Measuring the Costs of the U.S.-China Trade War of 2022-2025

Overview

Our new Andersen Institute White Paper on the U.S.-China trade war documents that the most recent round of the conflict—building on the Wassenaar Arrangement framework for dual-use export controls—began on August 15, 2022 with the announcement and implementation of the initial U.S. export controls policy toward China. ^[1],[2] Chinese retaliation, in the form of rare‑earth export controls, was first announced on July 3, 2023. After more than two years of escalation, the United States and China reached a partial truce on October 29, 2025. To quantify the economic consequences for U.S. firms caught in this conflict, we apply the standard event study toolkit used in financial economics.

Event Study Methodology and the Event Window

In an event study, one measures how stock returns respond to an event by comparing actual returns with a counterfactual benchmark. Typically, the first step is to construct that counterfactual benchmark of normal returns—the returns each firm would have earned in the absence of the event—by estimating, over a pre‑event period, the firm‑specific loadings on the market factors that drive its stock’s reactions to market-wide variation during the event window (without the influence of the event). The daily abnormal return is then defined as the difference between the actual return on an event‑day and the counterfactual normal return. For each affected firm, summing these daily abnormal returns over the event window yields the cumulative abnormal return (CAR), which captures the total valuation effect of the event. CARs can also be computed for any subset of the event window.

The valuation effects of the export control war between the U.S. and China can be gauged by identifying a group of firms that were particularly exposed to the conflict and calculating their CARs over the event window.

For example, we know that the window from August 15, 2022 to October 29, 2025 is the relevant period to gauge the market effects of U.S. export controls on firms involved in exporting products subject to export controls–-specifically, Ultra-wide bandgap semiconductor substrates, EDA (Electronic Design Automation) software for advanced chip design, and pressure gain combustion technology, each identified as strategically sensitive dual-use technologies. The October 7, 2022 policy change constituted a broader shock, extending restrictions on exports to China to advanced AI chips and semiconductor manufacturing equipment. Nevertheless, August 15, 2022 remains the appropriate starting date for the event window because it marks the date at which the market became aware of the rule and affected firms began publicly commenting on and reacting to the proposed controls. Accordingly, we treat the October 2022 rule as a compounding shock within the same regulatory episode rather than a separate event.

Identifying Firms Affected by Export Controls

What is the best way to identify the list of affected firms? Our approach to identifying firms affected by U.S. export controls uses two criteria, and we report estimates for three alternative samples: firms that satisfy one criterion (the AI-identified Sample), firms that satisfy the other criterion (the Comment-Identified Sample), and firms that satisfy both criteria (the Overlap Sample). To be included in the CommentIdentified Sample, the firm must have commented on the Commerce Control List rule during the Export Administration’s regulatory process. Firms with substantial exposure had the strongest incentives to comment, making the comment file a natural screen for identifying firms with meaningful stakes. To be included in the AIIdentified Sample, the firm must discuss its vulnerability to these export controls in its Form 10-K (as identified by the AI, Claude). Because publicly traded firms are obliged to disclose material risk and important contributors to their business environment, 10-Ks are a potentially useful source of information to identify firms that were substantially affected by the rule.

Requiring firms to meet both criteria (the Overlap Sample) may be desirable because each screen is potentially imperfect. With respect to relying on a list of commentators, some submissions were made by industry groups of firms rather than individual firms, and not all members of a commenting group may have been equally exposed to the export-control costs. The 10‑K screen is also potentially imperfect because it relies on an AI system to detect whether a firm explicitly discusses the relevant export controls. To assess the reliability of the AI’s classifications, we examined a small number of firms that appeared on the “commentators” list but not on Claude’s list. For example, one of those firms, Cisco, discussed export controls, political risk, and China in its 10-Ks, but did not explicitly link those three elements in the same passage. That omission likely led Claude to classify Cisco as unaffected, although we believe it was exposed to the rule, as its presence on the commentators list indicates.

Another factor weighing in favor of using the Overlap Sample is that, in our empirical exercise, errors of omission—leaving out some affected firms—are less serious than errors of commission, which involve adding firms to the affected list even though they were not actually exposed to the export controls. For that reason, it seems preferable to adopt a conservative identification strategy when constructing the average CAR, so long as the resulting sample remains large enough to achieve reliable inferences. By requiring firms to satisfy both criteria, we obtain a sample of affected firms that is smaller but more reliably inclusive of genuinely affected firms. Nevertheless, we report results below for all three samples.

Claude’s reading of 10-Ks identified 52 (AI-Identified Sample) firms that explicitly discussed vulnerability to the export controls. The second (Comment-Identified Sample) approach produces a list of 76 firms that either submitted comments individually or participated as members of commenting groups. Of these, we were able to identify 70 public firms with available stock price data. The Overlap Sample contains 25 firms. We construct equally weighted average CARs for each of these three alternative samples of firms, using two different methods for measuring CARs, and plot the results in the top panels of Figures 1 through 3.

Statistical Challenges

In constructing the CARs and interpreting them we faced two major problems, both of which result from performing an event study on a very long event window: First, when using a sample period this long for an event window, there is the well-known problem of “alpha bias.” Alpha is the estimated intercept term in a regression that identifies factors that affect stock returns, in which estimated betas measure the covariance of a stock’s return with various factors. When constructing an estimate of normal returns for a short event window, one estimates alpha and betas from the pre-event window, and defines the normal daily return during the event window as alpha plus each of the estimated betas multiplied by its respective factor during the event window (the five Fama-French factors plus momentum). When event windows are long, because alphas can differ substantially between the pre-event period and the event window, CARs can be strongly affected by the use of the pre-event window as the estimate of alpha in the long event window. Variations in the value of alpha in the data likely capture omitted factors whose covariances matter for returns.

In the Appendix, we describe two alternative methods we adopted for estimating CARs in light of the alpha bias problem. One method attempts to estimate the event-window alpha for the affected firms by assuming it is the same as the average alpha for other firms during the event window. The other method uses a matched-controls approach that relies on Fama-French characteristics to identify matches, which thereby avoids estimating alphas and betas in the pre-event window to construct CARs for affected firms during the event window. According to the matched-control approach, CARs are defined as the difference between the raw returns of affected firms and the raw returns of each affected firm’s matched group of control firms. Reassuringly, we find that our results are quite similar across the two different methods for measuring CARs.

The second problem we faced was that a long event window makes it likely that subsets of the firms identified as belonging to the affected group may have abnormal returns that reflect shared idiosyncratic factors unrelated to other firms in the affected group. A natural concern with interpreting the average CARs as the treatment effect is that the average CARs may mask distinct subgroup patterns related to changing earnings prospects for different subgroups over the event window. We discovered from examining individual firms’ CARs that there were indeed such clusters of firms that systematically deviated from others in the affected group.

To address this second problem, we derive the first principal component of the CARs for each sample, which captures the common factor explaining the largest share of variation across all firms over the event window. This first principal component is our estimate of the effect of the U.S.-China export-controls trade war, under the assumption that the trade war is the most important common factor connecting all the firms in each sample during the event window.

Results

For each of the three samples, we report the first principal component in the lower panels of Figures 1 through 3 (where each Figure makes use of a different sample of affected firms). Panel (c) shows the first principal component using the factor approach methodology that assumes all firms have identical alphas in the event window, and panel (d) shows the first principal component when using the matched-control method. For the three samples, the first principal component captures between 22 and 39 percent of the variation in the respective individual CAR paths for each sample.^[3]

As Figure 1 shows, the first principal component of CARs for the 25 affected firms in the Overlap Sample is negative during the export control war, reaching a trough of roughly -26 percent on April 3, 2025, when CARs are estimated using the factor approach (where we estimate alphas from the Russell 2000) and -42 percent on that same date when CARs are estimated using the matched-control approach. By the end of the event window, the first principal components are roughly zero, according to both methods for estimating CARs. When we construct similar measures for the two alternative samples (AI-Identified, or Comment-Identified), we obtain similar results for the first principal component, both in terms of the timing of cumulative peaks and troughs and in terms of their rough magnitude, as shown in Figures 2 and 3.

April 3 is the first trading day after President Trump’s “Liberation Day” speech, in which he announced an aggressive new tariff policy aimed at rebalancing international trade. It is not surprising that the same policy environment that produced sharply higher tariffs also coincided with the most severe valuation losses for firms exposed to the export- control measures. It is also interesting to note that—again, according to all three Figures—this low point in the first principal components was largely anticipated: market prices had been drifting downward for weeks and months leading up to Liberation Day. 

All three Figures also show that the first principal component of affected firms’ CARs began recovering months before the October 2025 de‑escalation, indicating that the market gradually priced in the likelihood of a negotiated resolution from April to October 2025. That expectation was ultimately validated by the October 29, 2025 agreement, which paused both sides’ escalation paths.

In principle, one could repeat the above exercise to identify and measure the CARs of U.S. firms that were exposed to Chinese export controls on rare-earth products. In practice, however, this is a much more challenging task for at least three reasons.

First, firms affected by China’s rare-earth export controls were exposed through a restriction on an input, rather than an output. As a result, the list of potentially affected firms spans a much more diverse group of industries, with correspondingly wider likely variation in the cost of a reduction in rare-earth availability. Second, unlike the limits on U.S. firms’ exports to China analyzed above, the cost impact of a rare-earth input restriction may be small for some firms compared to other shocks they experience, even if the aggregate consequences of the limited access to rare-earth products are large. Third, because China’s export controls did not involve a U.S. regulatory process, there is no analogous “commentators” list of U.S. firms that responded to the rules, which leaves one entirely reliant on AI-based reading of firms’ 10-Ks to judge which firms to consider affected. We also considered employing an alternative AI-Identification method to validate Claude’s identification of affected firms. One recent academic study applies AI to detect mentions of Chinese export controls in publicly traded firms’ earnings calls (“Geoeconomic Pressure,” by Christopher Clayton, Antonio Coppola, Matteo Maggiori, and Jesse Shreger, NBER Working Paper No. 34020, July 2025). Surprisingly, the two groups of firms—the one identified by Claude’s 10-K readings and the one identified by the earnings-call data in Clayton et al.—had no firms in common. Given the three challenges listed here, we concluded that we could not reliably estimate the valuation effects of Chinese rare-earth restrictions, although we continue to grapple with the problem. In conclusion, the effects of U.S. export controls on high-tech exports to China appear to have been substantial for the firms directly exposed to them. At the trough on April 3, 2025, the estimated effect is between -17 percent and -42 percent, depending on which method one uses to identify affected firms and to construct CARs. The negative valuation effect reversed after April 3, 2025, and rose to roughly zero by the end of October 2025, as trade negotiations between the U.S. and China produced a de-escalation. Markets appear to have anticipated both the April 2025 trough and the October 2025 resolution months in advance of each of those dates.

---