A beta-1 stock earns about 1.25% a month when the market factor's flow quantity is high, about 0.25% when it is low, and an unconditional 0.75%. That conditional spread carries the paper. It revives a claim about expected returns that most of us have treated as dead ground since Black.
An, Su and Wang argue that the price of factor risk changes with the amount of that risk held by the marginal investor. Retail money leaves funds. Funds sell in proportion to prior holdings. The investor taking the other side acquires more of the embedded systematic exposure and demands a higher price for carrying it next month.
The authors call their implementation beta times quantity (BTQ). It replaces the constant factor premium in the arbitrage pricing condition with a linear function of a factor-level quantity variable q. Expected return is the sum, across factors, of beta times lambda times q. The cross-sectional no-arbitrage restriction remains intact while the premium varies. Betas come from single-factor realized cov/var, rather than the multi-factor first stage in the original Fama-MacBeth procedure, a deviation the authors flag themselves.
Building q
The mutual fund dollar flows are derived from TNA changes net of returns, using 1,707,742 fund-month observations after cleaning. Holdings weights lagged two full quarters push those flows into stocks. This is more conservative than the one-quarter convention in Lou. The resulting stock-level dollar flows are aggregated by factor, with each stock weighted by its realized covariance with the factor return from 12-month rolling daily windows. The factor-level flow is then scaled by total US market cap and averaged over six months.
So q measures a beta-weighted aggregation of trading induced by mutual fund flows. The authors' aggregate flow series has a 0.63 correlation with ICI equity fund flows and 0.47 with the Flow of Funds line. They describe their measure as broadly consistent with both, attributing the differences to mutual fund coverage. We take those two correlations as evidence of a narrow channel.
The sample runs from January 2000 to December 2022: 276 months, about 1,644,000 stock-months and roughly 6,000 stocks a month. Training covers 2000-2009. Testing covers 2010-2022 and includes 849,519 stock-months. One split, one parameter set, no rolling re-estimation.
The reported fit
In the single-factor specifications, BTQ produces out-of-sample panel R-squared, with no demeaning under the Gu-Kelly-Xiu convention, of 0.75% for MKT, 0.84% for HML, 0.65% for MOM and 0.60% for SMB. Matching beta-only regressions return 0.05%, 0.15%, 0.02% and -0.10%. The FF3C multi-factor BTQ model reaches 1.07% OOS, compared with 0.22% for beta-only.
Across the 153 Jensen-Kelly-Pedersen factors, 139 of 153 single-factor BTQ specifications have positive OOS R2. The median is 0.38%. For the market factor, full-sample lambda is 1.80% with a t of 4.18. The beta-only mu is 0.38% with a t of 1.07, while 90 of 153 JKP factors have negative mu point estimates.
Lasso considers all 159 candidates and retains five at the penalty where OOS R2 peaks: market, betting-against-beta, 21-day return volatility, 21-day idiosyncratic volatility from the HXZ q-factor model, and book-to-market on enterprise value. All five lambdas are positive. Choosing factor names from the OOS peak introduces look-ahead, as the authors acknowledge before re-tuning through ten-fold in-sample cross-validation. At the cross-validated penalty, the model delivers 0.81%.
The authors also substitute the first 50 principal components of factor returns, estimated over 1970-2009. Only PC1 and PC2 survive selection. Both are positive, and the model records 0.77%.
The sign restriction is the result I find hardest to wave away. Reversing a factor's sign reverses beta and q together, leaving lambda greater than zero invariant to sign convention. Mu greater than zero lacks that property. Every selected factor comes out positive, including both PCs. The prediction could have failed. It passed.
Placebos mostly fail
Sending stock-level flow straight into expected returns and bypassing factor aggregation gives in-sample R2 of at most 0.006%. Allowing a separate coefficient for every stock raises in-sample fit to 0.4%, then produces OOS results between -81% and -232%.
A second exercise matches each factor's beta with every other factor's q. Across 25,122 combinations, roughly half the R2 mass falls below zero. Beta multiplied by the best of 126 FRED-MD macro series reaches 0.28%. The chosen series is unemployment insurance initial claims, selected ex post, as the authors say, and the median is -0.12%. Beta times factor momentum peaks at 0.12%.
Large caps carry the result
For the Lasso-selected model, OOS R2 rises from 0.46% in the smallest NYSE quintile, covering 323,617 observations, to 2.16% in the largest, with 103,927. The PC version moves from 0.44% to 2.09%. Characteristic and machine-learning predictors usually show the reverse size pattern because limits to arbitrage bind hardest in microcaps.
The middle of the test window is much weaker. Sub-period OOS R2 for the selection model is 1.16% in 2010-2014, 0.15% in 2015-2018 and 1.00% in 2019-2022. The authors link the middle trough to quieter variation in q, visible in their time-series plot. A harsher reading is that the model needs violent flows. Within the test window, spring 2020 is the only such episode, and 2019-2022 OOS R2 is 1.00%.
The lookback evidence is loose at both ends. Across h from 1 to 12, selected-factor OOS R2 ranges from 0.33% to 0.85%, with its peak at h=5 instead of the h=6 benchmark. The PC specification ranges from -0.15% at h=3 to 0.88% at h=7, while h=1 gives exactly 0. Six months was chosen, in the authors' words, for simplicity and transparency.
An appendix contains a harder result for the proposed mechanism. The time-series regression of each factor return on its own q mostly fails out of sample. HML records -14.70% and SMB -1.05%; only MKT is positive at +6.74%. In-sample R2 is 5.05% for MKT and 5.59% for HML.
The authors describe this as a much weaker argument for the pricing power of quantity and peripheral to their main focus. Their defence is reasonable. The cross-sectional lambda need not follow from the time-series coefficient because beta dispersion exists in the cross-section and disappears from the time series. We still regard this as the most direct test of the mechanism. It leaves the pricing result dependent on cross-sectional spread rather than demonstrated factor timing.
Still no portfolio
No portfolio return appears anywhere.
The predictive evidence rests on panel R2 near 1%, alongside the conditional SML and the Fama-MacBeth premium plots. In those plots, the market factor premium moves from below -2% a month to nearly +3% across q bins. R2 is defensible because the machine-learning literature uses it, and the beta-only comparison is apples to apples. Yet the gap between this evidence and a tradable book remains unmeasured. Averaging q over six months and estimating betas through 12-month rolling windows also suggests a slow signal, which the paper does not cost.
We could not test it. Constructing q requires fund-level subscriptions and redemptions together with stock-level holdings history, and we have neither. Volume or dollar-flow proxies would examine a different mechanism. Exact reproduction would remain difficult even with the right vendors because cleaning includes manual decimal-point corrections and switches to Morningstar TNA whenever CRSP deviates by more than 50%.
A rolling-window re-estimation that kept the same five factors in 2015 and in 2020 would move me. Their identities emerged from a 10-year training sample screened against 159 candidates. Betting-against-beta, two volatility measures and a value variant form a suspiciously familiar low-risk cluster. The market factor looks firmer than the rest, with a t of 1.96 surviving inside FF3C.
The authors claim novelty from bringing quantities into the factor pricing framework. q enters the factor premium, then predicts next month's cross-section. The model survives a placebo battery that defeats the more casual form of the same idea. Its underlying exposure is plain equity beta, which trades fine, unlike the premium in the correlation-rotation note. Access fails only at the conditioning variable.