XSci (AI review)

September 26, 2026

Referee Report

Manuscript: "Cross-industry Technological Information Spillovers through Analyst Networks" Field: Empirical finance / financial economics (sell-side analyst research, corporate innovation, asset pricing) Recommendation: Major revision

Summary and Recommendation

This paper asks whether sell-side equity analysts acquire technological knowledge from colleagues who cover firms in other industries, and whether that knowledge reaches the market. The author constructs an analyst-firm-quarter measure, TechConnect, defined as the patent-value-weighted cosine similarity between a focal firm's distribution across 669 Cooperative Patent Classification classes and the corresponding distributions of firms in different Fama-French 48 industries covered by the focal analyst's colleagues over the preceding four quarters. Across 869,651 analyst-firm-quarter observations spanning 1995 to 2021, drawn from 10,589 analysts, 714 brokerages and 5,056 firms, higher TechConnect is associated with smaller absolute forecast errors. The relation strengthens at longer horizons and for firms whose earnings are harder to predict. Identification rests on 36 brokerage mergers between 1997 and 2020, where analyst-firm pairs experiencing a net increase in technologically relevant colleague connections exhibit a 5.082 basis point larger decline in forecast error, with the effect driven by arriving rather than departing colleagues. Two downstream analyses follow: firms that lose high-TechConnect analysts through brokerage mergers and closures increase low-impact patenting, and technology momentum attenuates among firms covered by analysts with higher average connectedness.

The paper is ambitious, and the underlying idea is a good one. Technological relatedness genuinely does cross industry boundaries — the author's own Appendix A1 shows that 27 of 48 Fama-French industries display higher cross-industry than within-industry technological similarity — and the proposition that brokerage research departments aggregate latent technological knowledge is original and testable. The data assembly is substantial, the attention to competing within-brokerage channels is conscientious, and several methodological choices, including the use of fixed-effects Poisson estimation for patent counts following Cohn et al. (2022) and the merger × analyst × firm fixed-effects structure, reflect real care. Nonetheless, three sets of problems stand between the manuscript and publication. The central construct may not measure what the paper claims it measures; both quasi-experimental designs assign treatment in ways that are not clearly exogenous and are supported by comparisons that are never formally tested; and the interpretation of magnitudes, together with a number of internal inconsistencies in the exposition, makes it difficult for a reader to verify the reported findings. A major revision is required, though the tests needed to address most of these concerns appear feasible with the data already assembled.

Major Concern Category 1: Measurement and Construct Validity of Technological Connectedness

1.1 Patent-value weighting and the shape of the measure

The paper's entire argument depends on TechConnect capturing the intensity of colleague-mediated technological information flow, and two features of Equation (3) make that interpretation difficult to sustain on current evidence. The first is the weighting scheme. Each colleague-covered firm enters weighted by the Kogan et al. (2017) dollar value of its patent portfolio, a variable that Appendix A2 reports with a mean of 22,286, a median of 778, a 99th percentile of 340,986 and a skewness of 12.40. With weights that concentrated, the weighted average in Equation (3) approximates, for many observations, the technological similarity between the focal firm and a single dominant patentee that recurs across brokerages and across time. TechConnect would then be closer to a firm characteristic — the focal firm's distance in technology space from a handful of very large innovators — filtered through whether some colleague happened to cover one of them. Footnote 17 reports that equal weighting does not change the results, but no equal-weighted estimates are tabulated, and given how much rides on this choice the untabulated claim cannot substitute for evidence. The second feature is the distribution of TechConnect itself, which is severely zero-inflated: a mean of 0.088, a median of 0.027, and a 25th percentile of exactly zero. At least a quarter of the estimation sample has no measured technological connection at all. The headline magnitude is nonetheless computed from a one-standard-deviation move of 0.135, which exceeds the 75th percentile of 0.120 and is roughly five times the median. That is not a marginal comparative static; it moves an analyst-firm pair from essentially no overlap to the upper decile of the distribution, and it conflates an extensive margin with the intensive margin the theory describes. Reporting equal-weighted and log-value-weighted variants, a decomposition showing how much of each observation's TechConnect comes from its top one and top five contributors, and effects estimated across TechConnect quintiles would settle much of this.

1.2 The cross-industry framing rests on an industry boundary the paper's own robustness calls into question

The paper's distinguishing claim is that it isolates cross-industry transmission, mechanically separating itself from the within-industry channel of Phua et al. (2023) by excluding same-Fama-French-48 firms. Three pieces of the author's own evidence suggest this separation is less clean than the framing implies. Appendix A4 reports that the TechConnect coefficient is significant at the 5% level in ten of twelve alternative-classification specifications, with the exceptions arising under the Fama-French 17 classification and two-digit NAICS codes. Section 6.2 attributes the nulls to coarse definitions excluding too many firms, which is plausible, but so is the opposite reading: coarser boundaries absorb precisely the economically adjacent pairs that drive the effect, and once those are reclassified as same-industry the result disappears. These readings have opposite implications for the contribution, and the paper does not distinguish them. Appendix A1 points the same way, since the strongest cross-industry matches are frequently economically proximate — Banking with Insurance at 0.35, Candy & Soda with Beer & Liquor at 0.32, Computers with Business Services at 0.23, Retail with Banking at 0.26. The appendix also raises a construct question that goes unaddressed: financial-sector firms display among the highest within- and cross-industry similarities in the table, yet their patents are predominantly business-method and software filings whose class overlap reflects the architecture of the classification system rather than shared scientific capability. Since the mechanism requires analysts to exchange substantive judgements about technological trajectories, showing that the result survives the exclusion of financial-sector firms, and among pairs at genuine economic distance, would considerably strengthen the claim.

1.3 Nothing in the paper establishes that information changes hands

Section 3 candidly acknowledges, citing Manski (1993), that transmission among colleagues is unobservable. The deeper problem is not that the channel is unobserved but that no test in the paper can separate colleague-mediated transmission from alternatives requiring no communication at all. The most natural falsification is absent. TechConnect is computed from the coverage portfolios of the focal analyst's colleagues; the identical quantity can be computed from the portfolios of matched non-colleagues at other brokerages, producing a placebo that captures technological relatedness in the coverage landscape without any within-firm channel. This is the direct empirical counterpart of the paper's own Figure 2, which contrasts an analyst at brokerage X who has a technologically similar colleague-covered firm against an analyst at brokerage Y who does not. A second class of evidence is also missing. If analysts absorb technological information from colleagues, the absorption should leave traces beyond the forecast error — in the timing of revisions relative to technologically linked firms' earnings announcements, in the co-movement of a focal analyst's revisions with those of connected colleagues, or in the boldness of forecasts. The author has already constructed BoldPositive, BoldNegative, Revision and SignedRevision, which appear in Appendix A2 and Appendix A6 but in no reported table. Reporting even one such test would convert a correlation between two slowly moving firm-level characteristics into evidence of an informational response. It is also worth noting that in Table 2, adding analyst fixed effects in Column (4) attenuates the coefficient from −2.586 to −1.359, a 47% reduction that is reported without comment and that is itself informative about the scale of analyst-level selection.

Major Concern Category 2: Identification in the Two Quasi-Experimental Designs

2.1 Merger treatment is assigned by post-merger retention, not by the merger

The identification argument in Section 3.2 is that mergers are driven by business-level strategy rather than by any particular analyst-firm pair, following Derrien and Kecskés (2013). That justifies treating the occurrence of the merger as exogenous, but the treatment variable is not the merger. It is an indicator built from exactly which analysts remain at the acquiring brokerage afterward and which firms they continue to cover. Post-merger retention is among the most consequential decisions a merged research department makes: redundant coverage is consolidated and the coverage universe is reoptimised around the combined franchise's strengths. If the merged entity retains analysts covering technologically coherent sets of firms in its strongest sectors, then treatment identifies pairs located in the parts of the department the firm chose to invest in — precisely where resources, attention and effort would be expected to rise for reasons unrelated to technological information. The merger × analyst × firm fixed effects absorb the level of any such advantage but not its post-merger change, which is what the design estimates. Two construction details compound this. Arriving and departing analysts are classified by appearance in I/B/E/S, so an analyst who remains employed but ceases issuing forecasts is misclassified, and an analyst dropped for redundancy is by construction one whose portfolio overlapped an incumbent's, which correlates with technological similarity. And the benchmark in Equation (6) is the incumbent's own last observed pre-merger TechConnect, a quantity that is itself endogenous and heavily zero-inflated, so an incumbent previously at zero is classified as treated by almost any arriving colleague.

2.2 No first stage, no dose-response, and controls that the treatment mechanically changes

The paper never shows that TechConnect changed. No table reports the realised change in the continuous variable around mergers for treated and untreated pairs, which leaves the reduced-form estimate unscaled and unverifiable against the baseline. This matters because the two designs disagree by roughly fifteenfold: 0.349 basis points for a one-standard-deviation move in the panel against 5.082 basis points in the merger sample, an effect the author benchmarks as equivalent to issuing a forecast 43.8 days closer to the announcement — nearly half the 90-day window the sample is truncated at. A dose-response specification using the continuous change in TechConnect, together with the distribution of that change by treatment status, would place both designs on a common scale and permit a test of monotonicity that the binary treatment cannot deliver. A related specification issue concerns the control vector, which Table 3 imports wholesale from Table 2 and which includes BrokerageSize, NumFirms, NumIndustries, Broker_IndCon and all four colleague-connectedness measures. Every one of these is mechanically altered by a merger, so conditioning on them is conditioning on post-treatment variables and can bias the estimate in either direction. Given how demanding the fixed-effects structure already is, a specification with no time-varying controls, or with controls held at their pre-merger values, should be well identified and would be informative.

2.3 The firm-innovation design omits the covariate that matters most, and its sharpest claims are untested

The outcome in Section 5.1 is patent counts, yet the propensity-score matching in Section 5.1.1 balances on market capitalisation, book-to-market, past returns, analyst coverage, turnover and return volatility, and not on pre-event patenting or R&D. Panel A shows that treated firms are much larger and much more heavily covered than the unmatched universe, and Appendix A2 shows that the patent distribution is extraordinarily skewed, with a mean of 15.73, a median of zero, a 75th percentile of zero and a skewness of 28. Matching on size and coverage does not place treated and control firms at comparable points in that distribution, and a Poisson specification with firm fixed effects will be dominated by the small subset of substantial patentees. Equally important, the paper's central comparison is between Treatment^High and Treatment^Low, but matching is performed against the shared control group rather than between the two treated subgroups, and Treatment^High — defined as losing at least one above-median-TechConnect analyst — is mechanically increasing in coverage breadth. Table 1 establishes that TechConnect rises with firm size, R&D expense and analyst coverage, so the two treated groups are unlikely to be comparable. Beyond matching, the conclusions rest on comparisons that are never formally tested. The claim that the effect is concentrated in low-impact patents rests on 0.131 (t = 3.18) against 0.070 (t = 0.90) in a separate column, where the high-impact estimate is imprecise rather than zero. The placebo in Column (6) yields 0.152 with t = 1.54, a point estimate larger in magnitude than the main estimate of 0.122; its insignificance reflects a wider standard error, not an absence of effect. A joint specification permitting cross-equation tests would resolve both.

2.4 Patent truncation, inference, and an unreconciled tension between the two designs

Appendix A6 defines the outcome as patents "filed and eventually granted" in a rolling twelve-month window, and footnote 15 notes that the application-to-grant lag averages two to three years. With events running through 2020, event year +3 extending to 42 months, and the patent data ending in 2023, post-event filings for late cohorts will not yet have been granted and are therefore missing from the count. Truncation of this kind has a specific signature — measured filings decline toward the end of the data — and the dynamic estimates show exactly that pattern, falling from 0.161 in year +1 to 0.120 in year +2 to an insignificant 0.087 in year +3. Section 5.1.2 reads this decay substantively, as the effect dissipating while information environments adjust; mechanical truncation is at least as plausible, and cohort fixed effects will not absorb its differential incidence across treatment groups. Restricting to cohorts with a complete grant window, or applying a standard truncation adjustment, would distinguish the two. Inference is a further concern across both designs. In the merger analysis, treatment is assigned at the event level with only 36 events, yet standard errors are clustered by analyst and firm; event-level clustering or wild-cluster-bootstrap p-values would be more appropriate. The analyst-level specification also combines merger × analyst × firm fixed effects with a single set of calendar-time effects shared across staggered mergers, which permits comparisons between the post-period of an early merger and the pre-period of a later one — the contamination that Baker, Larcker and Wang (2022), cited for the firm-level design, warn against. Finally, the two designs disagree on a question the theory makes central. Table 3 reports a departure coefficient of 1.840 (t = 0.80), insignificant in every event quarter, and explains it by suggesting analysts retain informal access after a merger. Section 5.1 is built entirely on the premise that losing a high-TechConnect analyst has a prompt and economically meaningful effect. Both propositions cannot comfortably be held at once, and the tension deserves direct treatment rather than a passing remark.

Major Concern Category 3: Interpretation, Magnitudes, and Internal Consistency

3.1 Benchmark magnitudes are drawn from specifications other than the one cited

Section 3.1 states that "using the coefficients from Column (2)," a one-standard-deviation increase in TechConnect reduces forecast error by 0.349 basis points, an effect "about 83% as large as the difference in forecast accuracy between star and non-star analysts, equivalent to issuing a forecast 3.5 days closer to the earnings announcement or having an additional 6.8 years of firm-specific experience." Only the horizon benchmark is recoverable from Column (2), where ForecastHorizon is 0.099 and 0.349/0.099 ≈ 3.5. The Star coefficient in Column (2) is −0.238 and statistically insignificant (t = −1.42), implying roughly 147% rather than 83%; the figure of 83% follows from Column (3), where Star is −0.419. Similarly, FirmExperience in Column (2) is −0.061, implying 5.7 rather than 6.8 units; the figure of 6.8 follows from Column (4), where the coefficient is −0.051. The benchmarks should be recomputed from a single specification, and benchmarking against a coefficient that is indistinguishable from zero in the stated column should be reconsidered. The same issue recurs in the horizon analysis. Section 4.1 reports a 36.93 basis point effect for the 1080–1260 day bracket from an interaction coefficient of −273.558, but this is a differential relative to the omitted 0–180 day bracket, and the main effect of TechConnect in that column is +71.421, implying a net effect of approximately −27.3 basis points. Reporting the interaction alone as a total overstates the magnitude by roughly a third, and the positive main effect — implying slightly worse accuracy at the shortest horizons — warrants explicit discussion.

3.2 An unexplained return predictor sits inside the market-efficiency test

Table 6 reports a positive and significant coefficient on firm-level TechConnect in every column, ranging from 0.889 to 2.931, which Section 5.2 addresses in a single clause. Using the standard deviation of 0.063 that the text itself employs elsewhere, the Column (2) coefficient of 2.443 implies that a one-standard-deviation increase in firm-level connectedness raises monthly returns by roughly 0.154 percentage points, on the order of 1.9% annualised. FirmTC is also the only regressor in the table that is not decile-scaled, so its coefficient is not comparable with the others and the reader must perform this arithmetic unaided. The finding sits awkwardly with the paper's own interpretation: if technologically connected analysts make prices more efficient, firms they cover should not earn systematically higher subsequent returns. More consequentially, whatever characteristic drives the main effect may also drive the interaction. The saturated fixed effects in Columns (3) through (6) address size, book-to-market, coverage and innovation intensity one at a time but not jointly, and never address liquidity, institutional ownership or idiosyncratic volatility — the characteristics most closely tied to the speed of information diffusion and hence to the strength of any lead-lag effect. A useful falsification would run the same interaction using firm-level averages of the paper's other connectedness measures; if supply-chain or geographic connectedness attenuates technology momentum equally well, the technological reading would be hard to sustain. The test would also be more persuasive in portfolio form evaluated against a factor model, particularly since the introduction invokes an 84 basis point six-factor alpha from Lee et al. (2019) while Table 6 estimates a raw-return panel regression, and since a DGTW benchmark-adjusted return is already defined in Appendix A6 but never used. Finally, FirmTC is measured contemporaneously in quarter t while every other regressor is lagged, which introduces an avoidable look-ahead into a return-predictability regression.

3.3 The interpretation of the innovation result outruns the evidence

Section 5.1 concludes that high-TechConnect analysts discipline low-value patenting, and that firms expand lower-impact filings once that scrutiny is removed. The pattern is consistent with the hypothesis, but it is equally consistent with at least two alternatives the paper does not test. The first is the mechanism the paper itself cites: He and Tian (2013) argue that coverage suppresses innovation through short-term earnings pressure, and the paper's own Column (1) reproduces that result. Since Table 1 shows that TechConnect rises with firm size, R&D and coverage, high-TechConnect analysts are not a random subset, and the earnings-pressure account predicts the same sign. The second is general monitoring: when coverage falls, managerial discretion rises and managers expand visible output regardless of value. Distinguishing these requires evidence the paper does not yet supply — that high-TechConnect analysts were observably more critical of low-quality innovation before the shock, that the response is concentrated where managers had prior incentives to file low-value patents, or that R&D spending responds upstream of filing. There is also a framing problem. He and Tian treat coverage-induced reductions in innovation as a cost; this paper treats a directionally identical result as a benefit on the grounds that the suppressed patents fall outside the top citation decile. That reading requires evidence that the marginal suppressed patent had negative value, which a citation-percentile split does not establish. Reporting the Kogan et al. dollar value of the marginal patents, or a downstream R&D-productivity outcome, would be more convincing than the citation rank alone.

3.4 Conditioning tests and inconsistencies in the exposition

Several of the conditioning tests in Section 4 are constructed in ways that make a supportive result close to mechanical. The most important is the technology-specialisation split, which classifies colleagues by whether their portfolio-average PatentValue exceeds the median — the very quantity that weights Equation (3). Splitting on the magnitude of the weights guarantees that the high group carries most of the variation in the pooled measure, so the resulting coefficient ordering says little about whether specialised colleagues are better information sources. Defining specialisation on a dimension orthogonal to the weighting, such as the concentration of a colleague's coverage within technology space, would make the test informative. More broadly, Section 4 reports eleven conditioning results with no adjustment for multiple testing and no pre-specification, and several of the significant interactions carry t-statistics between 2.0 and 2.4.

A number of internal inconsistencies also need correction, because collectively they make the reported numbers hard to verify. Section 4.2 directs readers to Panel A for results that appear in Panel B, and Section 4.3 directs them to Panel B for results in Panel C. Section 4.3 states that the benefit is larger for firms with "more analyst coverage," whereas the interaction coefficient of +0.175 and the introduction both imply less coverage. InnovationIntensity is defined as a log of a patent-value ratio in Section 4.3 and as a ratio of logs of citation counts in Appendix A6; the descriptive statistics, with a mean of −1.81 and a minimum of −9.63, are consistent only with the former, and the variable is used both as a conditioning variable in Table 4 and as the basis for fixed effects in Table 6. Section 6.1 reports that "in five out of six columns, the coefficients on TechConnect are negative and significant at the 1% level," whereas Appendix A3 shows four of the six as positive — correctly so, since Score and Accuracy are increasing in accuracy — and all six as significant at the 1% level, so the text understates its own results while misdescribing their signs. Finally, sample sizes shift across tables without documentation: 869,651 in the main analysis, 860,732 in Table 4 Panels B and C, 856,216 in Appendix A3, 894,895 in Table 1 Column (2) — exceeding the stated final sample — 1,043,699 in Table 2 Column (1), and 2,582,456 and 2,651,813 in Table 4 Panel A. A short data-construction appendix documenting attrition to each estimation sample would resolve this.

Minor Issues

1. Sample period. The abstract, Section 2.1 and Appendix A2 describe the sample as January 1995 to December 2021, while Section 2.1 also states that the final sample runs from January 1995 to March 2022. These should be reconciled and the treatment of the partial quarter clarified.

2. Descriptive statistics in Appendix A2. ForecastHorizon is reported with a mean of 5.48 against percentiles of 23, 61 and 86, which is arithmetically impossible. The rightmost observation-count column also appears to carry 99th-percentile values shifted one position for several variables, leaving sample sizes unreported. Separately, Star and BoldNegative share three moments to near-identical precision (0.163, 0.370/0.369, 1.82), and PastPerformance reports an unusual profile for a percentile-type score, with a 99th percentile of 65 against a minimum of 0.

3. An apparently mis-scaled coefficient. In Table 2, Column (4), the coefficient on AnalystTenure is 27.831 against 0.009 and 0.031 in Columns (2) and (3). Tenure is nearly collinear with analyst fixed effects, but a coefficient three orders of magnitude larger than its neighbours suggests a scaling problem.

4. The motivating anecdote sits awkwardly with the measure. Footnote 5 reports that Apple–Tesla technological similarity rose from 0.07 in 2020 to 0.35 in 2023. The Huberty–Jonas exchange described in Section 1 is dated 2020, when measured similarity was near the unconditional cross-firm average of 0.061, so the paper's own measure would not have flagged it.

5. Event-window descriptions conflict. Section 3.2 and the Table 3 note describe a 24-month window, while the Figure 3 note describes "eight event quarters within a 12-month window." The time fixed effects are labelled year-quarter in Equation (5) and the Table 3 note but calendar-quarter in the discussion of Equation (7). Table 3 also reports 20,974 observations in Column (1) and 20,768 in Columns (2) and (3) without explanation.

6. Clustering is described two ways for the same coefficients. The Table 5 note states that standard errors are clustered at the event level; the Figure 4 note, which plots the Column (3) coefficients, states that they are clustered at the firm and year levels. The Figure 4 note also appears to omit "the median of" when defining high- and low-TechConnect analysts and refers only to the closure date although the sample combines mergers and closures.

7. The merger treatment definition cannot be read literally. Section 5.1.1 states that for mergers the analysis focuses on "firms that are covered by two analysts and one analyst is dropped due to redundancy," which cannot be reconciled with the mean treated-firm coverage of 17.4 analysts in Panel A. The intended meaning is presumably coverage by an analyst at each merging brokerage. The number of treated firms from mergers and from closures should also be reported separately.

8. Sample construction in the innovation analysis. Fifty-nine events, 1,662 firms, five matched controls and six event years do not obviously reconcile with the 34,016 observations reported, and Column (3) reports 28,099 without explanation. An attrition table would help.

9. Date conventions and thresholds in the patent analysis. Technological similarity is built from grant dates while the innovation outcome uses filing dates; the inconsistency should be acknowledged and the measure shown robust to an application-date construction. The top-10% citation threshold defining high-impact patents is also unmotivated, and no sensitivity to alternative cut-offs or to value-based quality proxies is reported.

10. Table 6 reporting. The note lists R&D intensity among the controls, but no such variable appears in the table or in the Section 5.2 control list. The MV coefficient in Column (3) is −56.254 against values between −0.166 and −0.279 elsewhere. The classification underlying the industry fixed effects is unspecified, and the standard deviation of FirmTC used in the magnitude calculation is never reported.

11. Calibration of language. Section 5.2 describes coefficients as remaining "largely unchanged" although the interaction falls to −2.317 (t = −1.94) once size-quintile fixed effects are added, and concludes that the results provide "strong evidence" on the basis of interaction t-statistics between −1.94 and −2.31. Section 3.2 attributes the fall in the event-quarter +4 coefficient from −10.212 to −4.226 to imprecision without reporting standard errors that would distinguish decay from noise.

12. Undocumented variable construction. Analyst gender and cohort membership drive two of the strongest results in Table 4 Panel B, but I/B/E/S records neither; the inference procedures, their coverage and their likely error rates should be described. Several coefficients in Appendix A3 print as "0.000" with negative t-statistics, obscuring their sign. SupplyChainConnect is built from a single 2002 benchmark input-output table across a 27-year sample. The pooled sample median AFE used for two magnitudes in Section 4.1 is never reported, and Appendix A1 tabulates 47 industries with Coal reported as "NA" while footnote 3 computes a figure out of 48.

13. Unused variables. DGTW-adjusted returns, three-day cumulative abnormal returns around recommendations, BoldPositive, BoldNegative, Revision and SignedRevision are all defined in Appendices A2 and A6 but appear in no table. Either the corresponding analyses should be reported, which would strengthen the mechanism and market-efficiency evidence, or the variables should be removed.

14. Design elements that would strengthen the identification. No cross-sectional heterogeneity is examined inside the merger design, although Table 4 reports extensive heterogeneity for the panel result. Only three pre-treatment quarters support the parallel-trends claim. A 15-day exclusion window is applied when classifying analyst movements but not to the outcome sample, unlike the six-month gap used at the firm level. And the exclusion restriction for closures, which cluster in market downturns and reduce coverage by 2.22 analysts on average, is not discussed.

Missing and Inadequately Integrated Literature

Cited but Inadequately Integrated

The paper draws its similarity measure from Jaffe (1986) and Bloom et al. (2013) but does not engage with what those studies established about its properties. Jaffe's construction was designed to capture technological opportunity for R&D spillover estimation rather than to serve as a proximity metric for information transmission, and Bloom et al. explicitly separate technology-space proximity from product-market proximity and show that the two carry opposite-signed effects. The present paper treats technological proximity as though it mechanically implies informational relevance, and should address whether patent-class overlap identifies firms whose shocks are correlated — Bloom et al.'s concern — or firms about which knowledge transfers, which the argument requires. A related point applies to Kogan et al. (2017): the patent-value weights are constructed from announcement-window stock returns, which introduces a mechanical link between the independent variable and equity-market information, and this is particularly consequential for Table 6, where the dependent variable is also a return.

Several analyst-literature citations are similarly underused. Phua et al. (2023) is invoked to establish that the within-industry channel is mechanically excluded, but exclusion at the industry-code level is not economic exclusion, as Appendix A1 makes clear. Huang et al. (2022) supplies the supply-chain analogue and the SupplyChainConnect measure, yet the question of whether the two results reflect distinct phenomena or two measurements of the same colleague-interaction propensity is never tested; entering both measures interactively, or showing the technological effect among firms with no supply-chain link to colleague coverage, would establish separability. Hwang et al. (2019) is cited once in passing although its subject — distinguishing genuine information sharing from common information sources — is precisely the identification problem this paper faces. Groysberg (2010) is invoked to explain the null co-location result, while its central finding, that analyst performance is largely firm-specific and does not port across employers, bears directly on the premise that colleague composition drives individual performance. Martens and Sextroh (2021) is discussed as the closest related work, but their finding that coverage overlap predicts cross-citation of patents is an alternative mechanism for the present results rather than only a contrast. On the asset-pricing side, Ali and Hirshleifer (2020) show that shared analyst coverage unifies a family of momentum spillover effects including technology momentum, which is arguably the closest competing explanation for Table 6, and Engelberg et al. (2020) document that analysts are systematically misaligned with anomaly signals, which raises the question of why technological information should be different. Finally, Manski (1993) is cited to motivate the identification discussion but never revisited; the paper should state which of the endogenous, contextual and correlated effects the merger design resolves and which it does not.

Truly Missing Papers

Four omissions bear directly on measurement. Lerner and Seru (2022) documents that firm-level patent and citation data suffer truncation and composition biases that survive standard adjustments and correlate with firm characteristics, and supplies an explicit checklist for finance researchers; this is directly relevant to the truncation concern in the innovation analysis and to the citation-decile definition of patent impact. Kelly, Papanikolaou, Seru and Taddy (2021) construct text-based measures of patent novelty and impact designed to overcome the limitations of classification-based similarity, with public patent-level data, offering a natural robustness check on the 669-class cosine measure. Kadan, Madureira, Wang and Zach (2012) examine analysts' across-industry versus within-industry expertise directly and find genuine cross-industry ranking ability, providing both a foundation for the paper's premise and an alternative empirical handle on it. Bradley, Gokkaya and Liu (2017) show that analysts with pre-analyst industry work experience forecast more accurately and that the exogenous loss of such analysts has measurable market consequences, which is the closest existing evidence that domain expertise rather than network position drives analyst performance.

Four further omissions bear on identification and interpretation. Sun and Abraham (2021) and Callaway and Sant'Anna (2021) develop heterogeneity-robust estimation for event studies and difference-in-differences with staggered timing, and the former shows specifically that flat pre-period coefficients cannot be read as validating parallel trends when treatment effects are heterogeneous — directly relevant to Figure 3 and to Table 3, Column (3). Chen, Harford and Lin (2015) use the same brokerage closures and mergers to show that coverage declines are followed by increased managerial expropriation, value-destroying acquisitions and earnings management, supplying a competing non-technological account of the "removal of informed scrutiny." Israelsen (2016) shows that stocks with overlapping analyst coverage exhibit excess comovement because individual analysts' forecast errors are correlated across covered firms, identified using brokerage mergers, which supplies a competing mechanism for the technology-momentum result in which shared coverage generates correlated pricing errors rather than improved price discovery.

Reference List

Bradley, D., S. Gokkaya, and X. Liu. 2017. Before an Analyst Becomes an Analyst: Does Industry Experience Matter? Journal of Finance 72(2):751–792.

Callaway, B., and P. H. C. Sant'Anna. 2021. Difference-in-Differences with Multiple Time Periods. Journal of Econometrics 225(2):200–230.

Chen, T., J. Harford, and C. Lin. 2015. Do Analysts Matter for Governance? Evidence from Natural Experiments. Journal of Financial Economics 115(2):383–410.

Israelsen, R. D. 2016. Does Common Analyst Coverage Explain Excess Comovement? Journal of Financial and Quantitative Analysis 51(4):1193–1229.

Kadan, O., L. Madureira, R. Wang, and T. Zach. 2012. Analysts' Industry Expertise. Journal of Accounting and Economics 54(2):95–120.

Kelly, B., D. Papanikolaou, A. Seru, and M. Taddy. 2021. Measuring Technological Innovation over the Long Run. American Economic Review: Insights 3(3):303–320.

Lerner, J., and A. Seru. 2022. The Use and Misuse of Patent Data: Issues for Finance and Beyond. Review of Financial Studies 35(6):2667–2704.

Sun, L., and S. Abraham. 2021. Estimating Dynamic Treatment Effects in Event Studies with Heterogeneous Treatment Effects. Journal of Econometrics 225(2):175–199.

Conclusion and Path Forward

This manuscript pursues a question that deserves an answer, and it does so with considerable effort. The observation that technological relatedness crosses industry boundaries more often than it respects them is genuinely interesting; the proposition that brokerage research departments aggregate latent technological knowledge is original; and the attempt to trace a single information channel from analyst forecasts through corporate innovation to asset prices is more ambitious than most papers in this literature attempt. Several design choices — the merger × analyst × firm fixed-effects structure, the arrival-versus-departure decomposition, the use of fixed-effects Poisson for count outcomes, and the reporting of covariate balance and dynamic effects rather than merely asserting them — reflect the instincts of a careful empiricist.

The concerns raised here are foundational rather than cosmetic, but most of them can be addressed with the data already assembled. On measurement, the priorities are to report equal-weighted and distribution-robust versions of TechConnect, to show what the effect looks like across the measure's quintiles rather than only at a one-standard-deviation move that exceeds the 75th percentile, and to demonstrate that the result survives among genuinely distant industry pairs and outside the financial sector. On mechanism, a placebo built from non-colleagues at other brokerages is the single most valuable test the paper could add, and at least one revision-based outcome would convert a correlation into evidence of an informational response. On identification, the merger design needs a first stage, a continuous dose-response specification, event-level inference, cohort-interacted time effects, and a controls-free variant; the innovation design needs matching on pre-event patenting and R&D, descriptive comparisons of the two treated subgroups, a sample restricted to cohorts with complete grant windows, and joint specifications that permit formal tests of the high- versus low-impact and placebo comparisons. On interpretation, the economic benchmarks should be recomputed from a single specification, the horizon effects reported net rather than as differentials, the positive FirmTC return coefficient confronted directly, and a consistency pass carried out over cross-references, variable definitions and sample sizes.

That is a substantial programme, but it is a well-defined one, and the author is clearly capable of executing it. Whether the internal organisation of research departments shapes which kinds of information reach the market is a question worth settling, and a revision along these lines would go a long way toward settling it.