💰 quantitative finance

A Cap-Axis Integral Diagnostic of Factor Models

This paper introduces a cap-axis integral diagnostic that evaluates factor models by analyzing pricing errors along the market-capitalization rank axis, revealing distinct insights into model performance and zero-alpha violations that traditional metrics like the Sharpe ratio fail to capture.

Original authors: Useong Shin

Published 2026-07-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Useong Shin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a chef trying to perfect a signature soup. You have a recipe (a financial model) that claims to predict how the soup will taste. To test it, you usually take a big spoonful from the middle of the pot, taste it, and say, "Yep, it tastes like the recipe says."

This paper argues that taking one big spoonful isn't enough. Just because the average taste is right doesn't mean the soup is perfect. Maybe the top layer is too salty, the bottom layer is too bland, and the middle is just right. If you only taste the middle, you miss the flaws in the rest of the pot.

Here is a simple breakdown of what the paper does, using that soup analogy:

1. The Problem: The "Average" Lie

In finance, researchers use "factor models" (recipes) to explain why stocks go up or down. They often check if a model works by seeing if it prices the entire stock market correctly.

The Paper's Insight: A model can be "right" on average (pricing the whole market) but still be "wrong" in specific, organized ways. It might consistently overprice small companies and underprice big ones, or vice versa. Standard tests often miss these hidden patterns because they look at the whole pot at once.

2. The New Tool: The "Straw" Test

The author invents a new diagnostic tool called the Cap-Axis Integral Diagnostic.

The Metaphor: Imagine the stock market is a giant tower of blocks, sorted from the biggest block (the largest companies) at the bottom to the tiniest block (the smallest companies) at the top.
The Test: Instead of tasting the whole tower, the author builds a "bridge." They take a slice of the tower (say, the bottom 30% of the biggest companies) and compare it to a "fair" slice of the whole market.
The Curve: They do this for every possible slice size (1%, 2%, 3%... all the way to 100%). This creates a wavy line (a curve) that shows exactly where the model gets the price wrong as you move from the biggest companies to the smallest.

3. What They Found: The "Time of Day" Mystery

The author tested this on 154 different financial models using data from 1967 to 2024. They found something surprising: The time of day matters.

The Daily "Glitch": When looking at daily data, one popular model (called q5) showed a massive, consistent error. It was like the soup tasted consistently too salty only when you tasted it every single morning. The error was so strong it looked like a straight, downward-sloping line.
- The Fix: The author realized this might be due to "lag." Small companies often react to news a few minutes or hours later than big companies. When the author adjusted for this delay (like waiting for the soup to settle), the daily error almost disappeared. It was just a timing issue, not a broken recipe.
The Monthly "Glitch": When looking at monthly data, a different set of models (the Fama-French family) showed a different kind of error. They had a "hump" of error in the middle of the tower (mid-sized companies). This error was weak during the day but became very clear when looking at the month as a whole.

The Takeaway: No single model is "the best." Some models have hidden errors that only show up on a daily clock, while others have errors that only show up on a monthly clock.

4. Why This Is Better Than Old Tests

Old tests are like checking a map with a low-resolution satellite image. You can see the continents (big trends), but you can't see the cities or the streets.

The Old Way: "Does this model work for the top 10% of companies? Yes/No."
The New Way: "Does this model work for the top 10%? Yes. But wait, it fails specifically for the companies at the 33% mark, and it succeeds again at the 50% mark."

The author shows that standard tests often hide these errors because they average them out. By looking at the "curve," they can pinpoint exactly where the model fails.

5. The "Size" Misunderstanding

A common worry is: "Isn't this just testing if the model likes big or small companies?"

The Answer: No. The author proved that you can have a model that loves big companies (a "size" factor) but still leaves a messy, wavy error line on this new test. The test measures the shape of the error, not just the direction. It's like measuring the texture of the soup, not just whether it's salty or sweet.

Summary

This paper gives financial researchers a new, high-resolution microscope. It shows that:

Models can be "right" on average but "wrong" in specific places.
Errors depend on time: Some models fail daily, others fail monthly.
Location matters: We can now see exactly which part of the market (big, medium, or small) is causing the problem, rather than just guessing.

It doesn't say one model is perfect; it just says we finally have a way to see exactly where the cracks are in the wall.

Technical Summary: A Cap–Axis Integral Diagnostic of Factor Models

Problem Statement
Standard asset pricing evaluation relies on criteria such as factor spanning, maximum-Sharpe-ratio comparisons, and joint alpha tests on finite sets of test assets. While a correctly specified Stochastic Discount Factor (SDF) implies both zero alphas and mean–variance efficiency, empirical factor models are low-dimensional approximations where these criteria can separate. A model may successfully expand the mean–variance frontier (improving the maximum Sharpe ratio) while leaving systematic pricing errors on economically meaningful subspaces that standard tests compress into a single scalar verdict or hide within aggregated portfolios. Existing tests often fail to reveal where pricing errors reside along the market's internal structure, particularly regarding the distribution of errors across the market-capitalization axis.

Methodology
The paper proposes a cap-axis integral diagnostic that lifts factor-model pricing errors into a continuous function along the market-capitalization rank axis.

Construction of the Bridge:
- Assets are sorted by descending market capitalization.
- For any cumulative market-value share $p \in [0, 1]$ , a zero-investment "bridge" portfolio is formed: long the top- $p$ prefix of the market and short an equal market-value exposure of the aggregate market.
- This bridge, $D_t(p)$ , is zero at both endpoints ( $p=0$ and $p=1$ ) by construction.
The Bridge-Alpha Curve:
- The bridge return is regressed on a factor model $f_{m,t}$ to estimate an alpha function $\alpha_m(p)$ .
- The resulting curve, $p \mapsto \alpha_m(p)$ , represents the pricing error distribution along the cap-rank axis.
- Under an "aggregate-market gate" (requiring the model to price the aggregate market with zero alpha), a zero bridge-alpha curve is necessary and sufficient for the model to price the closed linear subspace generated by the aggregate market and all cap-axis bridges.
Functional Statistics:
The curve is summarized by four functionals to separate direction, magnitude, and local distortion:
- Signed Area ( $\widehat{SA}$ ): Measures directional tilt; equivalent to the alpha of a single "rank-area" portfolio with weights $(1/2 - u)$ .
- Integrated Absolute Error ( $\widehat{IAE}$ ): Measures total distortion magnitude without sign cancellation.
- Integrated Squared Error ( $\widehat{ISE}$ ): Measures total squared distortion.
- Supremum ( $\widehat{SUP}$ ): Records the largest local distortion.
Inference and Placebos:
- HAC Inference: Tests if the signed area is statistically significant relative to time-series uncertainty (Newey–West standard errors).
- Ordering Placebos: A matched benchmark where dates, returns, and market weights are fixed, but the stock ordering is randomized. This tests whether the observed bridge shape is specific to capitalization rank rather than a general feature of the return distribution.
- Lead–Lag Correction: Applied to daily data to address non-synchronous trading, aggregating factor leads and lags to correct for beta attenuation.

Key Results
The diagnostic is applied to CRSP investable market data from 1967–2024 across 154 factors and several standard models (CAPM, Fama–French, Carhart, q5).

Frequency-Dependent Distortions:
- Daily Frequency: The q5 model exhibits a robust, one-sided negative bridge-alpha curve (signed area $\approx -39.6$ bp). This distortion is highly significant and passes ordering placebos. However, it attenuates rapidly under lead–lag corrections, suggesting it is driven by short-horizon non-synchronous trading.
- Monthly Frequency: The q5 distortion nearly vanishes. Conversely, Fama–French (FF3, FF5, FF6) and Carhart models display a significant positive, one-sided bridge (signed area $\approx +32$ bp) that is stronger at the monthly frequency than daily.
- Conclusion: Pricing errors are horizon-localized; different models leave errors at different frequencies.
Factor-Level Coordinates:
- Across 154 factors, the cap-axis footprint (measured by $\widehat{IAE}$ ) is distinct from the maximum-Sharpe-ratio gain ( $\Delta SR$ ). The rank correlation is low (0.15).
- A factor can significantly improve the Sharpe frontier while leaving a large zero-alpha footprint on the cap-rank subspace (e.g., q5 Expected Growth has the highest Sharpe gain and the largest footprint).
- The footprint is not mechanically explained by conventional size exposure. Regressing the footprint magnitude on size-factor loadings yields near-zero $R^2$ , indicating the diagnostic captures a distinct feature of factor behavior beyond simple size exposure.
Resolution vs. Size Bins:
- Standard size-decile tests compress the market into 10 bins, with the top decile holding ~60% of market value. This aggregation hides internal structure.
- The cap-axis diagnostic resolves the top decile into a continuous path, revealing that pricing errors often peak inside the dominant bin (e.g., at $p \approx 0.33$ or $0.45$) rather than at the bin boundaries. A model might pass a decile test while failing the cap-axis diagnostic due to hidden internal errors.

Significance and Claims
The paper claims to add a restricted zero-alpha complement to standard model evaluation tools.

Distinctness: It demonstrates that aggregate-market fit, maximum-Sharpe improvement, and restricted zero-alpha consistency on the cap-rank subspace are distinct empirical criteria. A model can pass the aggregate market and expand the frontier while failing the cap-axis diagnostic.
Localization: Unlike scalar tests, the diagnostic identifies the direction, magnitude, location, and frequency of pricing errors along a fixed market-internal coordinate.
Non-Substitutability: It does not replace spanning tests or size-bin tests but reveals information they compress. It shows that "approximate SDFs" can place their pricing errors at different horizons (daily vs. monthly) and that model-level consistency can arise from offsetting marginal footprints of constituent factors.
Utility: The bridge-alpha curve serves as a functional object that can be used for factor construction (minimizing IAE) or decomposing factor blocks, offering a high-resolution view of pricing errors that standard tests obscure.

The paper remains modest in its scope, acknowledging that passing the diagnostic does not imply full SDF validity (as it ignores momentum, value, and other non-cap-rank directions), and failing it does not imply a lack of mean–variance value. Its primary contribution is the measurement device itself.