💰 quantitative finance

Anatomy of the Market: A Body-Tail Test of Factor Models

This paper demonstrates that while factor models may successfully price the aggregate market, they often fail to price its internal components, as evidenced by systematic, offsetting pricing errors between size-ranked "body" and "tail" portfolios that cancel out in the aggregate.

Original authors: Useong Shin

Published 2026-06-23

📖 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 taste a giant, complex stew. You take a big spoonful from the middle of the pot, taste it, and say, "This tastes perfect! The seasoning is spot on."

Now, imagine that same stew is actually made of two very different parts: a huge pile of potatoes (the "body") and a small handful of spicy peppers (the "tail"). If you taste the whole pot, the potatoes might be slightly too salty, and the peppers might be slightly too bland. But because the saltiness of the potatoes cancels out the blandness of the peppers, the average taste of the whole spoonful seems perfect.

This is exactly what the paper "Anatomy of the Market: A Body–Tail Test of Factor Models" by Useong Shin is investigating.

The Big Question

In the world of finance, experts use mathematical recipes (called "factor models") to predict how stocks will perform. These recipes usually include a "market factor," which is just a way of saying "the average performance of all stocks."

The author asks a simple but tricky question: If a recipe works perfectly for the whole pot of stew (the entire stock market), does it also work perfectly for the individual ingredients inside it?

The Experiment: The "Body" and the "Tail"

To test this, the author didn't just look at random stocks. He took the entire US stock market and split it into two distinct groups based on how big the companies are:

The Body: The massive, well-known companies (like Apple, Microsoft, Amazon). These make up the bulk of the market's value.
The Tail: The smaller, less famous companies.

Crucially, if you put these two groups back together, you get the exact same market you started with. It's like taking a puzzle apart and putting it back together; the picture is the same, but now you can look at the pieces individually.

The Surprise Discovery

The author tested several famous financial recipes (like the CAPM, Fama-French, and the newer "q5" model).

The Good News: Every single recipe passed the test for the whole market. The "average" error was zero. The stew tasted perfect.
The Bad News: When he looked at the Body and the Tail separately, one specific recipe (the q5 model) failed miserably.

Here is what happened with the q5 model:

It predicted the Body (big companies) would do worse than they actually did (a negative error).
It predicted the Tail (small companies) would do better than they actually did (a positive error).
The Magic Trick: Because the big companies are so huge, their negative error canceled out the small companies' positive error. The math added up to zero, making the whole market look "priced correctly" even though the model was wrong about the pieces.

The "Random Split" Check

You might think, "Maybe it's just hard to split a market in half; maybe any split causes these errors."

To check this, the author did a control test. He took the exact same stocks and split them into two groups completely at random (ignoring size).

Result: When the groups were random, the q5 model worked fine. The errors disappeared.
Conclusion: The problem wasn't the act of splitting the market. The problem was specifically how the model handled the size ranking (Big vs. Small). The model was confused by the specific structure of big companies versus small companies.

The Culprit: The "Growth" Ingredient

The author then played detective to find out which part of the q5 recipe was causing the trouble. He removed ingredients one by one.

He found that the trouble came from the model's focus on Expected Growth (EG) and Return on Equity (ROE).
Specifically, the "Expected Growth" factor seemed to be the main villain. It was pushing the model to misprice the big companies and small companies in opposite directions.

Why This Matters (In Simple Terms)

The paper teaches us a valuable lesson about looking at the big picture:

Just because the average looks good doesn't mean everything inside is correct.

In finance, a model can pass the "sanity check" of pricing the whole market while still being fundamentally broken when it comes to the specific parts of that market. The author calls this the "Body-Tail Test." It's a way to stress-test financial models to see if they are hiding errors by averaging them out, rather than actually understanding how different parts of the market work.

Summary

The Setup: Split the stock market into "Big Guys" (Body) and "Small Guys" (Tail).
The Test: See if financial models price the whole market and the two groups correctly.
The Result: Most models worked. The q5 model looked perfect for the whole market but was actually wrong for the pieces. It thought big stocks were overpriced and small stocks were underpriced, but the errors canceled out.
The Cause: The model's focus on "Expected Growth" was the specific ingredient causing the confusion.
The Lesson: Don't just trust the average. If a model prices the whole market, check if it also prices the parts correctly, or it might just be hiding its mistakes.

Technical Summary: Anatomy of the Market: A Body–Tail Test of Factor Models

Problem Statement
The paper addresses a fundamental limitation in standard factor model evaluation: the reliance on aggregate market pricing as a primary sanity check. Most linear factor models (e.g., CAPM, Fama–French, q-factor) include the market return as a core factor. Consequently, a model's ability to price the aggregate market portfolio is often a "built-in sanity check" rather than a demanding test of pricing ability. The author argues that a small or insignificant alpha for the aggregate market does not imply that the model correctly prices the market's internal components. Pricing errors within the market's constituent portfolios may have opposite signs and cancel out in the aggregate, masking internal inconsistencies. The central question is: If a factor model prices the aggregate market, does it also price the market's internal components?

Methodology
The author proposes a "body–tail test" to evaluate internal pricing consistency without introducing external anomaly portfolios.

Data Construction:
- An investible market portfolio is constructed from CRSP common stocks (NYSE, AMEX, NASDAQ) from January 1967 to December 2024.
- The universe is screened for liquidity and market capitalization to exclude extreme microcaps while preserving 99.7% of total market capitalization.
- A custom market return, cMKT, is constructed as a value-weighted buy-and-hold portfolio. It is verified to track standard Fama–French and q5 market factors with extremely high correlation ( $R^2 > 0.999$ ), ensuring that subsequent pricing errors are not artifacts of market factor implementation.
The Body–Tail Decomposition:
- At each formation date, the investible universe is split into two legs based on cumulative market capitalization rank:
  - Body: The top $p$ percent of cumulative market cap.
  - Tail: The remaining stocks.
- The split ratios $p$ range from 0.50 to 0.99.
- Crucially, the body and tail legs are tradable portfolios that, when recombined with their time-varying market weights, exactly reconstruct the original aggregate market portfolio.
- The test estimates time-series alphas for the body and tail separately and tests the joint null hypothesis that both alphas are zero ( $H_0: \alpha_{body} = \alpha_{tail} = 0$ ).
Matched Null Benchmark (Random Splits):
- To distinguish pricing errors linked to the size-ranked structure from mechanical effects of splitting a portfolio, the author employs random splits.
- These splits use the same universe, formation dates, and split ratios but assign stocks to legs randomly rather than by market cap rank.
- If rejections occur in body–tail splits but not in random splits, the errors are attributed to the size-ranked structure, not the act of splitting.
Models Evaluated:
- CAPM, Fama–French 3-factor (FF3), Carhart 4-factor, Fama–French 5-factor (FF5), Fama–French 6-factor (FF6), and the q5 model (market, size, investment, ROE, expected growth).

Key Results

Aggregate vs. Internal Consistency:
- All candidate models pass the aggregate market sanity check, exhibiting high $R^2$ and statistically insignificant alphas for the full market portfolio.
- However, the body–tail decomposition reveals stark differences. While CAPM, FF3, FF5, and FF6 are broadly stable (failing to reject the joint zero-alpha null at most split ratios), the q5 model rejects the joint zero-alpha restriction at all nine split ratios.
Systematic Offset in q5:
- Under the q5 model, the body leg consistently exhibits negative alphas, while the tail leg consistently exhibits positive alphas.
- These errors are of sufficient magnitude to cancel out when the legs are recombined, explaining why the aggregate market alpha remains insignificant.
Role of Size-Ranking:
- The rejection is specific to the size-ranked structure. In matched random splits, the q5 joint rejection rate drops to the nominal level (approx. 4.1%), indicating that the error is not a mechanical consequence of splitting the market but is tied to the size hierarchy.
Factor Block Diagnostics:
- The rejection is not driven by the market factor. Replacing the q5 market factor with the author's cMKT does not eliminate the rejection.
- Ablation studies identify the ROE–EG (Return on Equity – Expected Growth) block as the source of the tension.
  - The MKT-only q5 variant passes the test.
  - Adding ME (Size) and IA (Investment) maintains stability.
  - The inclusion of EG (Expected Growth) and ROE drives the rejection.
  - Dropping EG substantially reduces the rejection rate (from 9/9 to 2/9 split ratios), while dropping ROE leaves significant rejections.
Time Variation:
- The q5 body–tail rejection is not time-invariant. It is most pronounced in the 1980s through the mid-2000s and weakens significantly in recent years (2010s–2020s). This aligns with the weakening of the EG and ROE factor premia in recent data.

Significance and Claims
The paper makes three primary contributions:

Internal Consistency Check: It introduces the body–tail test as a diagnostic tool to evaluate whether a factor model prices the market's internal components, not just the aggregate. It demonstrates that aggregate market fit is a necessary but insufficient condition for internal pricing consistency.
Isolating Structural Errors: By using random splits as a matched benchmark, the paper isolates pricing errors specifically linked to the size-ranked structure of the market, ruling out mechanical split effects.
Diagnosing Model Specifics: The study identifies that the q5 model, despite its strong performance in mean–variance spanning tests (where it expands the investment opportunity set relative to FF6), leaves systematic, offsetting pricing errors on the market's size-ranked components. This suggests that a model can be "strong" in terms of Sharpe ratio expansion while failing to price the internal relative returns of the market portfolio.

Conclusion
The author concludes that the aggregate market can appear priced because internal pricing errors cancel out. The body–tail test reveals that the q5 model's profitability–growth block (specifically Expected Growth) creates a tension where large-cap stocks are underpriced (negative alpha) and small-cap stocks are overpriced (positive alpha) relative to the model's predictions. The paper does not claim q5 is generally inferior in a mean–variance sense but argues that factor model evaluation must extend beyond aggregate market fit to include the pricing consistency of the market's tradable internal components.