The asset growth premium that sits inside the Fama-French and q-factor models is better understood as the residue of two mispricing curves bending in opposite directions. In a new study, Zhuo Cheng, Daniel Cohen and Jing Fang find that the familiar downward slope from low-growth to high-growth firms shows up only when two very different populations are pooled. Separate them and neither group has a straight negative line. That matters for anyone holding a conservative-investment tilt. If the premium comes from prices correcting in small, opaque, costly-to-trade stocks, then its persistence depends on those frictions persisting. The theory both benchmark models rest on says nothing about that.
What the Balance Sheet Line Actually Measures
The standard account treats asset growth as a clean proxy for corporate investment. Fama and French tie it to expected book equity growth through the dividend discount model: higher expected growth, holding price constant, implies a lower expected return. Hou, Xue and Zhang tie it to q-theory: firms invest less when their cost of capital is high, so low investment marks a high discount rate. Both stories assume that total asset growth is mainly a capital-spending decision responding to the cost of capital.
That assumption had already been weakened before this paper. In their 2024 study, Michael Cooper, Huseyin Gulen and Mihai Ion rebuilt the investment factor from conventional and intangibles-inclusive measures of investment and found that both benchmark models lost much of their explanatory power. Most of the pricing information in the asset growth factor came from growth in inventory and accounts receivable.
That finding changes what the factor appears to measure. Inventory and receivables are working-capital accruals. They are the balance sheet lines where management estimates, revenue recognition timing and channel stuffing show up first. A factor whose power sits mainly in receivables and inventory looks less like a sort on capital budgeting and more like a sort on how hard a company's reported growth is to verify. The q-theory label describes plant and equipment. The return information sits elsewhere in the filing.
Cheng, Cohen and Fang pick up from there. Their question is direct: if the investment channel is not driving the result, what is?
Hard-to-Value Firms at Both Tails
Their answer begins with who sits at the extremes. Across asset growth quintiles, firms that are shrinking sharply and firms that are expanding quickly share a profile. Compared with the middle of the distribution, they are smaller, younger, more R&D-intensive and more operationally levered. Their earnings and cash flows are more volatile. Their accounting is less predictable, and their accrual estimation errors are larger.
Measures of information quality point the same way. Firms in the top and bottom quintiles have wider analyst forecast dispersion, larger forecast errors and higher return volatility. They are also harder to arbitrage, with lower institutional ownership, thinner analyst coverage and wider bid-ask spreads.
The result is a U-shaped relation between asset growth and the severity of mispricing. The authors confirm it directly with price-to-value estimates built from two established valuation methods. Valuation error is smallest in the middle quintiles and largest at both ends.
From there the prediction follows. If realized returns contain a material mispricing-correction component, and mispricing is worst at both tails, then the asset growth-return relation should depend on which way a firm is mispriced. Among undervalued firms, returns should rise at both extremes as prices correct upward, producing a U. Among overvalued firms, returns should fall at both extremes as prices correct downward, producing an inverted U.
Two Curves That Average Into One Line
The obvious obstacle is that no one can observe in advance which firms are undervalued and which are overvalued. The authors handle this with finite mixture normal regression, an unsupervised method that sorts observations into latent groups and estimates a separate return relation within each group. The researcher does not assign firms to groups. On more than 172,000 firm-year observations from 1963 to 2022, the model is left to find its own structure.
The data strongly favor two groups over one. Log-likelihood, the Akaike Information Criterion and the Bayesian Information Criterion all improve sharply when a second group is added. The BIC falls by more than 10,000, far beyond the conventional threshold for decisive evidence in favor of the more complex model.
The two groups look as the mispricing account predicts. In one, next-year returns are U-shaped across asset growth quintiles, highest for the strongest shrinkers and the fastest growers and lowest in the middle. In the other, the shape inverts: returns are lowest at the extremes and least negative in the middle.
How these curves combine into the famous anomaly comes down to asymmetry. In both groups, the downward-sloping segment is steeper than the upward-sloping one.
- In the U-shaped group, the decline from extreme shrinkers toward the middle is steeper than the rise toward extreme growers. Undervalued shrinkers recover more than undervalued growers.
- In the inverted-U group, the decline from the middle toward extreme growers is steeper than the rise from extreme shrinkers. Overvalued growers fall the hardest.
Pool the two groups and run the single regression used throughout most of the literature, and you get the textbook result: returns decline steadily from the lowest asset growth quintile to the highest. The nonlinearity and the two-regime structure are hidden by averaging. CMA's monotone slope is a summary statistic of a more complex structure, not the structure itself.
Testing the Mispricing Read Against Risk
The authors run three tests aimed at separating mispricing from risk.
First, they remove each firm's exposure to 10 established factors, including CMA, momentum and four factors built specifically to capture mispricing and behavioral effects. If the patterns were compensation for priced risk, this adjustment should largely eliminate them. It does not. The U and inverted-U shapes and their asymmetry largely survive.
Second, they check the latent groups against an external measure. Firms with higher price-to-value ratios, which are more likely to be overvalued, are significantly more likely to be assigned to the inverted-U group. This test carries the most weight. A purely statistical clustering method produced groups that line up with an independent valuation signal in the direction the mispricing account requires.
Third, they use idiosyncratic volatility as a proxy for the cost of arbitrage. The nonlinear patterns are stronger among high-volatility stocks, where correcting a mispricing is expensive and risky. Mispricing that persists because it is costly to trade against is the standard signature of limits to arbitrage. The same pattern often appears in the value and momentum literatures.
The results hold across decile sorts, alternative asset growth definitions (including several from the Cooper, Gulen and Ion critique), trimming of extreme returns, exclusion of firms without a full 12 months of returns, and three separate 20-year subperiods. The effect does not weaken over time. That is notable, because many published anomalies have decayed as markets have become more efficient. An effect that holds up after decades of attention is consistent with frictions that are expensive to overcome.
The chart above provides context for the universe involved, not a test of the paper. Small caps are where the study locates the mispricing, and their returns differ sharply across trend and volatility conditions. The high-volatility state is where idiosyncratic risk makes correction most expensive.
Where the Case Is Weakest
The strongest objection is methodological. Mixture models fitted to fat-tailed return data can find multiple groups that partly reflect non-normality rather than distinct economic populations. A BIC improvement of 10,000 on 172,000 observations is strong evidence that one normal component fits worse than two. It is weaker evidence that the two components are specifically "undervalued" and "overvalued" firms. The price-to-value validation addresses this, but price-to-value estimates are themselves model outputs with their own errors, and those errors are probably largest in the young, R&D-heavy firms at the tails.
Second, the risk adjustment cannot settle the question alone. Showing that 10 factors fail to explain a pattern shows that those factors do not capture it. It does not show that no priced risk could. A q-theory defender can argue that the right risk factor has not yet been built.
Third, mispricing and risk are not mutually exclusive. Some of CMA's premium may still be compensation for risk, with the mispricing component dominating at the extremes. The paper argues that the mispricing component dominates for asset growth. It does not show that the risk component is zero.
Finally, this is a single new study. The Cooper, Gulen and Ion result makes its conclusion more plausible, but independent replication on international samples would test whether the two-regime structure holds where accounting regimes and arbitrage capacity differ.
What a Conservative-Investment Tilt Is Actually Holding
The practical consequence concerns where the premium comes from and how durable it is. A risk premium should persist as long as the risk does. A mispricing premium persists only as long as correction remains costly. On this evidence, the asset growth effect falls mostly in the second category, and its returns concentrate in small, thinly covered, high-volatility, low-institutional-ownership stocks. Those are the stocks where trading costs and implementation frictions consume the most of a strategy's paper return.
Two implications follow. First, a diversified large-cap fund with a conservative-investment tilt probably captures a smaller share of the premium than backtests on the full cross-section suggest, because the correction happens mainly in names such funds hold lightly or not at all. Second, asset growth by itself is a poor signal at the tails. An extreme shrinker can be a strong recovery candidate or a falling knife, depending on whether it sits in the U-shaped group or the inverted-U group. The paper suggests that a valuation estimate is what separates them. An extreme value of asset growth is better read as a warning that the company is hard to value than as a buy or sell signal.
The test that would confirm or overturn this view is clear. If combining extreme asset growth with independent price-to-value estimates produces out-of-sample return spreads that exceed CMA's, and those spreads shrink as liquidity and institutional ownership rise, then the investment factor is mainly a limits-to-arbitrage factor and should be modeled as one. If the conservative-minus-aggressive spread holds up with similar strength among liquid, well-covered large caps where arbitrage is cheap, the risk interpretation regains ground. The paper's evidence so far points to the first outcome.
