💰 quantitative finance

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

This paper introduces a bridge-alpha curve diagnostic to evaluate factor models across entire characteristic axes rather than isolated deciles, revealing that while value and investment factors exhibit systematic pricing errors and sign reversals in CRSP data from 1967 to 2024, profitability and momentum factors largely flatten their axes, with these pricing errors proving nearly orthogonal to maximum-Sharpe gains.

Original authors: Useong Shin

Published 2026-07-08

📖 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 the stock market as a giant, chaotic library. For decades, financial experts have tried to organize this library using a few simple "shelves" or "categories" (called factor models) to explain why some books (stocks) are more expensive or return more money than others. The big question is: Do these shelves actually work, or do they just look good on the surface?

This paper introduces a new, ultra-precise way to check if these shelves are truly holding up the books correctly. Instead of just checking a few random books, the author uses a method called the "Characteristic-Axis Integral Diagnostic."

Here is the breakdown in simple terms:

1. The Problem: The "Blunt Instrument"

Usually, to test if a model works, experts pick a few specific groups of stocks (like the top 10% of cheap stocks vs. the bottom 10%) and see if the model predicts their returns.

The Analogy: Imagine trying to test if a new paint color looks good on a wall. The old method is like painting just three small squares on the wall and saying, "Well, these three squares look fine, so the whole wall must be perfect."
The Flaw: You might miss a weird streak of bad paint right in the middle, or a patch that looks great but is actually the wrong shade. The old tests depend too much on where you decide to cut the wall into squares.

2. The Solution: The "Bridge" Test

The author proposes a new method that looks at the entire wall, not just a few squares.

The Setup: Pick one characteristic, like "Value" (how cheap a stock is). Line up every single stock from the cheapest to the most expensive.
The Bridge: Imagine building a bridge at every single point along that line. At any point, you take the "prefix" (all the stocks cheaper than that point) and subtract the "average" performance of the whole group.
The Result: This creates a Bridge-Alpha Curve. Think of this curve as a map of the "pricing errors."
- If the curve is a flat line at zero, the model is perfect. It prices every single stock correctly along that line.
- If the curve dips below zero, the model is "overcorrecting" (it thinks cheap stocks are too cheap and punishes them too hard).
- If the curve shoots up, the model is "undercorrecting" (it isn't giving enough credit to the cheap stocks).

3. The Findings: The "Overcorrection" Trap

The author tested four major categories: Value (cheap stocks), Profitability (companies that make money), Investment (companies that grow assets), and Momentum (stocks that are already going up).

Here is what they found, using the "Bridge" map:

The "Undercorrection" (The Empty Shelf): Models that don't have a specific factor for a category (like a model without a "Value" factor) leave a positive curve. They simply miss the boat; they don't explain why cheap stocks perform well.
The "Overcorrection" (The Broken Shelf): This is the paper's big surprise. When models do include the right factor (like adding a "Value" factor to the model), they don't just fix the problem—they often break it in the opposite direction.
- Value & Investment: The famous "Fama-French" models (which include Value and Investment factors) actually overcorrect. They push the curve so far below zero that they create new pricing errors. It's like trying to fix a crooked picture frame by tilting it so hard it falls off the wall the other way.
- Profitability & Momentum: In contrast, models with Profitability and Momentum factors (like FF5 and Carhart) actually flatten the curve. They get it just right, leaving the line close to zero.

4. The "Fingerprint" Concept

The paper concludes that there is no single "best" model for everything.

The Analogy: Think of a model like a pair of glasses.
- Model A might give you perfect vision for reading (Value) but make you dizzy when looking at the horizon (Investment).
- Model B might be great for the horizon (Profitability) but blurry for reading.
The Result: A model that looks great on a standard test (like having a high "Sharpe Ratio," which measures overall efficiency) might still have huge, systematic errors on specific "axes" (categories). The paper shows that efficiency (making money) and accuracy (pricing correctly along a specific line) are two different things.

5. Why This Matters

The author argues that we need to stop looking at models as "Pass/Fail" based on a few test portfolios. Instead, we should look at their "Fingerprint."

Some models are "over-aggressive" (they overcorrect).
Some are "under-aggressive" (they undercorrect).
Some are "just right."

The paper proves that simply adding a factor to a model doesn't guarantee it works perfectly on that specific characteristic. Sometimes, the way the factor is built (the "construction") matters more than the factor itself. For example, a factor built exactly like the test (using the same accounting rules) works better than a famous, complex factor that tries to do the same thing but uses a different method.

In short: The paper gives us a new, high-resolution microscope to see exactly where and how financial models fail. It shows that the most popular models often "overcorrect" on Value and Investment, while doing a great job on Profitability and Momentum, proving that there is no "one-size-fits-all" solution in asset pricing.

Technical Summary: A Characteristic–Axis Integral Diagnosis of Factor Models

Problem Statement
Standard asset-pricing diagnostics, such as the Gibbons–Ross–Shanken (GRS) test or characteristic-sorted portfolio tests, evaluate whether a factor model prices a selected set of test assets. However, these tests often depend on arbitrary portfolio construction choices (e.g., decile binning) and may fail to reveal the specific location or shape of pricing errors along an economic coordinate. Conversely, spanning tests and maximum-Sharpe ratio comparisons assess the expansion of the investment opportunity set but do not identify where systematic pricing errors (alphas) remain within a restricted subspace. There exists a gap between evaluating a model's factor span and verifying zero-alpha pricing on economically meaningful, pre-specified characteristic coordinates. This paper addresses the need for a diagnostic that measures pricing errors as a continuous function along a characteristic axis rather than as a discrete set of portfolio alphas.

Methodology
The paper introduces a characteristic-axis integral diagnostic that generalizes a previous "cap-axis" framework to arbitrary firm characteristics (e.g., book-to-market, profitability, investment, momentum).

Bridge Construction:
- Stocks are sorted ex ante by a specific characteristic $x$ (e.g., descending order of book-to-market).
- For any cutoff $p \in [0, 1]$ , a "prefix portfolio" is formed comprising the top $p$ fraction of the sorted universe, weighted by market capitalization.
- A "bridge" return $D^x_t(p)$ is constructed by subtracting an equal-exposure aggregate portfolio ( $p \times R^{A,x}_t$ ) from the prefix portfolio. This creates a zero-investment portfolio that isolates the deviation of the prefix from the aggregate.
- The bridge is closed at endpoints ( $p=0$ and $p=1$ ), ensuring $D^x_t(0) = D^x_t(1) = 0$ .
Bridge-Alpha Curve:
- The bridge return is regressed on a factor model $m$ at each cutoff $p$ :
  $D^x_t(p) = \alpha^x_m(p) + \beta^x_m(p)'f_{m,t} + \varepsilon^x_{m,t}(p)$
- The resulting function $p \mapsto \alpha^x_m(p)$ is the bridge-alpha curve.
- The null hypothesis is a zero-curve restriction: $\alpha^x_m(p) = 0$ for all $p \in [0, 1]$ . This implies the model prices every prefix, tail, interval, and step-function portfolio generated by the characteristic order.
Functional Summaries:
- Instead of pointwise tests, the curve is summarized by four functionals:
  - Signed Area (SA): $\int \alpha^x_m(p) dp$ , measuring directional distortion.
  - Integrated Absolute Error (IAE): $\int |\alpha^x_m(p)| dp$ , measuring total distortion without sign cancellation.
  - Integrated Squared Error (ISE): $\int \alpha^x_m(p)^2 dp$ , weighting large local errors.
  - Supremum Norm (SUP): $\sup |\alpha^x_m(p)|$ , recording the maximum local error.
- The Signed Area is equivalent to the alpha of a single "rank-area portfolio" with linear weights $(1/2 - u)$ .
Inference:
- Inference for SA uses standard HAC t-tests on the rank-area portfolio.
- Inference for IAE, ISE, and SUP (nonlinear functionals) utilizes a finite-grid HAC-Gaussian-process approximation. The long-run covariance of the bridge-alpha estimator is estimated, and critical values are derived via simulation (or block-bootstrap) under the zero-curve null.

Key Empirical Results
The diagnostic is applied to CRSP data (1967–2024) across four axes: Value (HML), Operating Profitability (RMW), Investment (CMA), and Momentum (UMD).

Sign Reversal and Overcorrection: A recurring pattern emerges. Models lacking the counterpart factor (e.g., CAPM on the Value axis) tend to leave positive bridge-alpha curves (undercorrection). Adding the counterpart factor shifts the curve downward. However, this shift often results in overcorrection, pushing the curve significantly below zero.
- Value Axis: Models with HML (FF3, FF5, FF6) exhibit significant negative signed areas and high IAE, indicating they overcorrect the value premium. q5, lacking an explicit value factor, leaves a near-zero curve.
- Investment Axis: Models with investment factors (FF5, FF6, q5) show significant negative overcorrection. The flattest model is Carhart (which lacks an investment factor), though its aggregate gate fails.
- Operating Profitability (OP) Axis: Unlike the other axes, models with profitability factors (FF5, FF6) successfully flatten the curve, leaving insignificant alphas. q5 shows slight negative overcorrection at the monthly frequency, but daily rejections disappear under Dimson lead-lag corrections, suggesting a timing artifact rather than a stable pricing error.
- Momentum Axis: This serves as a positive control. Models with UMD (Carhart, FF6) flatten the curve effectively without significant overcorrection. q5 also passes without UMD, likely due to its ROE/EG block absorbing momentum premia.
Orthogonality of Metrics: The paper finds that maximum-Sharpe gains ( $\Delta SR$ ) and axis-level pricing errors (IAE) are nearly orthogonal. Factors that significantly expand the opportunity set (e.g., q-EG) can leave large pricing errors on specific axes, while factors with modest $\Delta SR$ can produce flatter curves.
Model Fingerprinting: No single model dominates across all axes.
- q5: Passes Value and Momentum but is rejected on Investment.
- FF5/FF6: Flattest on OP but rejected on Value and Investment.
- Carhart: Stable on Investment and Momentum but rejected on Value.

Significance and Claims
The paper claims to provide a functional diagnostic that complements, rather than replaces, existing tests (GRS, spanning, Sharpe ratio). Its primary contributions are:

Resolution: It converts pricing errors from a discrete set of portfolio alphas into a continuous function, revealing the direction, magnitude, location, and local concentration of errors along a characteristic coordinate.
Restricted Pricing: It tests a specific, restricted subspace generated by a pre-specified characteristic. Passing an axis means the model prices all portfolios generated by that order, not that it is a valid Stochastic Discount Factor (SDF) for the entire market.
Factor Construction Sensitivity: The results suggest that pricing errors depend not just on the presence of a characteristic premium in a model, but on how that premium is implemented (e.g., 2x3 sorting vs. direct characteristic sorting). Overcorrection is attributed to a mismatch between factor construction (often size-neutralized) and the cap-weighted characteristic rank coordinate.
Distinct Evaluation Criteria: The paper establishes that maximizing the Sharpe ratio and minimizing axis-level distortion are separate objectives. A factor can improve the frontier while leaving systematic pricing errors on a specific characteristic axis.

The author emphasizes that the diagnostic is a "narrow but strong" statement about pricing on a restricted subspace. It does not claim to resolve the debate on whether characteristics proxy for risk or expected returns, but rather provides a high-resolution tool to measure where and how factor models fail to price specific economic coordinates. Robustness checks (factor replication, block-bootstrapping, and finite-sample calibration) confirm that the results are not artifacts of data construction or inference procedures.

1. The Problem: The "Blunt Instrument"

2. The Solution: The "Bridge" Test

3. The Findings: The "Overcorrection" Trap

4. The "Fingerprint" Concept

5. Why This Matters

More like this