Six of Zhou's seven Chinese studies measure only same-period crowding

A crowded-selling reversal trade can fail if the crowd you measure differs from the one you expect to unwind. Zhou's review makes that distinction worth taking seriously.

Across 65 references, this narrative review runs no estimation of its own. The trade behind it is familiar: funds buying or selling the same stock in the same window can move its price. What happens next decides the money. Trading on information should leave a persistent move; forced trading, such as selling to meet redemptions, should produce some reversal. Zhou groups the literature around three herding statistics and four possible mechanisms: information cascades, career concerns, relative performance with fund flows, and shared benchmarks or styles. The evidence table covers China.

The statistics come from Lakonishok, Shleifer and Vishny (1992), Wermers (1999) and Sias (2004). LSV measures how far the share of active funds buying a stock departs from the market-wide buyer share in a given period, such as a quarter. It takes the absolute difference, then subtracts the deviation expected from independent draws at that period's overall buying rate. Wermers separates the same count by trade direction. The buy-herding score (BHM) applies when buyers exceed the benchmark; the sell-herding score (SHM) applies when they fall short. Sias measures persistence by regressing the current period's standardised buyer share on the previous period's share. He separates the coefficient into funds repeating their own trades and funds following others. All three require holdings or transaction records with institution identifiers. Chinese holdings disclosures arrive quarterly or semi-annually.

Which crowd would the trade follow?

The three statistics answer different questions. Disagreement between them is therefore quite possible, and Zhou's example is one a desk would recognise. Funds might sell heavily during a stress quarter, then stop. HM and SHM could jump even as the Sias coefficient remains low in the recovery quarter. Gradual trimming over several quarters could instead produce a small HM alongside high Sias persistence.

The first pattern records a shock. The second records an unwind still in progress. Fading the first crowd and following the second amount to different bets.

That distinction runs through the Chinese evidence. Among the 7 micro-level studies in Zhou's summary table, 6 use LSV or an LSV-type measure. Just 1, Fan et al. (2024), uses a dynamic measure of fund-level following. Cross-checks are scarce too: only 2 of the 7, Caglayan et al. (2021) and Xue et al. (2023), compare LSV with adjusted LSV and the structural FHW measure. These are alternative measures used to check the herding finding. Zhou reports similar evidence across them. We found no magnitudes attached to that comparison.

Crashes do not reliably increase herding

The cited results resist a simple stress story. Foreign-investor herding in Korea weakened during the 1997 crisis relative to the pre-crisis period. Hwang and Salmon (2004), working at market level, also found their herding index falling during crises. Zhou states the sample limitation directly: "Neither of the two types of samples can directly represent China's public mutual funds, but these findings show that there is no fixed positive relationship between market pressure and herding."

The sentence's second half is defensible. A clean counterexample suffices to refute a fixed positive relationship. It cannot establish the sign for Chinese funds. Zhou acknowledges that limit too: "China's direct holdings-based evidence on extreme markets is still limited, and few studies jointly compare rising and falling markets with buying and selling within the same model."

This leads to the review's strongest design recommendation. Market direction and trade direction belong on separate axes. Stress studies should therefore report 4 cells: buying in rising markets, selling in rising markets, buying in falling markets, selling in falling markets. Calling a downturn sell-side herding by definition folds the price state into fund behaviour. The selling cells have an additional mechanical constraint. A fund generally needs holdings at the start of the period before it can sell, so SHM during a slump partly reflects the tradable set.

Redemptions compound the problem.

Outflows can force funds to sell overlapping liquid holdings together. The flow studies cited here, Coval and Stafford 2007 and Lou 2012, find temporary price impact followed by reversal. Common redemption selling can occur without imitation. Zhou also allows for funds responding to sales they expect other funds to make. Her warning that "common transactions cannot directly prove intentional imitation" applies directly to this case.

The buy-side crash-risk finding

Two cited results run against the usual prior. Deng et al. (2018) locate the association between fund herding and subsequent crash risk mainly on the buy side. Feng and Xiang (2023) find greater fund herding in stocks eligible for China's short-selling pilot, with no subsequent price reversal. Stock Connect (Zhao et al., 2021) and corporate site visits (Quan et al., 2023) are associated with less herding instead.

Zhou offers a plausible reconciliation. Better information may improve each fund's private judgment; it may also give many funds the same correlated signal. Feng and Xiang connect their finding to correlated signals. The review, however, reports no study testing those channels against each other.

No effect sizes carried forward from 65 references

Zhou carries forward no effect size from any cited study. We found neither a t-stat nor a herding level from any of the 65 references in the text. The Chinese summary table is qualitative throughout, and we found no account of how its studies were chosen. A desk weighing the turnover on a crowded-selling reversal has nothing here to size against.

We could not test the idea ourselves. Each of the three measures requires a fund-by-security panel of holdings or trades with stable institution identifiers. Separating flows from information also requires subscriptions and redemptions at fund level. We have neither dataset. The Chinese fund universe is also outside our coverage.

What a desk can use

Zhou's practical contribution is in study design. She recommends defining stress states independently of the herding statistic. Thresholds should follow out-of-sample rules or rolling windows, since choosing them over the full sample can introduce look-ahead. She also recommends applying LSV, Wermers and Sias to the same sample and reporting all 4 state-by-direction cells.

For a desk already running a crowding screen on 13F-style data, those checks are inexpensive additions. A Chinese study reporting magnitudes across all 4 cells, with flows controlled, would turn these design rules into something a trader could act on.