{"id":"adc67317-4aa7-4dea-94b5-42426660b5cb","arxiv_id":"2603.00248","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"Targeted local projections shrink LP impulse responses toward SVAR estimates at each horizon with data-driven weights, then use a double bootstrap for inference.","lead":"This paper proposes a new estimator, targeted local projections, that blends local projection and vector autoregression impulse-response estimates horizon by horizon to reduce long-horizon noise. Applied macroeconomists could get shorter confidence bands with near-nominal coverage, but the theoretical derivation of the optimal blend is flawed.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The risk criterion in Eq. (20) is not consistent: under Theorem 2 the gap term converges to a random variable, so the closed-form weight (21) is not MSE-optimal and Theorem 5's normality is unsupported.","rationale":"The reader's weakest assumption identifies the same load-bearing weakness: the squared gap cannot be a consistent estimator of squared bias under the paper's own local asymptotics. I agree with that diagnosis. This is central because the contribution is the closed-form, data-driven MSE-optimal weight; without Theorems 3–4, TLP is an ad hoc shrinkage estimator that happens to work in simulations. The finite-sample simulations and the MSDB bootstrap are credible and potentially useful, and I am not disputing the empirical evidence. But the theoretical claim—risk consistency, closed-form optimality, and asymptotic normality under local misspecification—is internally inconsistent with Theorem 2. The Appendix B proof confirms the problem through the E[β]^2/E[β^2] conflation. The idea may be salvageable with a SURE-style correction that subtracts the estimated variance of the gap from T·gap^2, and with a reworked asymptotic distribution for random weights. As written, the paper's main theoretical result is not supported, so the rejection stands.","tokens_in":18686,"tokens_out":8915,"duration_ms":81111,"concrete_test":"Independently re-derive the feasible risk from definition (18) under Theorem 2, explicitly computing E[(β̂_TLP − β*)^2] = (1−v)^2 E[(β̂_LP − β̂_VAR)^2] + v^2 Var(β̂_LP) + (1−v)^2 Var(β̂_VAR) + 2v(1−v) Cov(β̂_LP, β̂_VAR), using E[T(β̂_LP − β̂_VAR)^2] → aBias_h^2 + Σ_LP + Σ_VAR − 2Σ_COV. Then check whether (20) equals (1−v)^2 aBias_h^2 + [v^2Σ_LP + (1−v)^2Σ_VAR + 2v(1−v)Σ_COV]. If not, (20) is not consistent and (21) is not the MSE-minimizing weight. A numerical cross-check at T=800 on the VARMA(1,1) local DGP, comparing the argmin of (20) over a v-grid with the Monte Carlo argmin of (18), would settle the practical impact.","verdict_should_be":"REJECT","load_bearing_attack":"Under Assumptions 1–2 with ζ=1/2, Theorem 2 implies √T(β̂_LP − β̂_VAR) →_d Z ∼ N(−aBias_h, Σ_LP + Σ_VAR − 2Σ_COV). Hence T(β̂_LP − β̂_VAR)^2 in Eq. (20) converges in distribution to Z^2, not in probability to aBias_h^2. The weight in Eq. (21) therefore converges to a non-degenerate random variable, not to the constant v in Eq. (24), and Theorem 5's claim that v(λ̂) →_p v with asymptotic normality of β̂_TLP is invalid. The proof of Theorem 3 (Appendix B) also substitutes E[β̂^2] for (E[β̂])^2: Eq. (46) adds Σ_LP to T(β*)^2 under the label “squared expectation,” and Eq. (49) puts variance terms inside the squared-bias component. Consequently Eq. (20) double-counts variance—even with a correct gap estimator one would need to subtract Σ_LP + Σ_VAR − 2Σ_COV from T·gap^2, not multiply the variance block by 2. The central claim that (21) minimizes the MSE (18) is therefore not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Targeted Local Projections (TLP), a ridge-type shrinkage estimator that forms a horizon-specific linear combination of LP and VAR impulse-response estimates. The shrinkage weight is chosen by minimizing a feasible estimator of the asymptotic mean-squared error, yielding a claimed closed-form solution. The paper also proposes a Mean Symmetric Double Bootstrap (MSDB) for inference and reports simulations comparing TLP with LP, VAR, SLP, and BLP under locally misspecified VARMA DGPs.","tokens_in":19047,"tokens_out":10449,"duration_ms":101982,"significance":"The idea of shrinking LP toward a VAR target with data-driven, horizon-specific weights is practically appealing, and the simulation study is reasonably broad, including different sample sizes, lag lengths, misspecification strength, and GARCH errors. The double-bootstrap proposal is a useful practical ingredient. However, the central theoretical contribution is invalid: the feasible risk criterion is not a consistent estimator of the stated MSE, the closed-form weight does not minimize the population MSE, and the asymptotic-normality claim for TLP is unsupported. What remains is a heuristic estimator with simulation support, rather than the theoretically justified procedure claimed in the abstract and theorems.","major_comments":[{"comment":"The proof of Theorem 3 confuses the second moment with the squared mean. Eq. (46) assigns Sigma_LP to T(E[beta_hat_LP])^2, but this variance term belongs to E[(beta_hat_LP)^2], not to the square of the expectation. Eq. (49) then puts variance terms inside the squared-bias component, and Eq. (51) adds the same variance block a second time through the factor 2. Consequently, the feasible criterion R_hat_h in Eq. (20) is not a consistent estimator of R_h in Eqs. (18)-(19), and Theorem 3 is not established.","section":"Appendix B, Eqs. (46)-(51); Theorem 3, Eq. (20)"},{"comment":"Under Assumptions 1-2 with zeta=1/2, Theorem 2 implies sqrt(T)(beta_hat_LP - beta_hat_VAR) converges in distribution to a nondegenerate normal random variable Z. Hence T(beta_hat_LP - beta_hat_VAR)^2 converges in distribution to Z^2, not in probability to aBias_h^2. Therefore the 'bias' term in Eq. (20) and the weight in Eq. (21) converge to nondegenerate random variables, contradicting the definition of the constant v in Eq. (24). The claim in Theorem 5 that v(lambda_hat) ->_p v, and the resulting asymptotic normality of TLP, is unsupported.","section":"Theorem 2 vs. Theorem 3, Eqs. (20)-(21), (24)-(25)"},{"comment":"Even when the VAR is correctly specified, beta_hat_LP - beta_hat_VAR is O_p(T^{-1/2}), so T times its square does not vanish. Eq. (22) then yields a random limit weight in (0,1), not zero. Thus the statements that TLP 'reduces to SVAR' under no misspecification (Section 2.3 and the discussion after Theorem 4) are contradicted by the authors' own formula. This is not a minor caveat but a direct consequence of the same missing variance correction that invalidates Theorem 3.","section":"Section 2.3 and discussion after Theorem 4"}],"minor_comments":[{"comment":"The notation 'argmin T E[...]' is imprecise; the criterion is minimized over lambda_h, so it should be min_{lambda_h} or argmin_{lambda_h} of the expectation.","section":"Eq. (18)"},{"comment":"The statement uses v(lambda_hat) without subscript h, and beta^* without subscript h, although all objects are horizon-specific. Please make the notation consistent.","section":"Theorem 5"},{"comment":"The DGP labeled 'VARMA(1,100)' is referred to as having the 'highest degree of misspecification', but the truncation of the infinite MA polynomial at 100 lags is not explicitly defined in the text; please clarify.","section":"Section 5"},{"comment":"The bootstrap validity discussion cites Brüggemann et al. (2016) for the correct specification case, but no formal theorem is given for the double bootstrap under the local misspecification used in the main text; a precise statement of the regularity conditions and convergence result would help.","section":"Section 4.1"}],"recommendation":"reject","confidential_remarks":"The stress-test concern is accurate. The main proofs in Appendix B contain a fundamental confusion between squared means and second moments, and the consistency claim in Theorem 3 is not a minor technical gap: at zeta=1/2 the squared gap cannot consistently estimate the squared bias because its asymptotic variance is nonvanishing. The resulting weight is random in the limit, so Theorem 5 cannot be salvaged by small corrections. A rejection is appropriate; a future version could reframe TLP as a heuristic/bootstrap-based shrinkage method with weaker claims, but that would be a substantially different paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the TLP estimator is a plausible addition to the local-projections toolkit, and the simulation work is serious. But the paper's central theoretical result—the closed-form weight that minimizes MSE—does not hold, because the risk criterion in Eq. (20) is not a consistent estimator of the MSE in Eq. (18), and the proof in Appendix B confuses second moments with squared expectations.\n\nWhat's actually new: horizon-specific frequentist shrinkage of LP toward a VAR, with double-bootstrap inference. That specific combination is not in the cited literature. The positioning against BLP and SLP is fair, and the simulations cover relevant DGPs (VARMA with local misspecification, GARCH), with honest comparison across LP, SVAR, SLP, BLP, and TLP. The double-bootstrap procedure is described in enough detail to replicate.\n\nThe soft spot is load-bearing. Under the paper's own Theorem 2, √T(β̂_LP − β̂_VAR) has a non-degenerate normal limit, so T(β̂_LP − β̂_VAR)² converges to a random variable, not to the constant aBias_h². That means the weight in Eq. (21) converges to a random variable, not to the v in Eq. (24), and Theorem 5's asymptotic normality claim is unsupported. Appendix B makes the same problem worse: Eq. (46) writes T(E[β̂_LP])² as T(β*)² + asymptotic variance + o(1), which confuses E[β̂²] with (Eβ̂)²; Eq. (49) puts variance terms inside the squared-bias component; and Eq. (51) then adds variance twice. So the claim that (21) minimizes MSE is not established. This isn't a minor typo—it's the main theoretical contribution.\n\nThat said, the idea may be salvageable with a proper unbiased risk estimator (a SURE-style criterion that subtracts the variance terms correctly) and with inference that acknowledges the weight is random under local misspecification. The simulation evidence is not nothing; it suggests the method can work in practice even if the theory is currently wrong.\n\nWho this is for: applied macroeconomists looking for a variance-reduction trick for LPs would get something from the simulations, but they shouldn't cite the theory as it stands. It deserves a serious referee—this is a relevant, well-motivated question with extensive simulations, so a desk reject would be too harsh. I'd send it out, expecting major revision or a conditional reject. A careful referee should check whether the weight-selection theory can be fixed.","headline":"A plausible shrinkage estimator undermined by a load-bearing proof error: the risk criterion double-counts variance and the weight doesn't converge to a constant.","tokens_in":19524,"tokens_out":3400,"would_cite":false,"duration_ms":32420,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62P20","62F40","62J07"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes targeted local projections, a shrinkage estimator that combines LP and SVAR impulse responses horizon-by-horizon with a closed-form MSE-optimal weight, and a double bootstrap for coverage.","keywords":["targeted local projections","local projections","structural VAR","impulse response functions","mean squared error","shrinkage","double bootstrap","misspecification"],"falsifier":"Simulate the local-to-VAR DGP with bias of order T^{-1/2}, compute T(βhat_LP − βhat_VAR)^2 across many samples for large T, and check whether its distribution collapses to a point (the squared bias) or remains spread out; a spread-out limit means the weight is random and the claimed asymptotic normal approximation for TLP fails.","tokens_in":18557,"feed_emoji":"🎯","tokens_out":5655,"duration_ms":56356,"temperature":0.7,"pith_summary":"This paper proposes a frequentist shrinkage estimator—targeted local projections (TLP)—that replaces the plain local projection impulse response at each horizon with a linear combination of the LP estimate and the corresponding structural VAR estimate. The combination weight is chosen to minimize an estimated mean-squared error, balancing the LP's small bias against the VAR's lower variance, and it has a closed form. The paper argues this reduces variance at long horizons while introducing only modest bias, and that a mean-symmetric double bootstrap restores near-nominal coverage in small samples despite the VAR's asymptotic bias. If correct, the method gives macroeconomists a computationally cheap way to obtain tighter impulse-response confidence intervals than LP alone.","feed_headline":"Shrink local projections toward VARs to halve long-horizon uncertainty","feed_subtitle":"Targeting LP estimates toward SVAR counterparts preserves coverage while halving interval lengths in simulations.","key_machinery":"The key object is the shrinkage weight v(λ) = X'X/(X'X + λ), which after projecting out controls makes TLP a convex combination of the two estimators. The optimal λ (hence v) is obtained by minimizing a feasible risk criterion that estimates squared bias by the squared gap between the LP and VAR estimates plus twice the estimated variance of the combination. The closed-form weight in equation (21) and the mean-symmetric double bootstrap that centers on the bootstrap mean and uses symmetric intervals are what carry the argument.","core_discovery":"The central claim is that the TLP estimator, defined as the penalized LP solution after projecting out controls, equals v(λ) times the LP estimate plus (1−v(λ)) times the VAR estimate, with v determined by the data so as to minimize an asymptotic MSE criterion. Under local misspecification (VAR bias of order T^{-1/2}), the paper derives a unique closed-form weight, shows the estimator is asymptotically normal with bias smaller than the VAR's and variance smaller than the LP's, and provides a double-bootstrap inference procedure that maintains near-nominal coverage in simulations.","pith_inferences":["The same targeting idea could be applied to smooth local projections or to any pair of estimators with complementary bias-variance profiles, replacing the VAR target with any lower-variance estimator.","Because the weight depends on a squared gap that is only a noisy estimate of the bias in finite samples, the double bootstrap may be doing more work than the asymptotic theory suggests; a finite-sample study of the weight distribution would clarify this.","The horizon-specific weights could be exploited as a diagnostic: a sudden shift of weight toward the VAR at long horizons is a tell that LP variance is dominating, and the method's confidence intervals at those horizons reflect that trade-off."],"forward_implications":["Applied researchers can obtain materially narrower confidence intervals for long-horizon impulse responses without sacrificing nominal coverage, provided enough lags are used in the target VAR.","The closed-form weight makes the estimator as easy to compute as a ridge regression, with no numerical optimization per horizon.","The double-bootstrap procedure offers a general template for inference when an estimator is a linear combination of a robust and a biased-but-efficient estimator.","The paper's theory clarifies when it is optimal to trust the VAR: essentially when the two estimators are close relative to their variance difference."],"fun_headline_variants":["Optimal blend of local projections and VARs reduces uncertainty","Shrink local projection variance by mixing with VAR estimates","Data-driven LP-VAR combo cuts long-horizon variance","Targeted LP estimator balances bias and variance via VAR","Merging LP and VAR estimates yields lower MSE at long horizons"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole optimal-weight construction assumes that T(βhat_LP − βhat_VAR)^2 consistently estimates the squared asymptotic bias of the VAR, which requires the gap that appears in the numerator to converge to a constant rather than a random variable.","fun_headline_variants_meta":{"raw":{"variants":["Optimal blend of local projections and VARs reduces uncertainty","Shrink local projection variance by mixing with VAR estimates","Data-driven LP-VAR combo cuts long-horizon variance","Targeted LP estimator balances bias and variance via VAR","Merging LP and VAR estimates yields lower MSE at long horizons"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000175,"raw_usage":{"total_tokens":1087,"prompt_tokens":672,"completion_tokens":415,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":416,"completion_tokens_details":{"reasoning_tokens":333}},"tokens_in":416,"tokens_out":415,"duration_ms":4986,"temperature":1.0,"reasoning_tokens":333,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T19:59:25.552320+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the local-to-VAR DGP with bias of order T^{-1/2}, compute T(βhat_LP − βhat_VAR)^2 across many samples for large T, and check whether its distribution collapses to a point (the squared bias) or remains spread out; a spread-out limit means the weight is random and the claimed asymptotic normal approximation for TLP fails.","supporting_citations":[],"review_version":1}