Most of Gan's 13% two-day abnormal return arrived during the session before gazettal. The result therefore reads as leakage followed by an announcement. Gan acknowledges as much: "within the 2-day cumulative 13% increase, the day before the event (−1, −1) contributed the dominant portion, suggesting a potential anticipatory price adjustment." The paper's central issue is contained in that sentence.

A disclosure comes first. We could not trade the paper's universe, which consists of 30 mainland A-shares selected from an Eastmoney concept board. Substituting a US-listed crypto-infrastructure basket would lose the abnormal returns examined in the paper. We ran no substitute test.

The trade Gan tests

Hong Kong gazetted its Stablecoin Ordinance on 30 May 2025, making it the first dedicated stablecoin statute in a major financial center. The Ordinance leaves mainland firms' current cash flows and existing operating environments unchanged at once. Its effect is to replace an uncertain future with a defined compliance route for firms connected to stablecoin infrastructure.

The trade follows directly. Companies positioned to serve that route through payment rails, custody, security certification or hardware receive a larger expected-opportunity set. Their shares should rise when the text becomes public.

Gan tests the idea with A-share concept stocks. He reviews firm business descriptions, broker research and Eastmoney's thematic concept board, producing 32 names. After excluding ST stocks and companies with fewer than 100 trading days in the estimation window, 30 remain. OLS estimates normal returns over [-250, -11] using four benchmarks: CAPM, Fama-French three-factor, Carhart four-factor, Fama-French five-factor.

The factors are built in-house rather than purchased. Their construction uses roughly 6,300 A-share listed companies after excluding ST and financials. The resulting panel contains about 1.5 million firm-days across 242 trading days from July 2024 to June 2025. Gan applies the standard 2x3 sorts and rebalances at end-June. CSI All-Share serves as the market proxy, while the Chinese 10-year is converted to a daily risk-free rate.

CAPM produces CAAR of 0.130 over (-1, 0), with t = 8.230. The corresponding results are 0.141 for FF3, 0.139 for Carhart and 0.138 for FF5. CAAR peaks at (-1, +3): 0.188 under CAPM and 0.189 under FF3, reported in the conclusion as approximately 18.9%.

Before the event, (-5, -2) is flat and insignificant under every model. CAPM gives 0.005, t = 0.658; Carhart gives -0.010, t = -1.484. A 1,000-iteration placebo test samples 30 random non-concept A-shares, and the real CAAR falls wholly beyond the right tail of the simulated distribution. Gan also performs a 1:1 nearest-neighbour match using average market cap and average BM within the same industry letter code. Concept stocks remain significantly higher, although the paper reports no test statistic for this exercise.

A flat week is followed by no response to the May 21 draft passage, then nearly the entire move occurs during the lone session before gazettal.

Four models, one unresolved cause

Cross-model consistency supplies Gan's main robustness argument. In Figure 1, the four CAAR curves nearly coincide. He takes that overlap as evidence that factor pricing bias does not generate the abnormal returns. Across (-1,0), the estimates range from 0.130 to 0.141, leaving about a percentage point between the benchmarks. Yet their agreement cannot distinguish an Ordinance effect from early access to Gazette proofs or a thematic bid accumulating through a week of crypto headlines.

The paper already contains the more discriminating test. Along with conventional and BMP t-stats, Gan reports Kolari-Pynnonen versions that adjust for cross-correlation among stock residuals. For thirty stocks sharing one calendar date, much of the residual variance is common.

The contrast is sharp. The (-1, 0) window survives, with CAPM KP-t = 2.212, p = 0.035 and FF5 KP-t = 3.588, p = 0.001. Wider windows fail to do so. CAPM records KP p = 0.069 for (-1, +1), 0.064 for (-1, +3) and 0.118 for (-1, +5). Over that same (-1, +5) window, the naive t reaches 7.301 and the BMP-t reaches 6.872.

The 18.9% window has KP p = 0.064, missing 5% while remaining within 10%. The tougher result is (-1, +5), where CAAR is 0.174 and KP p = 0.118. Gan deserves credit for reporting all three columns, a practice many event studies omit. He calls the weaker KP results "unsurprising given the conservative nature of the cross-correction adjustment," before citing 18.9% in the conclusion. That conclusion also says CAARs are significant "at the 1% level across all reported windows". Table 1 lists KP p-values of 0.064, 0.069 and 0.118 for the wider windows. With 30 stocks shocked on the same date, accurate also describes the adjustment.

The 29 May constituent problem

We cannot reconstruct a 30-stock A-share concept basket as of 29 May. No point-in-time record is available to us for the Eastmoney stablecoin concept board's membership on 29 May 2025.

Gan confronts the issue directly. Eastmoney, he writes, identifies stocks from FY2024 annual report segment descriptions and company websites, "with classifications established before the event date." The supplementary appendix contains the complete 32-name list. Still, a third party maintains the board and the supplementary list carries no date, leaving readers unable to verify membership independently as of 29 May. We found no evidence that the list changed, while also lacking a way to check. This does not support an accusation.

The 13% also should not be read as a P&L figure. Daily price limits bind for Chinese small caps, and a limit-up print does not guarantee a fill. Transaction costs, daily price limits and capacity constraints are absent from the paper. Its figure captures the market's revaluation of these companies. A tradable book could have earned something different.

Hardware leads, small caps follow

The cross-sectional analysis offers the most economically credible section. Gan divides 30 names into five industry-chain segments. Hardware equipment records the highest CAAR and exceeds security certification at t = 3.978, p = 0.040. Against payment, hardware gives t = 2.309 (p = 0.073); against fintech, t = 2.266 (p = 0.092). FinTech against Payment shows virtually nothing, t = 0.283.

A p of 0.040 paired with a t of 3.978 reveals how few degrees of freedom remain.

The board comparison places ChiNext above the Shanghai main board at t = 2.358, p = 0.032. ChiNext versus Shenzhen is insignificant (t = 1.127), as is Shanghai versus Shenzhen (t = -1.472). Only one of three pairwise comparisons clears. The size split produces the same p, 0.032, with t = -2.970 under CAPM. Non-leading firms exceed leading firms during certain sub-windows. Gan says the difference holds at 5% under all four models, although Table 3 displays only the CAPM figure.

Both findings are interpreted through a retail-attention channel, with citations to Barber and Odean and the Chinese attention literature. Gan argues that smaller companies with concentrated stablecoin exposure allow investors to connect regulatory benefits to earnings more directly than they can for diversified conglomerates.

The explanation is plausible. A retail-driven theme bid passing through thinly traded small caps after a headline could produce the same pattern without regulatory-benefit content. The splits are univariate, and we did not find a joint cross-sectional regression combining size, board and chain position. Hardware companies in this sample may simply be the smallest and least liquid. Under that reading, the three heterogeneity results amount to one result counted three times.

Canada remains unexplained

The international analysis applies a single-index CAPM across 13 markets, each with at least three stablecoin-related listings. South Korea supplies the clearest positive response: 0.06 (t = 2.91) over [-1,+1], 0.09 (t = 4.60) over [-1,+5] and 0.13 (t = 2.85) over [-5,+5]. The US first turns negative, posting -0.04 (t = -5.37) over [-1,0], then reverses to +0.08 (t = 6.13) over [+0,+3].

Gan characterizes Germany, Italy and France as muted, with no window attaining conventional significance. Germany's estimate over [-5,+5] is 0.06 (t = 1.33), while France records 0.04 (t = 1.14). Calling the response absent becomes harder with a 6% point estimate on the table.

Gan states the section's purpose plainly: "We offer this section as a descriptive benchmark, a first-pass documentation of how stablecoin-related stocks in different jurisdictions moved around the same calendar date." He presents the US time-zone explanation as conjecture and lists alternatives he cannot exclude. For Europe, MiCA appears as "one plausible interpretation" of the muted response. That degree of caution fits three-stock country baskets lacking controls for local crypto or macro news. Indonesia uses 28 May as its event date because 29-30 May were Ascension Day closures.

Canada is the unaddressed observation. Its returns are significantly negative in both the pre-event and wider windows: -0.12 over [-5,-2] (t = -3.39), -0.05 over [-1,+1] (t = -3.25) and -0.17 over [-5,+5] (t = -2.53). The [-1,+3] window is flat (t = -0.79). In magnitude, that [-5,+5] result exceeds every positive country response. The paper records it without comment. Any account of Hong Kong's licensing framework as a global positive for stablecoin infrastructure has to explain Canada.

What would change the reading?

For me, the paper's most useful result is the KP column. Gan does not present it as a contribution. He identifies three others: the first event study of stablecoin legislation, cross-validation across four pricing models and gradient transmission along the compliance chain. Yet the KP results show what event-day clustering does to a t of 7.301 for (-1, +5). Once adjusted, it drops to 1.612. Researchers quoting t-stats from single-date event studies of concept baskets should examine that comparison first.

Gan's defense of the day -1 move is carefully phrased. It "does not undermine the event's value as an identification anchor, but it does imply that the estimated CAARs should be interpreted as capturing the cumulative effect of both anticipation and the official announcement". For the 2-day window, the anchor holds: this is the sole window that KP clears, with CAPM p = 0.035 and FF5 p = 0.001. At (-1, +3), p = 0.064; at (-1, +5), p = 0.118. The anchor no longer holds there.

Two pieces of evidence would change my view of the causal claim. The first is a timestamped Eastmoney constituent snapshot from 29 May 2025 that matches the supplementary list. The second is intraday evidence placing the day -1 move after a specific, identifiable disclosure instead of showing it dispersed across the session. Gan proposes high-frequency data for future work, and that is the right direction.

Without those additions, his own limitation remains the defensible reading: CAAR combines anticipation and announcement. Anticipation accounts for most of the 13%. The paper is therefore a well-executed study of information leakage in a market with a high retail share, published as a study of regulatory clarity.

We have previously identified small-sample Chinese equity evidence as the weak leg in otherwise careful work, including our review of tensor-BEKK covariance estimation (our note). Here, everything rests on 30 stocks during one day.