The aggregate Shenwan pharma index can send a trader the wrong way on a policy event. In three of six events, sub-sectors moved in opposite directions: the 2015 clinical-trial inspection, the 2020 stent procurement and the 2021 DRG/DIP plan. Li and Zhao make that point themselves. They also identify the 2017 review and approval reform as a case where the index works reasonably well: all seven indices had medium-frequency CARs of the same sign. Even when every medium-frequency CAR was negative, though, their tables show a wide spread beneath the index.

Six events, several return bands

Policy announcements can change expected cash flows by different amounts for drug makers, distributors and device firms. Li and Zhao examine six national pharmaceutical policy events in China:

The event dates are official release dates. For both procurement rounds, they use the announcement of provisional selection results.

The aggregate is the Shenwan first-level Pharmaceutical and Biotechnology Index, which covers the whole pharma and biotech sector. Li and Zhao compare it with the CSI 300 and run the same analysis for six second-level sub-sectors: chemical pharmaceuticals, traditional Chinese medicine (TCM), biological products, pharmaceutical distribution, medical devices and healthcare services. Their data are daily iFinD closes from June 4, 2014 to October 31, 2025.

EEMD supplies the split. It sifts many noise-perturbed copies of each return series and averages them, producing oscillating components (IMFs) with increasing periods. The authors group IMF1 to IMF3 as the "high-frequency" return and IMF4 to IMF6 as the "medium-frequency" return, leaving the slowest components aside. At each scale, they fit a market model over [−260, −11]. They then sum CARs over four symmetric windows, ranging from [−1,+1] to [−10,+10].

Their broader finding is that policy orientation and instrument type do not consistently predict abnormal-return patterns. The authors describe the exercise as descriptive and say they ran no formal tests across events or scales.

Procurement splits the sector

Both procurement rounds involved state-organized volume-based buying: firms bid for guaranteed volume at cut prices. After "4+7", the paper finds negative medium-frequency CARs for the first-level index and all six sub-sectors in all four windows. The losses still varied sharply. Over [−10,+10], chemical pharma was at −11.29%, medical devices at −9.73%, and TCM at an insignificant −0.36%.

The stent results differ. Medical devices fell −7.15% over [−10,+10] at medium frequency, the sector's largest drop. Pharmaceutical distribution was positive in every medium-frequency window and reached 3.31% over [−5,+5]. The same procurement instrument produced the opposite sign for one sub-sector. A desk watching only the aggregate could miss both the depth of the device move and the distributors' gain. This direct comparison is the paper's strongest evidence.

The qualification matters. These are the same medium-frequency CARs used elsewhere in the analysis, drawn from a two-sided filter and windows containing up to ten pre-event days. The paper notes that several sub-sectors already had negative medium-frequency ARs before the "4+7" date. Medical Devices gives the stent result some support across scales: its CARs were the lowest of the seven indices in all four high-frequency windows. The sub-sector split remains the sturdiest result, though the leakage concern below reaches it as well.

What remains unspecified?

The paper gives Nstd = 0.20, 100 ensemble members, a cap of 1000 sifting iterations, both windows, and a rule dropping a date from all indices if any index lacks it. Data are on Figshare and code on GitHub. The text leaves three matters open.

First are the IMF boundaries. The authors selected them using average periods and variance shares. Their sensitivity table shows the cost of moving a single IMF across a boundary: 88.1% to 91.1% of the 168 medium-frequency CAR signs survive, meaning roughly 15 to 20 flip. Changing Nstd or ensemble size has less effect, with 96.4% to 98.2% retained.

The seed is described as fixed, but we did not find its value in the paper text.

The significance stars are harder to assess. We did not find a description of the test behind them. Smoothing is built into the medium-frequency series, yet many CARs have three stars even in the three-day [−1,+1] window. If the stars come from a standard CAR t-test that assumes independent daily ARs, significance is overstated on an autocorrelated input.

When the bands disagree

The sign reversals appear in the authors' own account. Their abstract says high- and medium-frequency CARs "differ in direction for some events and sub-sectors". In Event 3's [−5,+5] window, every high-frequency CAR is positive, including chemical pharma at +3.35%. Every medium-frequency CAR is negative, with chemical pharma at −10.47%. Event 1 is more striking: over [−10,+10], pharmaceutical distribution has a high-frequency CAR of −15.15% and a medium-frequency CAR of +26.90%.

The authors argue that the scales "should not be directly equated with investors' short- and medium-term assessments". Their point is that direction at one scale "does not necessarily extend to other statistical scales". A trader, however, holds the raw return, the sum of the layers. Adding the two reported columns gives about −7.1% for Event 3 chemical pharma and about +11.8% for Event 1 distribution. These are rough proxies: each scale has its own alpha and beta, and the paper leaves out the low-frequency layer. A decomposition can spread one move across bands, leaving large offsetting components around a moderate net. We did not find CARs on undecomposed returns in the paper. That baseline belongs beside every row.

Did the 2025 weakness precede the news?

The 2025 innovative-drug policy is explicitly supportive. Yet its medium-frequency CAR over [−10,+10] is −4.00% for the first-level index and −4.22% for biotech, both three-star. The authors attribute the result to weakness before the event: medium-frequency ARs for most indices were already negative. From the second trading day after release, medium-frequency ARs turn positive for all seven indices. They cannot tell whether the earlier weakness reflects expectations, earlier information or other news.

The filter offers another possible explanation. The authors say EEMD "was applied ex post to the full-sample return series, so its frequency components should primarily be interpreted as statistical scales". Their methods section restricts those components to "ex post characterization" and rules out real-time trading signals and causal identification. Even for an ex post description, timing matters. Sifting forms envelopes using extrema on both sides of a date. A medium-frequency value on day −5 thus depends partly on returns after day 0, allowing the filter to carry a post-event move backwards. That bears directly on the pre-event drift the paper infers for Event 6.

The authors address multiple comparisons by stressing patterns "corroborated by daily AR trajectories". Those trajectories come from the same two-sided decomposition, so they cannot settle the filter concern. The paper also reports generally positive high-frequency AR on the event date and CARs over the shorter windows. That layer, too, was fitted on the full sample.

No Shenwan replication

We did not replicate the Shenwan event study. Our price coverage is US-focused; it lacks the Shenwan sub-sector indices and CSI 300 history. US healthcare assets would substitute a different market's reaction to Chinese policy.

Across the two procurement rounds, sub-sector dispersion is the sturdiest finding. Before trading on the band-level sign flips or the Event 6 pre-event drift, a reader would need a one-sided decomposition and a raw-return CAR alongside them. The authors disclaim real-time use. This caution is for their readers.