Wei and Chi's ranking of Chinese sell-side reports is convincing; the evidence for a trade stops well short. At 63 trading days, upgrades beat downgrades by 3.00 percentage-point in gross event returns. That long-short difference comes before costs. The paper makes no attempt at a trading case, and we found nothing in it that would support a position you could hold after costs.
What counts as news?
An analyst changing their view gives the market fresh information. Repeating a Buy repeats a position already known. The authors test this familiar, sound idea using 504,600 analyst-report events from the iFinD terminal, covering July 2009 to March 2026. Each report counts as one stock-date event. They distinguish the recommendation level from the recommendation action. Levels range from Score1, the most favourable, to Score5; Score3 is neutral. The action comes from the vendor's label: upgraded, maintained or downgraded. They measure forward returns over 1, 5, 21 and 63 trading days after the report becomes tradable, subtracting the equal-weighted market return for each window.
Among the 350,989 events with action labels, the upgrade-minus-downgrade spread reaches 0.58pp after one day (t=13.60). By 63 days it reaches 3.00pp (t=11.35). Regressions with report-year fixed effects control for log size and log prior-365-day report count. Those give a 63-day spread of 3.253pp (t=12.45). Adding stock fixed effects produces 2.803pp; broker fixed effects produce 3.170pp.
The authors also construct a forecast-fundamental revision. This measures whether a report's forward P/E and PEG look more or less attractive than the previous forecast information for that stock.
Actions hold up; levels wander
I believe the action result. The recommendation levels are harder to trust. Market-adjusted Score1 minus Score3 starts at 0.20pp after one day, then falls to -0.74pp at 21 days (t=-5.81). At 63 days it is -0.09pp (t=-0.51). Across a month, the most favourable rating trails neutral. With year fixed effects and controls, however, the 21-day level effect changes sign to +0.463pp (t=4.48). Clustering by report-month reduces that to t=1.59. The authors describe levels as "less stable", even as their regressions keep the result in play. The action spread retains t=8.10 at 63 days under the same month clustering.
The benchmark comparison makes the case for focusing on the long-short spread. At 21 days, Downgrade minus Maintain is -0.995pp raw and -0.34pp after adjustment. Upgrade minus Maintain rises from 0.802 to 1.46. Each leg shifts by about 0.66pp, leaving the spread almost unchanged: 1.796 raw and 1.802 adjusted. Returns measured against Maintain therefore depend on the benchmark. At 21 and 63 days, the spreads scarcely change (1.796 vs 1.802; 3.059 vs 3.001).
The small-stock claim is thinner
The authors find stronger effects in smaller stocks and names with less coverage. The coverage evidence holds up: at 63 days, the action spread's low-minus-high bucket difference is 2.050pp (t=2.97). The regression estimate is 2.00pp (t=2.93).
Size has weaker support. Small-stock events account for just 8.7% of the sample (43,978). The small-minus-large difference at 63 days is 1.88pp with t=1.84 in buckets, and 1.56pp with t=1.53 in regression. The case for size depends more on shorter windows, including 1.82pp at 5 days (t=4.68). Larger gross returns in small names tell us nothing about returns after their trading costs. We found no figures for liquidity, turnover or capacity in the paper.
Reading an unchanged rating
The third claim concerns information inside Maintain reports. Those with favourable forecast revisions outperform those with unfavourable revisions by 0.153pp after one day and 1.104pp after 63. Their t-statistics range from 11.776 to 15.576. The authors also form a confirmed action spread: upgrades accompanied by favourable revisions against downgrades accompanied by unfavourable revisions. At 63 days, that spread reaches 3.708pp (t=8.644), exceeding the 3.00pp baseline.
I did not find the control I wanted. The revision measure is scaled by price. In the authors' words, it "captures changes in the relationship between the current stock price and analyst-forecast earnings and growth". Built as described, it would record a favourable revision when the stock price falls and forecasts stay unchanged. Short-term reversal or value could therefore explain part of the signal.
The sizes of the effects fit that concern. The Maintain signal delivers 1.104pp at 63 days. Favourable revisions offset downgrades by 1.095pp (t=2.541), and predict 1.315pp (t=8.318) even for reports with no action label. Across different report types, the same price-scaled measure earns roughly the same return.
Upgrades are the exception. Favourable revisions add just 0.489pp to upgrades at 63 days (t=1.007). A revision driven entirely by price should help upgrades as well, so this estimate works against the mechanical explanation. It is also the table's noisiest estimate. The equal-weighted market is the only return benchmark, and the regressions contain no reversal or momentum control. Reconstructing revisions from forecast earnings changes alone, with price held fixed, would settle the question. Even that reproduction requires iFinD's report-level forecast tables.
A 3-point spread, built over 63 trading days
The cumulative 3.00pp starts at entry. Excluding day one leaves about 2.42pp of drift from day 2 to day 63. The stretch between days 21 and 63 contributes 1.20pp. With reports arriving daily, holding for 63-day periods would create substantial overlap among positions.
The downgrade side requires short exposure. The authors cite Jiang, Lu and Zhu when discussing how short-sale constraints affect the incorporation of news into Chinese prices. We found no discussion of daily price limits or T+1. Nor could we establish from the text whether the one-day return begins at the open or at the prior close of the first tradable day. That distinction determines how much of the 0.58pp could disappear into a limit-up open.
We could not backtest this.
Our price data excludes Chinese A-shares. We also lack the iFinD report history and report-level forecast tables needed for the Maintain and confirmation signals. Rerunning it in the US would test a different experiment.
The action spread is strong enough to use for ranking events. For the Maintain result, I would change my mind if an earnings-only revision retained a t-statistic near 12.