsilon^{x}_{m,t}(p_{l})$$

(23)

at each cutoff, collecting the intercepts into

$$\widehat{\bm{\alpha}}^{x}_{m}=\left(\widehat{\alpha}^{x}_{m}(p_{1}),\ldots,\widehat{\alpha}^{x}_{m}(p_{L})\right)^{\prime}.$$

(24)

Applying numerical integration weights and a grid maximum to this vector gives the sample counterparts of $SA$, $IAE$, $ISE$, and $SUP$.

Inference treats $\widehat{\bm{\alpha}}^{x}_{m}$ as a finite-dimensional estimator process on the grid. Under standard weak-dependence and moment conditions,

$$\sqrt{T}\left(\widehat{\bm{\alpha}}^{x}_{m}-\bm{\alpha}^{x}_{m}\right)\Rightarrow N(0,\Omega^{x}_{m}),$$

(25)

where $\Omega^{x}_{m}$ is the long-run covariance matrix of the bridge-alpha process. I test $SA$ with the HAC alpha $t$-test of the rank-area portfolio. Since $IAE$, $ISE$, and $SUP$ are nonlinear, nonnegative functionals of the curve, I do not assign them pointwise $t$-statistics; instead, nonlinear functional inference uses a finite-grid HAC-Gaussian-process approximation. I estimate the HAC long-run covariance matrix $\widehat{\Omega}^{x}_{m}$ of the intercept process and, under the zero-curve null, repeatedly draw

$$Z^{x}_{m}\sim N(0,\widehat{\Omega}^{x}_{m})$$

(26)

and apply the same integration and grid maximum to each draw, producing simulated null distributions for $IAE$, $ISE$, and $SUP$. The reported p-value is the upper-tail probability of the observed functional in this null. This Gaussian approximation is not an assumption that raw returns are normal; it approximates the sampling distribution of the grid-level bridge-alpha estimator process, with serial correlation and heteroskedasticity entering through $\widehat{\Omega}^{x}_{m}$ and the nonstandard functional distributions handled by simulation.

The p-values are per-axis statements: even when several axes are tested, I do not combine them into a joint zero-curve null, but read the direction, magnitude, and local distortion of each model–axis curve as a model-by-axis fingerprint.

Finite-sample performance is checked in a separate calibration. Holding the empirical monthly bridge-return panel and factor matrix fixed as the sampling environment, I simulate the zero-curve null and structured axis-local alternatives of strength $\lambda\in\{0,0.25,0.5,0.75,1\}$, where $\lambda=0$ is the size case and positive $\lambda$ scales a systematic axis-local bridge-alpha curve. For each simulated sample I recompute the $IAE$, $ISE$, and $SUP$ tests and record 5% rejection frequencies under both HAC-GP and block-bootstrap critical values. This asks whether the tests are near nominal size under the null and whether rejection rises against structured axis alternatives.

Passing the diagnostic is a narrow but strong statement. If a model prices the axis aggregate portfolio and has a zero bridge-alpha curve at all cutoffs, no systematic alpha remains in any prefix, tail, interval, or step-function portfolio generated by that order. Passing the value axis does not merely mean HML alpha is small; it means that, within the book-to-market order, any cutoff-defined portfolio is priced. Where a conventional decile test is a low-resolution inspection, the integral diagnostic traces the full pricing-error path by moving the cutoff continuously. The statement remains local: it implies neither a valid SDF nor pricing of all anomalies or industry directions, and with a valid subuniverse it is a statement about that universe. A pass therefore means not that the model explains the market, but that it leaves no visible pricing-error structure along the selected characteristic coordinate.

The diagnostic is also distinct from mean–variance performance. For a candidate factor $g_{t}$, define the maximum-Sharpe gain as

$$\Delta SR(m,g)=SR(m+g)-SR(m).$$

(27)

$\Delta SR$ asks how much the factor expands the opportunity set; $IAE^{x}_{m}$ asks how much zero-alpha violation remains on a given axis. Neither implies the other: a factor can greatly increase $\Delta SR$ while leaving a bridge-alpha curve unchanged, and another can add little to the frontier while sharply reducing axis distortion. I treat the two as complementary, not competing, model-evaluation criteria.

The sample consists of NYSE, AMEX, and NASDAQ common stocks from January 3, 1967 to December 31, 2024. The universe screen preserves nearly all economic scale and trading activity while removing the extreme microcap and illiquidity tails. Investibility is updated at each month-end using market capitalization, $ME=|PRC|\times SHROUT\times 1{,}000$, and liquidity, the average dollar volume over the previous 63 trading days. An incumbent exits if its cumulative market-capitalization share exceeds 99.9% or its liquidity falls below the monthly 2.5th percentile; an outside stock enters only if its cumulative share is within 99.5% and its liquidity is at or above the fifth percentile. This hysteresis reduces boundary churning. For early NASDAQ observations the liquidity screen is deferred until 63 trading days of volume history are available. All screening variables use information available by month-end, and the month-end universe is applied from the first trading day of the following month.

Returns are adjusted CRSP daily returns combining total and delisting returns: when both are observed they are compounded, and when only the delisting return is observed that return is used, with no additional negative correction for missing delisting returns. Positions with missing total returns are liquidated at their last observed value and reinvested proportionally into surviving names in the same portfolio. Axis-portfolio membership is fixed at the formation date; stocks are not forced out during the holding period because of later universe updates. The final universe contains, on average, 77.6% of the common-stock base sample while preserving about 99.7% of total market capitalization and 63-day dollar volume; at the end of 2024 it includes 2,426 of 3,804 common stocks and preserves 99.78% of market capitalization and 99.28% of dollar volume. Since the screen uses only market capitalization and liquidity, it does not directly use any anomaly characteristic.

Because characteristic-axis bridges are benchmarked to aggregate portfolios built directly from CRSP stocks, I first check whether market-return implementation could generate alphas. The daily value-weighted buy-and-hold market return from the investible universe, $cMKT$, correlates 0.999778 with the Fama–French market return and 0.999799 with the q5 market return over the common sample, with daily RMSEs of 2.21 bp and 2.10 bp. Regressing the full market return on each candidate model gives $R^{2}$ above 0.999, and Newey–West 21-day alphas are insignificant at the 5% level, so the axis results are unlikely to be artifacts of market-factor implementation.

The candidate models are CAPM, FF3, Carhart, FF5, FF6, and q5. FF6 adds momentum to FF5; q5 uses the provider’s market, size, investment, return-on-equity, and expected-growth factors. The main frequency is monthly, with daily results as companion diagnostics used when signs or significance differ from the monthly baseline or when high-frequency implementation matters. Monthly tests use published monthly factor series rather than ex post aggregation of daily factors, and all series are audited for units, frequency, and sample coverage before estimation.

The empirical axes are value, operating profitability, investment, and momentum. Accounting-based characteristics are built from Compustat annual fundamentals and CRSP market equity, momentum from CRSP monthly returns, and all use only information observable at the formation date. For annual axes, June formation follows the standard Fama–French timing convention: financial statements for fiscal year $Y-1$ are assumed available by June of formation year $Y$. Book equity is stockholders’ equity plus deferred taxes and investment tax credit, less preferred stock, using the standard Compustat hierarchy for each component; missing preferred stock and deferred taxes are set to zero.

For the value axis, at the end of June in formation year $Y$, I compute

$$x^{BM}_{i,Y}=\log\left(\frac{BE_{i,Y-1}}{ME^{Dec}_{i,Y-1}}\right)$$

(28)

and sort stocks with positive book equity and positive December market equity from high to low book-to-market, holding from July of year $Y$ to June of $Y+1$.

For the operating profitability axis, I use

$$x^{OP}_{i,Y}=\frac{REV_{i,Y-1}-COGS_{i,Y-1}-XSGA_{i,Y-1}-XINT_{i,Y-1}}{BE_{i,Y-1}}$$

(29)

with missing cost items set to zero, including stocks only when revenue and positive book equity are observed, and sorting from high to low profitability.

For the investment axis, I require positive total assets in fiscal years $Y-1$ and $Y-2$ and compute

$$Inv_{i,Y}=\frac{AT_{i,Y-1}-AT_{i,Y-2}}{AT_{i,Y-2}},\qquad x^{INV}_{i,Y}=-Inv_{i,Y},$$

(30)

so descending order runs from conservative to aggressive investment; the holding period matches value and operating profitability.

The momentum axis uses monthly formation and next-month holding. At the end of formation month $m$, I compute the 12–2 cumulative return,

$$x^{MOM}_{i,m}=\prod_{s=2}^{12}\left(1+R_{i,m-s}\right)-1,$$

(31)

requiring at least 8 valid monthly returns in the 11-month window, and sort from winners to losers.

The baseline diagnostic is implemented on the valid-characteristic subuniverse, so the test asks whether the model prices the axis generated by the observed characteristic ranks without assigning missing-characteristic stocks to arbitrary ranks. This follows the same principle as Fama–French factor construction, which excludes firms with missing book equity. A full-universe version that places missing-characteristic stocks into a separate terminal bucket is produced as a robustness check, with similar main sign patterns.

Before testing the bridge curve, I estimate the aggregate gate

$$R^{A,x}_{t}-R^{f}_{t}=a^{A,x}_{m}+b^{A,x\prime}_{m}f_{m,t}+e^{A,x}_{m,t}.$$

(32)

If a model does not price the valid aggregate itself, a nonzero curve may reflect subuniverse alpha rather than the internal ordering—especially for accounting-based axes, where the valid subuniverse is the CRSP$\cap$Compustat intersection and aggregate alpha may reflect coverage selection. I therefore report aggregate-gate statistics separately and interpret curve results for gate-failing models conditionally.

The engine forms characteristic ranks at the formation date and freezes the order during the holding period. Daily weights are updated as start-of-day buy-and-hold wealth shares, normalized to sum to one within the valid subuniverse, and aggregate, bridge, and rank-area returns follow the finite-sample definitions in Section 3.4. The baseline grid uses 201 equally spaced cutoffs on $[0,1]$. Returns use dividend-inclusive adjusted returns with an ex-dividend-return dividend-pool accounting rule, reinvesting each day’s dividend cash flows into surviving stocks in proportion to their ex-dividend values. Each run verifies the body–tail recombination identity and the agreement between the numerically integrated bridge path and the rank-area return, up to whole-stock boundary approximation error. Daily bridge returns feed the companion diagnostics directly; monthly bridge returns are built from the same daily path, compounding long-only aggregate returns within each month and summing zero-investment bridge and rank-area returns, so the monthly bridge is a lower-frequency version of the same daily payoff.

For each model $m$, axis $x$, and cutoff $p$, I estimate the bridge regression (7). The main text reports monthly results, annualizing monthly intercepts by 12 into basis points with a baseline Newey–West lag of six months; daily companion diagnostics annualize by 252 with a baseline lag of 21 trading days, and alternative lags are audit items. When the monthly baseline and daily diagnostic agree qualitatively, I omit additional lead–lag tables to preserve space; when a model passes the monthly test but rejects daily, I apply a Dimson lead–lag correction to check whether the daily rejection reflects short-horizon non-synchronous exposure rather than a stable axis-level pricing error. This rule is relevant only for the operating-profitability axis, where q5 passes monthly but rejects in the baseline daily diagnostic.

The signed area is tested through the rank-area regression

$$Y^{x}_{t}=\alpha^{Y}_{m}+\beta^{Y\prime}_{m}f_{m,t}+\varepsilon^{Y}_{m,t},$$

(33)

for which I report the annualized alpha, HAC standard error, $t$-statistic, $p$-value, and $R^{2}$; the $R^{2}$ describes how much time-series variation in the axis-integrated payoff the model explains, while the alpha test asks whether that payoff has a nonzero mean pricing error. For the full curve I report $SA$, $IAE$, $ISE$, and $SUP$, testing $SA$ by the rank-area regression and the nonlinear functionals by the finite-grid HAC-Gaussian-process approximation of Section 3.4. The baseline simulation uses 50,000 antithetic draws with positive semidefiniteness imposed through an eigenvalue floor, and HAC-GP is the baseline pricing-null inference; I also compute block-bootstrap critical values for sensitivity to finite-sample dependence and resampling. In the size and power calibration I hold the empirical monthly bridge-return panel and factor matrix fixed, inject structured axis-local alternatives indexed by $\lambda\in\{0,0.25,0.5,0.75,1\}$ ($\lambda=0$ size, $\lambda>0$ power), and record 5% rejection frequencies.

The main empirical object is a model-by-axis fingerprint: for each model and axis I report the sign and significance of the rank-area alpha, the magnitudes of $IAE$, $ISE$, and $SUP$, and their finite-grid p-values, and I classify each model–axis pair by rank-area alpha significance and relative $R^{2}$ within the axis-frequency panel. This quadrant view is an interpretive aid, not a replacement for the bridge-functional tests.

I also compare the axis diagnostic with factor-level mean–variance gains. For each candidate factor $g$ in the monthly library, I add $g$ to CAPM and estimate

$$\Delta SR(g)=SR(CAPM+g)-SR(CAPM)$$

(34)

together with the corresponding axis functional. $\Delta SR$ measures expansion of the opportunity set while $IAE$ measures distortion remaining on the axis; this is not a factor-discovery exercise but a diagnostic showing that Sharpe improvement and axis pricing are distinct criteria. Each axis has a natural counterpart factor—HML for value, RMW or a profitability factor for operating profitability, CMA or IA for investment, and UMD for momentum—and the reducibility question is whether adding the representative factor removes the pricing-error structure on its own axis. Comparing the CAPM curve with the CAPM-plus-counterpart curve: if the signed-area alpha disappears and the functionals become insignificant, the distortion is reducible; if the curve crosses zero to the opposite sign, the factor absorbs the average premium but overcorrects on the cap-weighted rank coordinate.

The value axis sorts stocks each June by $\log(BE/ME^{Dec})$ from high to low book-to-market and holds that order from July through the following June. The value-valid subuniverse covers, on average, 81.1% of investible market capitalization after 1967 and 90.3% at the final formation date, so the test concerns the intersection for which the value characteristic is observed, not the entire market.

The operating-profitability axis sorts stocks each June by

$$(REVT-COGS-XSGA-XINT)/BE$$

from high to low OP and holds the order from July through the following June. The construction requires positive book equity and a valid profitability measure, with missing cost items treated as in Section 4.3. The OP-valid subuniverse covers, on average, 80.9% of investible market capitalization after 1967 and 88.6% at the final formation date.

The investment axis computes total asset growth each June, $INV=(AT_{Y-1}-AT_{Y-2})/AT_{Y-2}$, and uses $-INV$ as the sort key, so the axis runs from conservative to aggressive investment. I include only stocks with positive lagged and current assets, holding from July through the following June. The investment-valid subuniverse covers, on average, 79.2% of investible market capitalization after 1967 and 92.9% at the final formation date—the highest final coverage among the three accounting-based axes.