{"id":"a1aff270-7ee5-407a-9a0f-f051f6197b6c","arxiv_id":"1908.05476","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"First-price auction primitives, including the distribution of the number of active bidders, are nonparametrically identified from winning bids alone when bidders observe competition, via density discontinuities at conditional support boundaries.","lead":"This paper shows that in first-price auctions where bidders know how many rivals they face, the distribution of the (unobserved) number of bidders and the distribution of private values can be recovered nonparametrically from winning bids alone, using kinks and jumps in the winning bid density. The method matters for empirical auction work in procurement, timber, and housing, where only transaction prices may be recorded.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Identification hinges on Assumption N's contiguous support: if P(N=k)=0 for any intermediate k, the discontinuity count misidentifies the maximum and the iterative value-quantile recursion cannot start; no relaxation is provided.","rationale":"I read the paper in good faith and checked the main identification chain. The equilibrium mappings (2) and (4), the jump formula (9), the identification of the support endpoints, and the iterative quantile extrapolation of Lemma 2.3 are all internally consistent within the stated model. The formulas for v and p_n in Lemma 2.2 follow algebraically from the jump sizes and the normalization that the p_n sum to one. The theorem is therefore correct under Assumption N. The load-bearing concern is that Assumption N is not a minor regularity condition: it requires the support of N to be a full integer interval with every interior probability positive. The discontinuity count identifies the number of support points, not their labels, so a gap in the support breaks both the support recovery and the recursion that starts from the largest number of bidders. The paper offers no test for such gaps and no relaxation, so the advertised result is narrower than the abstract suggests. This is a limitation of scope rather than a flaw in the proof, so it does not overturn the theorem. The reader's weakest assumption identifies the same issue and the empirical and abstract-to-body mismatches also support a conditional verdict, so I recommend no change to the reader's conditional assessment.","tokens_in":33519,"tokens_out":12907,"duration_ms":126981,"concrete_test":"Simulate the model with private values uniform on [0,1], Assumption IPV, and N taking values 2 and 4 with equal probability. Compute the unconditional winning bid density from equation (10): it has exactly two discontinuities, at b_2 and b_4. Apply Lemma 2.2: the tail index gives the minimum as 2; with two discontinuities the formula gives maximum 3; equations (11)-(12) then return probabilities p_2 and p_3 summing to one, which cannot equal the true (1/2, 1/2). Then run the Step 1-3 recursion treating the second discontinuity as b_3: Lemma 2.3's sequence uses B_2, whose quantile is identified only on a wrong interval, and the recovered V differs from the true V(alpha)=alpha on a positive-length quantile interval. Confirming these discrepancies shows that Assumption N is load-bearing and needs either a relaxation or an explicit scope statement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central identification argument is internally coherent, but it is only as strong as Assumption N, which is a substantive support condition rather than a regularity condition. Lemma 2.1 and Lemma 2.2 identify the support of N by counting density discontinuities and then setting the maximum as the minimum plus the number of discontinuities minus one. This identifies the number of support points, not their integer labels. If P(N=k)=0 for some k between the minimum and the maximum, no jump occurs at b_k, so the count is too small and Lemma 2.2 returns a wrong maximum; the 'identified' probabilities in equations (11)-(12) are then fitted to a support that is not the true one. The difficulty is not merely that one probability is zero: the recursive construction in Lemma 2.3 also fails because Step 1 needs the interval [b_{n-1}, b_n] starting from the top, and the upper boundary for the missing intermediate number is not a discontinuity. Since the minimum and maximum of N are not directly observed, the data cannot distinguish a genuine gap from a zero-probability support point, and the paper gives no test or relaxation. The abstract's general claim that the resulting discontinuities can be used to identify the distribution of N therefore overstates robustness: the identified object is the distribution conditional on a full-interval support assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies nonparametric identification of first-price auction primitives when the analyst observes only winning bids and the number of active bidders is unobserved. Under the maintained assumptions that N is exogenous, independent of private values, has support equal to a full integer interval, and is observed by the bidders, it shows that discontinuities in the winning bid density at conditional bid upper bounds identify the support and probabilities of N, and that an iterative use of the equilibrium bid-quantile and value-quantile mappings identifies the private value distribution. The framework is then extended to endogenous participation through a reserve price or an entry cost, with and without bidder knowledge of participation, yielding testable restrictions for information and entry models. An empirical illustration using USFS timber auctions concludes that many reported three-bid auctions have only two competitive bidders and that risk-aversion estimates are sensitive to this form of unobserved competition.","tokens_in":33711,"tokens_out":7451,"duration_ms":72014,"significance":"If the identification result holds, it is a substantial contribution: the winning bid alone, through the location and size of density discontinuities, identifies both the distribution of competition and the private value distribution without instruments or multiple bids. The proofs are coherent and carefully structured; the jump formula (9) is correctly derived from the equilibrium conditions, and the iterative identification argument in Lemma 2.3 is rigorous. The extension to endogenous participation with testable restrictions for whether bidders observe competition is valuable and connects cleanly to the discrete-mixture literature. The empirical application is suggestive but not decisive because of the very small subsamples and the heuristic nature of the discontinuity-detection algorithm.","major_comments":[{"comment":"Assumption N requires P(N=k)>0 for every integer k between n and n. The identification of n as n plus the number of density discontinuities minus one, and the formulas for p_n in equations (11)-(12), rely on a jump appearing at every intermediate boundary b_k. If P(N=k)=0 for some intermediate k, no jump occurs at b_k, so the count understates n and the probabilities are assigned to an incorrect support. The iterative construction in Lemma 2.3 also fails, because Step 1 uses the interval [b_{n-1}, b_n] starting from the top boundary of the next-lower component. Since n and n are not directly observed, the data cannot distinguish a genuine support gap from a zero-probability support point, and the paper provides no test or relaxation of this condition. The abstract's claim that the resulting discontinuities identify the distribution of N therefore overstates robustness; the identified object is the distribution of N conditional on a full-interval support assumption.","section":"Section 2.1, Lemma 2.2, Theorem 2.1"},{"comment":"The empirical conclusions rest on very small subsamples (45, 53, and 44 auctions), a k-NN discontinuity-detection algorithm with hand-set bandwidths h0=0.2 and h1=0.5 and epsilon=0.01, and no reported standard errors or sensitivity analysis. The point estimates in Table 1 (e.g., 0.90-0.95 for unobserved competition versus 0.45-0.53 from observed bids) and the CRRA lower bound of 0.9 in Section 4.2 are therefore not robustly supported; the text itself acknowledges that the variance is likely to be very high. The paper should provide bootstrap confidence intervals and a sensitivity analysis over (h0, h1, epsilon, K, M) before drawing conclusions about non-competitive bidding or risk aversion.","section":"Section 4, Tables 1-3, Appendix"}],"minor_comments":[{"comment":"The text says 'Expression (9) allows for the identification of p1 = p2 = 1/2', but p_1 is undefined because the support of N starts at n=2; the intended statements concern p_2 and p_3.","section":"Section 2.5.1"},{"comment":"The abstract included at the top of the submission mentions a parametric Bayesian estimation procedure and an application to Shanghai Government IT procurements, but the full text proposes no Bayesian procedure and applies the methods to USFS timber auctions. These statements should be reconciled.","section":"Abstract and Section 4"},{"comment":"The sentence 'we obtain the best lower bound for the CRRA coefficient θ 0.9' should read 'a lower bound of θ = 0.9' or 'the best lower bound is 0.9'.","section":"Section 4.2"},{"comment":"The displayed formula for the critical value c(epsilon;h0) is difficult to parse because of the mixed square-root and logarithmic terms; please format it more clearly and define each term.","section":"Appendix"}],"recommendation":"major_revision","confidential_remarks":"The theoretical core is sound under the stated Assumption N, and the proofs are careful. The main concern for publication is the gap between the general framing of the abstract and the genuine support condition of Assumption N, which is a substantive restriction rather than a regularity condition. The empirical section needs substantial revision (or clear repositioning as illustrative) before the applied claims can be taken as evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the central identification theorem is real: when N has contiguous support {n,...,n}, the winning-bid density discontinuities identify p_n, and the iterative quantile argument recovers F. I checked the algebra in equations (7), (9), (11), (12) and the induction in Lemma 2.3; it hangs together. Second, the paper does not deliver the Bayesian estimation or Shanghai procurement application advertised in its metadata; the full text is a USFS timber illustration with small samples and no standard errors. That mismatch needs fixing.\n\nWhat is new: the jump-size formula (9) is a genuinely new way to identify the mixture weights from a single bid, and the mix of support-location and jump-size information is clever. The extensions to reserve price and entry cost, and the nonidentification result when buyers don't observe N, are useful and clearly argued.\n\nSoft spots. The load-bearing assumption is Assumption N: every integer in [n,n] must have positive probability. If one intermediate N has probability zero, there is no discontinuity at its would-be boundary, so the count of jumps underestimates n-n and the recursion in Lemma 2.3 cannot start. This is not a niche technicality; it's the mechanism of the identification. The paper gives no test for gaps and no relaxation. The theorem is correct as stated, but the scope is narrower than the abstract's general claim. That should be made explicit and ideally relaxed.\n\nThe empirical application is illustrative at best. Sub-samples are 44-53 auctions, bandwidths are hand-chosen, and the estimated probabilities in Tables 1-2 have no confidence intervals. The CRRA bounds in Table 3 are built on the same fragile estimates. I wouldn't take any of the applied numbers as evidence about timber auctions.\n\nAlso, the omitted proofs of Proposition 3.6 and Lemma 3.2 matter little; those are supporting.\n\nWho is this for? Auction theorists and empirical IO researchers working on unobserved competition. The identification idea is worth engaging with seriously. A referee should spend time on Section 2 and ask for a statement of the support assumption's bite, plus a clean version with a coherent abstract. I would send it to peer review.\n\nRecommendation: engage with it. It deserves a serious referee, but only after the abstract/body mismatch is corrected.","headline":"The density-discontinuity identification of unobserved N is new and the theorem is sound, but the full-support assumption is narrow and the version I read does not match its own abstract's empirical claims.","tokens_in":34287,"tokens_out":3350,"would_cite":true,"duration_ms":37430,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91B26","62G05"],"pacs":[],"model":"deepseek-v4-flash","headline":"The winning bid alone can identify both the bidder-count distribution and the private-value distribution in a first-price auction.","keywords":["auction models","unobserved competition","nonparametric identification","density discontinuities","endogenous participation","unobserved heterogeneity","discrete mixture models","first-price auctions"],"falsifier":"Simulate the benchmark model with $F$ uniform on $[0,1]$, buyers observing $N$, and $N$ taking values 2 and 4 with equal probability. The winning-bid density has jumps exactly at $b_2$ and $b_4$, so the paper's rule $\\bar n = \\underline n + \\#\\{\\text{jumps}\\} - 1$ returns $\\bar n = 3$, and the estimated $p_3$ is positive although $P(N=3)=0$; this violates the true data-generating process and shows that Assumption N's full-support requirement is what carries the jump-counting step.","tokens_in":33247,"feed_emoji":"🔨","tokens_out":9974,"duration_ms":95371,"temperature":0.7,"pith_summary":"The paper tries to establish that, in a symmetric independent-private-value first-price auction, observing only the winning bid—not the number of bidders, not the losing bids—is enough to identify the full model. The reason is that each possible number of active bidders generates its own upper endpoint for the bid support, and the winning-bid density jumps at those endpoints; the locations and sizes of the jumps encode the support and the probabilities of the unobserved competition variable. A second iterative step, built from two equilibrium quantile mappings, then recovers the private-value distribution over its whole support. The same logic extends to endogenous participation through a reserve price or entry cost when participants observe competition, and the paper derives testable restrictions for whether they do. This matters because many real markets—informal procurement, housing 'bidding wars', subsidy competitions—record only the transaction price, and because recorded bid counts can misstate true competition.","feed_headline":"Winning bids alone reveal the hidden number of bidders","feed_subtitle":"Density jumps in transaction prices let analysts recover bidder counts and private values.","key_machinery":"The load-bearing object is the unconditional winning-bid density $g(b)$, which is a finite mixture $g(b)=\\sum_{n=\\underline n}^{\\overline n} p_n\\, n\\, G_n^{n-1}(b)\\, g_n(b)$. Equilibrium bidding makes the conditional support upper bounds $b_n$ strictly increasing in $n$, and Corollary 2.1 shows each conditional density is positive at its upper boundary, $g_n(b_n)=1/((n-1)(\\bar v-b_n))$. Hence $g$ has a jump at every $b_n$, with size $\\Delta_n=n p_n/((n-1)(\\bar v-b_n))$; the locations give the support of $N$, the sizes give the probabilities after $\\bar v$ is solved from $\\sum p_n=1$, and the iterative quantile mappings $V(\\alpha)=B_n(\\alpha)+\\alpha B_n'(\\alpha)/(n-1)$ and $B_n(\\alpha)=\\frac{n-1}{\\alpha^{n-1}}\\int_0^\\alpha t^{n-2}V(t)\\,dt$ extend identification of $V$ from the top quantile interval down to $\\alpha=0$.","core_discovery":"The paper's central claim is Theorem 2.1: under Assumptions N and IPV, if buyers observe the number of active bidders $N$, then the private-value c.d.f. $F(\\cdot)$ and the distribution of $N$ are identified from the winning-bid distribution alone. The argument uses the fact that the conditional bid quantile $B_n(\\alpha)$ is strictly increasing in $n$, so the support upper bounds $b_n=B_n(1)$ satisfy $\\underline v=b_{\\underline n}<\\cdots<b_{\\overline n}<\\bar v$, and that the conditional bid density is positive at each upper bound, $g_n(b_n)=1/((n-1)(\\bar v-b_n))$. These ingredients force the unconditional winning-bid density to jump at every $b_n$, with jump size $\\Delta_n=n p_n/((n-1)(\\bar v-b_n))$. Jump locations identify the support of $N$, jump sizes identify the probabilities $p_n$ once the upper bound $\\bar v$ is recovered from $\\sum p_n=1$, and an iterative quantile argument extends identification of $V(\\alpha)=F^{-1}(\\alpha)$ from the top bid interval down to $\\alpha=0$.","pith_inferences":["The same jump-counting logic could identify any finite mixture whose components have strictly ordered support endpoints and positive densities at those endpoints, so the auction setting is one instance of a more general mixture-identification principle.","If an intermediate bidder count has zero probability in the population, the jump count would misreport the support of $N$; a robust alternative would treat detected jumps as a subset of possible boundaries and use the jump-size inequalities to prune impossible configurations.","The USFS finding that a third recorded bid is often non-competitive suggests a testable screen for passive or coordinated bidding in procurement data, since the model itself does not explain why the dominated bid occurs."],"forward_implications":["In markets where only transaction prices are recorded, competition intensity and bidder valuations become estimable without observing the number of bidders or any losing bids.","When active buyers observe the number of competitors, endogenous participation through a reserve price or entry cost does not block identification from winning bids; when they do not observe it, identification requires an instrument or auxiliary observations such as unsold objects.","The derived inequalities on jump sizes and locations provide a direct specification test: winning-bid data violating them cannot be rationalized by the benchmark model.","Empirically, using the recorded number of bids as the competition level can overstate true competition; in the USFS timber data, most three-bid auctions behave as if only two bidders compete, and risk-aversion bounds move toward risk neutrality."],"supporting_citations":[{"why":"Supplies the equilibrium quantile mapping between bid and value distributions used in the iterative identification.","marker":"Guerre, Perrigne and Vuong (2000)"},{"why":"Establishes that symmetric Bayesian Nash equilibrium bids are strictly increasing in private values under compact support.","marker":"Maskin and Riley (1984)"},{"why":"Provides the tail-index estimator used to identify the lowest number of bidders.","marker":"Hill and Shneyerov (2013)"},{"why":"Prior winning-bid identification with a misclassification proxy that the new discontinuity argument extends and contrasts.","marker":"An, Hu and Shum (2010)"},{"why":"Instrument-based identification for auctions with selective entry, relied on for the entry-cost extensions.","marker":"Gentry and Li (2014)"},{"why":"Defines the affiliated-signal two-stage entry model used in the entry-cost analysis.","marker":"Ye (2007)"},{"why":"Supplies the USFS timber data and risk-aversion estimates used as the empirical benchmark.","marker":"Lu and Perrigne (2008)"},{"why":"Shows expected payoff does not depend on whether buyers observe competition, which frames the observability assumption.","marker":"McAfee and McMillan (1987)"}],"fun_headline_variants":["Density jumps reveal hidden bidder count in auctions","Winning bids alone unmask the number of bidders","Unobserved competition identified via bid density jumps","From single bids to bidder numbers: a density trick","Hidden bidders exposed by winning-bid density kinks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Identification collapses if some intermediate number of bidders never occurs, because that number would produce no density jump, the support count would be wrong, and the iterative expansion down the bid distribution could not begin.","fun_headline_variants_meta":{"raw":{"variants":["Density jumps reveal hidden bidder count in auctions","Winning bids alone unmask the number of bidders","Unobserved competition identified via bid density jumps","From single bids to bidder numbers: a density trick","Hidden bidders exposed by winning-bid density kinks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000275,"raw_usage":{"total_tokens":1649,"prompt_tokens":956,"completion_tokens":693,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":614}},"tokens_in":572,"tokens_out":693,"duration_ms":6448,"temperature":1.0,"reasoning_tokens":614,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:13:54.979972+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the benchmark model with $F$ uniform on $[0,1]$, buyers observing $N$, and $N$ taking values 2 and 4 with equal probability. The winning-bid density has jumps exactly at $b_2$ and $b_4$, so the paper's rule $\\bar n = \\underline n + \\#\\{\\text{jumps}\\} - 1$ returns $\\bar n = 3$, and the estimated $p_3$ is positive although $P(N=3)=0$; this violates the true data-generating process and shows that Assumption N's full-support requirement is what carries the jump-counting step.","supporting_citations":[],"review_version":1}