{"id":"c9676bea-df7f-45db-8505-d850efdb53c1","arxiv_id":"2411.16553","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Activists' proxy communications are more aligned with the preferences of larger mutual fund shareholders, and this alignment predicts more attention, more votes, and higher success in proxy fights.","lead":"This paper finds that hedge fund activists tailor their proxy-fight communications to the preferences of mutual funds holding large stakes in the target company. Campaigns that align better with those preferences attract more shareholder attention, receive more votes, and are more likely to succeed, especially when mutual funds hold a large combined stake.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Align measure's transfer from shareholder-proposal voting to proxy-contest language is weakly validated; the main holdings-alignment result could be an artifact of SVR prediction error.","rationale":"The reader's weakest assumption is also the most load-bearing concern: the entire central claim rests on Align measuring true institution preferences in proxy contests. The SVR is necessarily trained on shareholder proposals because proxy fights are rare, but the training and prediction samples differ in genre, stakes, and decision context. The paper's own limitation discussion in Internet Appendix C.4 acknowledges this discrepancy and mitigates only by noting similar aggregate against-management rates, which does not ensure phrase-level preference weights transfer. Table 6 is the right validation, but it is small and selected, covering only fights that reached a vote, while the main Table 4 sample includes all fights. The manual method in Section 6.3 is reassuring but inherits the same transfer assumption. My proposed holdout test directly checks whether the measure generalizes across time and sample domains; if it fails, the paper's central result would not support the stated causal interpretation. I do not change the reader's CONDITIONAL verdict because the paper has multiple corroborating results, including actual-vote validation and a simpler non-machine-learning measure, and the concern is addressable with additional validation rather than fatal.","tokens_in":38156,"tokens_out":8745,"duration_ms":96131,"concrete_test":"Holdout validation: train the SVR on shareholder proposals from only July 2003 through June 2014, freeze the phrase coefficients, and compute Align for proxy fights with meeting dates in 2015 through 2019. Then estimate Table 6 (actual SupAct on Align) and Table 4 (Align on holdings) in this holdout period. If either coefficient attenuates by more than half or loses significance, the transfer from proposal voting to proxy-contest language is not stable, and the main result is not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3 trains SVR on shareholder proposal texts and votes, then applies the phrase coefficients to proxy communications. This is a domain shift: precatory, one-paragraph shareholder proposals differ from adversarial, multi-issue proxy solicitation materials, and institutions may weight issues differently in contested director elections. The paper's validation (Table 6) uses only 1,457 institution-fight records from 199 fights that reached a vote, less than 40% of the 522 fights in Table 4, and those fights are selected by the decision to proceed to a ballot. The validation confirms that predicted Align correlates with actual votes in that subsample, but it does not establish that Align is unbiased in the full sample, nor that the Table 4 holdings coefficient reflects true preference alignment rather than SVR prediction error correlated with holdings. The manual proposal-type check (Section 6.3) also relies on the same transfer assumption. If Align is a noisy or systematically biased proxy, the central claim that activists tailor communications to larger holders' preferences is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies whether activist hedge funds tailor their proxy-fight communications to the preferences of the institutional investors that hold large stakes in the target company. The author constructs a text-based measure, Align, by training a support vector regression on mutual funds' votes on shareholder proposals and applying the phrase coefficients to proxy communications. Using a sample of 522 proxy fights from 2004 to 2019, the paper reports that alignment increases with an institution's stake in the target, and that better-aligned campaigns attract more attention from institutions, receive more actual votes, and are more likely to succeed when mutual fund ownership is high. The paper also presents robustness checks using lagged holdings, a small set of merger-based shocks, a Russell index reconstitution case study, and a non-machine-learning proposal-type measure.","tokens_in":38385,"tokens_out":5968,"duration_ms":58025,"significance":"If the central claim holds, the paper makes a useful contribution to the activism literature by moving beyond target and activist characteristics and quantifying how campaign language is tailored to the preferences of large shareholders. The paper has several genuine strengths: the Align measure is validated against actual institution votes in a subsample (Table 6); the manual proposal-type measure in Section 6.3 provides an independent non-machine-learning check; and the main alignment-holdings correlation is robust to proxy-fight and institution fixed effects and to six-month lagged holdings. The paper is also transparent about limitations, including the training/prediction domain shift and the lack of counterfactual data. However, the causal interpretation of the main result rests on very limited exogenous variation, and the invested-only subsample does not robustly reproduce the headline coefficient with both fixed effects.","major_comments":[{"comment":"The Align measure is trained on shareholder proposal texts and applied to proxy communications, a domain shift that the paper explicitly acknowledges in Appendix C.4. The validation against actual votes in Table 6 is limited to 1,457 institution-fight records from 199 fights that reached a ballot, a selected subsample of the 522 fights in Table 4. If SVR prediction error is correlated with holdings—for example, because larger institutions have more training observations or more systematic voting patterns—the coefficient in Table 4 could partly reflect measurement bias rather than true tailoring. The paper should report whether SVR prediction errors are orthogonal to holdings in the full sample, or provide a validation covering a broader set of fights, before the central claim can be regarded as established.","section":"§3.1–3.2, Table 6, Appendix C.4"},{"comment":"The causal interpretation in the abstract and in Section 6—that activists 'design' campaigns around the preferences of large shareholders—relies on very weak exogenous variation: 11 fund-merger proxy fights in Table 9 and one Russell reconstitution case study involving three institutions in Figure 5. The merger specification yields an Acquired×Post coefficient that is only marginally significant in the baseline (0.157, t=1.79), and the Russell evidence is a single case with no inferential power. These designs cannot rule out time trends or activist selection of merger periods. The paper should either substantially soften the causal language to 'suggest' or 'are consistent with,' or provide a stronger identification strategy.","section":"§6.2"},{"comment":"The main holdings-alignment result is not robust in the subsample of institutions that actually hold target shares when both proxy-fight and institution fixed effects are included: Table 14, column (4), reports a coefficient of 0.0065 with t=1.64, which is not significant at conventional levels. The paper's statement that the appendix 'shows that the results hold' for the invested subsample is therefore overstated. Because 81% of the main sample has zero holdings (Table 3), the Table 4 result may be driven largely by the contrast between invested and non-invested institutions rather than by the size of the stake. This sensitivity should be discussed and the invested-only specification should be displayed as a main robustness table.","section":"§4 vs. Internet Appendix C.2, Table 14"},{"comment":"The reported 9.4 percentage-point effect for above-average-ownership fights is computed incorrectly. For OwnDum=1, the marginal effect of a one-standard-deviation increase in AgAlign is γ+β = −0.0218 + 0.111 = 0.0892, not −0.0218 + 0.0048 + 0.111 = 0.094; the ownership-dummy main effect is not part of the marginal effect of AgAlign. The text and conclusion should report approximately 8.9 percentage points (or the correct interaction effect) and should ideally provide a confidence interval. This is a small but real arithmetic error in a headline number.","section":"§5.3, Table 7"}],"minor_comments":[{"comment":"The observation count is inconsistent: the text reports 66,836 observations and 12,582 with non-zero holdings, while Table 3 and Table 4 report 66,432 observations; please reconcile these numbers.","section":"§3.3"},{"comment":"Panel (a) of Figure 5 labels the target as 'Leap Water,' while the text and panel (b) refer to 'Leap Wireless'; this typo should be corrected.","section":"Figure 5"},{"comment":"The citation 'Li, Patel, and Ramani Li et al.' is malformed and appears to merge authors and a subsequent citation; this entry should be fixed.","section":"References"},{"comment":"The paper reports 533 proxy fights in Section 2.2 but the main analysis sample in Section 3.3 contains 522 fights; the source of the difference (for example, institutions failing the 100-proposal voting threshold) should be stated explicitly.","section":"§2.2, §3.3"},{"comment":"The validation sample of 1,457 institution-fight records from 199 fights is described as covering less than 40% of the 522 fights; the paper should state this explicitly when discussing the generalizability of the validation.","section":"§5.2, Table 6"}],"recommendation":"major_revision","confidential_remarks":"The paper's empirical design is careful in many respects, but the gap between the abstract's causal 'demonstrates that hedge funds tend to design' and the actual identification is substantial. The invested-only subsample result and the domain-shift validation issue are the two points that most need attention before publication. The manuscript is likely within scope for a finance journal; the weak exogenous variation is not fatal if the claims are reframed as descriptive and the limitations are prominently acknowledged."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a real empirical contribution, not a buzzword paper. It builds a phrase-level measure of institution preferences from full shareholder proposal texts via SVR, then shows that proxy communications align more with institutions holding larger stakes in the target. That correlation is robust to proxy-fight and institution fixed effects, six-month lagged holdings, subsamples, and a manual proposal-type measure. I believe the descriptive result. The paper is also careful to use outcome variables—attention, votes, success—that are measured separately from the voting data used to construct Align, so the circularity concern that often plagues text-based measures does not really land here.\n\nThe genuinely new piece is the measure itself: nobody else has pulled full proposal text and used supervised ML to recover institution-level phrase weights, then applied those weights to activist communications. The validation against proxy voting guidelines in the appendix is a nice touch and gives me more confidence that the SVR coefficients capture something real. The attention result using EDGAR logs is clever and the magnitudes are plausible.\n\nNow the soft spots, in proportion. The transfer assumption is the biggest one: SVR is trained on precatory, often one-paragraph shareholder proposals and applied to adversarial, multi-issue proxy solicitation materials. The paper acknowledges this and offers Table 6 as validation, but that validation covers only 199 fights that actually went to a vote—less than half the main sample. That is enough to show the measure has predictive content in a selected subsample, but not enough to rule out systematic bias in the full sample. It is a reason to treat coefficient magnitudes cautiously, not a reason to dismiss the paper.\n\nThe causal language is stronger than the identification. The phrase \"activists design campaigns\" is repeated throughout, but the exogenous variation is eleven fund mergers and one Russell reconstitution case study. The lagged-holdings result helps rule out the most obvious reverse-causality story, and the author is candid about not being able to separate selection from tailoring. Still, I would want the conclusion to say \"associated with\" rather than \"design.\"\n\nOne technical issue the paper does not address: Align is a generated regressor, and the standard errors in the attention and vote regressions do not appear to account for first-stage estimation error. That is fixable with bootstrap or a correction, and probably does not change the qualitative findings, but a referee should ask for it.\n\nWho is this for? Corporate governance and activism researchers, definitely. It deserves a serious referee. I would send it out.","headline":"A genuinely new text-based measure and a robust holdings-alignment correlation, with a causal claim that outruns the identification.","tokens_in":38851,"tokens_out":1516,"would_cite":true,"duration_ms":20495,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that activists design their proxy campaigns to align with the preferences of institutions holding larger stakes in the target company, and that such alignment raises attention, votes, and success.","keywords":["hedge fund activism","shareholder preferences","proxy contests","text analysis","machine learning","mutual fund voting","institutional investors","Align score"],"falsifier":"A direct test would be to estimate Align from pre-fight shareholder-proposal voting for the universe of proxy contests with disclosed votes, then compare it with the votes actually cast; if Align has no predictive power for institutions whose proxy-contest voting diverges from their shareholder-proposal voting, or if the holdings-alignment gradient reverses sign in that subsample, the central claim would be refuted.","tokens_in":37963,"feed_emoji":"🗳️","tokens_out":7286,"duration_ms":66937,"temperature":0.7,"pith_summary":"This paper claims that activist hedge funds design their proxy campaigns around the preferences of the institutions that own the largest stakes in the target firm. It estimates each institution's preferences from how it voted on shareholder proposals in the prior two years, then measures how closely the activist's written communications match those preferences using a text-based score called Align. The central finding is that a one-standard-deviation increase in an institution's stake is associated with a 0.4 to 0.7 percentage point rise in alignment, and that campaigns aligning better with large holders draw more attention, more votes, and a higher chance of success. If true, this means persuasion in proxy fights is aimed at a small set of powerful mutual funds, and activist language is a strategic tool rather than just rhetoric.","feed_headline":"Activists write proxy fights to please their largest fund holders","feed_subtitle":"Text-aligned campaigns draw more attention, gain 3pp more votes, and succeed 9.4pp more often.","key_machinery":"The central object is Align, a score between 0 and 1 that measures how close a proxy communication is to an institution's revealed preferences. It is built in two steps: a Support Vector Regression is trained on each institution's votes against management on shareholder proposals in the two years before the fight, with up to five-word phrases as features, so every phrase receives a coefficient; then the frequencies of those phrases in the activist's proxy filings are combined with those coefficients, $\\mathrm{Align}_{p,i} = \\alpha_i + \\beta_i \\cdot \\mathbf{x}_p$, to predict the chance that the institution supports the activist. A holdings-weighted aggregate version, AgAlign, weights each institution's Align by its stake in the target, and this aggregate is what links alignment to campaign success.","core_discovery":"The paper's central claim is that activists tailor their campaign communications to the preferences of institutions holding larger stakes in the target. The author constructs Align, a 0-to-1 measure of how close a proxy communication is to an institution's revealed preferences, by training a support vector regression on each institution's votes against management on shareholder proposals with matched proposal text, then applying the resulting phrase coefficients to the activist's communications. The paper finds that a one-standard-deviation increase in target ownership (about 0.63 percentage points, versus a mean of 0.09%) is associated with a 0.4 to 0.7 percentage point increase in Align; an institution owning ten percent of the target sees average alignment of 65%, compared with 46% for institutions holding under 0.01%. Better-aligned campaigns also receive more attention (a 23% increase in EDGAR views per standard deviation), more actual votes (a three-percentage-point increase in institution support per standard deviation), and are 9.4 percentage points more likely to succeed when mutual funds hold above-average stakes. The author argues that the direction runs from activist tailoring to support, using six-month-lagged holdings, merger shocks to holdings, and a Russell reconstitution case study.","pith_inferences":["A testable extension the author does not run: applying the same Align construction to plentiful say-on-pay or ESG shareholder votes could reveal whether the holdings-alignment gradient holds outside the small proxy-fight sample.","If the mechanism is genuine, activists' issue selection should track the proxy guidelines of their largest mutual fund holders, so future campaign themes could in principle be predicted from those public guidelines.","The concentration of the success effect in above-average-ownership fights implies that in diffusely held targets, catering to big funds is cheaper but less decisive, so activists may rely on other persuasion channels such as media or direct board negotiations, which this paper does not measure."],"forward_implications":["Activist proxy communications are a measurable strategic input: any campaign's text can be scored against institution preferences using public SEC filings.","Alignment predicts actual votes, not just attention: a one-standard-deviation rise in Align is associated with a three-percentage-point increase in institutions' activist support.","The success effect is conditional on mutual fund ownership: alignment raises win probability by 9.4 percentage points only when mutual funds hold above-average stakes, with no detectable effect otherwise.","Institutions pay attention selectively: for similar holdings, funds access proxy filings more often when the text aligns with their revealed preferences."],"supporting_citations":[{"why":"Supplies the Support Vector Regression method that converts phrase frequencies into institution-specific preference coefficients.","marker":"Drucker et al. (1997)"},{"why":"Provides the benchmark of voted proxy fights (207) and prior evidence that activists target firms with supportive investors.","marker":"Brav et al. (2023)"},{"why":"Supports the premise that activists initiate fights where institutional investors are receptive, which the paper extends to issue-level tailoring.","marker":"Kedia et al. (2021)"},{"why":"Shows ownership structure shapes activist strategy at the institution-type level, which the paper pushes down to individual institution preferences.","marker":"Appel et al. (2019)"},{"why":"Supplies the EDGAR log-file method used to measure institutions' attention to proxy communications.","marker":"Iliev et al. (2021)"},{"why":"Provides the Bayesian persuasion foundation for why communication content should move voting behavior.","marker":"Kamenica and Gentzkow (2011)"},{"why":"Documents the costs of proxy fights, motivating why activists would tailor communications to win support efficiently.","marker":"Gantchev (2012)"},{"why":"Supplies the mutual fund merger list used as an exogenous shock to holdings in the robustness tests.","marker":"Lewellen and Lowry (2021)"}],"fun_headline_variants":["Activists court big funds with tailored proxy pitches","Proxy fights aligned with big investors win more votes","Funds' voting history shapes activist campaigns","Tailored proxy campaigns boost votes and success","Activists mirror large holders' views to win support"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's load-bearing premise is that preferences revealed in institutions' votes on routine shareholder proposals transfer to contested proxy fights, so that the Align score built from proposal texts measures true alignment in campaigns; the validation sample of actual proxy-fight votes is small (1,457 institution-fight records from 199 fights).","fun_headline_variants_meta":{"raw":{"variants":["Activists court big funds with tailored proxy pitches","Proxy fights aligned with big investors win more votes","Funds' voting history shapes activist campaigns","Tailored proxy campaigns boost votes and success","Activists mirror large holders' views to win support"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000476,"raw_usage":{"total_tokens":2312,"prompt_tokens":851,"completion_tokens":1461,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":467,"completion_tokens_details":{"reasoning_tokens":1391}},"tokens_in":467,"tokens_out":1461,"duration_ms":10216,"temperature":1.0,"reasoning_tokens":1391,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:58:53.160301+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would be to estimate Align from pre-fight shareholder-proposal voting for the universe of proxy contests with disclosed votes, then compare it with the votes actually cast; if Align has no predictive power for institutions whose proxy-contest voting diverges from their shareholder-proposal voting, or if the holdings-alignment gradient reverses sign in that subsample, the central claim would be refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports the premise that activists initiate fights where institutional investors are receptive, which the paper extends to issue-level tailoring."}],"review_version":1}