{"id":"d3d3f1c0-a922-42db-9fd4-b845c4ab7917","arxiv_id":"2502.09420","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"XOR package bid selection in electricity auctions can be solved as a linear program under finite package sets, and the welfare loss from bid limits is bounded by a Wasserstein distance between true and scenario price distributions.","lead":"This paper studies how electricity traders should pick the small set of package bids (XOR bids) they are allowed to submit in European power auctions, and how much efficiency is lost when bid counts are capped. It gives a fast linear-programming method for picking bids, a bound linking price forecast error to profit loss, and simulations for a generator, a battery, and a flexible heating utility.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Wasserstein bound only covers non-binding bid caps (B≥S); loss from a binding cap (B<S), the paper's main regime, is unproven.","rationale":"The reader's verdict is CONDITIONAL, and my stress-test does not move it; the concern I identify is related to but distinct from the reader's weakest_assumption. The reader identifies the Walrasian/no-market-power assumption as the weakest point, while also noting in the rationale that 'the theoretical bound does not cover the recommended heuristic algorithm used in the main simulations.' My concern sharpens this: Theorem 1's bound is valid only for strategy (5), which requires B ≥ S. The paper's main numerical regime is B=24 and S up to 400, where the bid cap binds and Algorithm (9) is used; no theorem bounds the loss in that regime. This is a genuine scope gap in the central welfare-loss claim, but it does not invalidate the LP reformulation for finite packages (Proposition 2), which appears sound, nor the Wasserstein bound itself under its explicit assumptions. The paper is transparent about using Algorithm (9) as a heuristic and about the assumptions behind the theorem, so the issue is one of overclaimed applicability rather than internal inconsistency. The CONDITIONAL verdict remains appropriate: the authors should either extend the bound to the B<S regime or soften the claim that bid-limit welfare losses are bounded. My proposed check—constructing a small finite-package instance with B<S and comparing the optimal LP loss to L·d_W(P,Q)—would settle whether the loss can exceed the theorem's bound in the binding regime, thereby determining whether the overclaim is substantive.","tokens_in":20287,"tokens_out":24628,"duration_ms":220149,"concrete_test":"Take the finite-package model of Section 2.4 with K=3 candidate packages, S=2 price scenarios, B=1 (a binding cap), a non-atomic price distribution P, and Q the scenario distribution. Compute the optimal expected loss of the LP (8) with B=1 and compare it with L·d_W(P,Q) for the same P,Q. If the loss can exceed L·d_W(P,Q), Theorem 1 cannot be used to bound bid-limit losses in the B<S regime. Alternatively, run the Section 4.4 battery case with B=24 and S=400 and compare the empirical average loss to 2·max||x||·d_W between the empirical scenario distribution and the realized-price distribution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3's Theorem 1 bounds E[Γ(λ)] ≤ L·d_W(P,Q) for the profit loss Γ(λ) of the simple strategy (5). Strategy (5), defined in Section 2.3, is only available when B ≥ S: the agent submits one scenario-optimal profile per price scenario. In the paper's motivating setting, the bid cap binds: Section 4.4 uses B=24 with S=24,50,100,200,400, so for S>24 strategy (5) is infeasible and the recommended Algorithm (9)—generate S profiles, then select B of them—is used instead. Theorem 1 therefore does not bound the loss caused by the bid limit when that limit is actually restrictive; it bounds only the loss due to approximating the true price measure P by the scenario measure Q under an unlimited ability to submit all S profiles. The abstract's and Section 1's claim of 'an upper bound on the profit loss caused by the limit on the number of bids' is thus not supported in the B<S regime. The empirical results in Section 4 are suggestive, but no theorem connects them to the Wasserstein bound.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies how a price-taking participant in a uniform-price combinatorial electricity auction should select XOR package bids when the auctioneer caps the number of bids per exclusive group. It proposes a stochastic bilevel model, argues that truthful bidding dominates, and claims that with a finite package set the selection problem can be solved as an LP because the constraint matrix is totally unimodular. For infinite package sets it proposes a heuristic (Algorithm 9) that generates scenario-optimal profiles and then selects B of them. The main theoretical result is a Wasserstein-distance bound on the expected profit loss of the simple strategy that submits one optimal profile per scenario. The paper also reports simulations for a thermal generator, a battery, and a flexible load in the German 2023 day-ahead market, and discusses OR bids, convex package bids, and portfolio bidding.","tokens_in":20510,"tokens_out":35391,"duration_ms":320663,"significance":"If the LP reformulation and the Wasserstein bound were fully established, the paper would offer a practically appealing polynomial-time method for package-bid selection and a quantitative link between price-forecast error and efficiency loss. The paper is explicit about its Walrasian-equilibrium assumptions, provides reproducible code, and makes a genuine attempt to connect auction design to real European market rules. However, the two main theoretical pillars are currently not established in the regime that motivates the paper: Theorem 1 covers only the non-binding case S ≤ B, while the simulations and the recommended algorithm operate in the binding case S > B, and the total-unimodularity proof in Proposition 2 is demonstrably incorrect. The contribution is therefore conditional on substantial revision.","major_comments":[{"comment":"The theorem bounds the expected profit loss of the simple strategy (5), which is only feasible when B ≥ S, as stated in Section 2.3. In the paper's main motivating regime, where the bid cap binds, Section 4.4 uses B = 24 with S = 50, 100, 200, and 400, and the recommended algorithm (9) is used instead. Thus Theorem 1 does not bound the loss caused by a binding bid cap; it bounds only the loss from approximating the true price distribution P by the scenario distribution Q when the agent may submit all S scenario-optimal profiles. The abstract's and Section 1's claim of an upper bound on the profit loss caused by the limit on the number of bids is therefore not supported in the B < S regime. The authors should either restrict the claim explicitly to S ≤ B or prove a bound for the actual selection algorithm (8)/(9).","section":"Section 3, Theorem 1"},{"comment":"The proof of Proposition 2 asserts that the constraint matrix formed by (7c)-(7e) is totally unimodular. This assertion is false. For K = 1, S = 2, B = 1, the submatrix using rows (7c) for s = 1, (7d) for (k,s) = (1,1) and (1,2), and columns γ11, γ12, δ1 is [[1,1,0],[1,0,-1],[0,1,-1]], whose determinant is 2. Therefore the claimed equivalence between (8) and its LP relaxation is not established by the total-unimodularity argument, and the paper's central claim of a polynomial-time LP reformulation is unsupported. The authors need to provide a correct integrality proof for the selection polytope or revise the algorithmic and complexity claims. If the LP relaxation is not exact, Algorithm (9) as described may produce fractional package selections, and the simulation results in Section 4 would need to be re-examined.","section":"Section 2.4, Proposition 2"},{"comment":"The simulation results in Figures 4-6 are obtained with Algorithm (9), which Section 2.5 explicitly states does not guarantee an optimal solution to the original infinite-package problem. There is no theorem connecting the empirical profit losses to the Wasserstein bound, and the experiments measure the heuristic's performance on a finite candidate set rather than the economic loss from a binding bid cap. The paper should state this limitation in Section 4 and avoid presenting the numbers as validation of Theorem 1.","section":"Section 4"}],"minor_comments":[{"comment":"The variable 'usj' in equation (17l) appears to be a typo for 'u_j' or 'u_t'; please correct it.","section":"Appendix D, equation (17l)"},{"comment":"Equation (19d) contains the string 'eT s=' which seems to be a formatting error; it should read 'e_T = E_0'.","section":"Appendix F, equation (19d)"},{"comment":"The Wasserstein distance definition has a typo: '∥λ1, λ2∥2' should be '∥λ1 − λ2∥2'.","section":"Section 3.3, equation (12)"},{"comment":"The passage from the per-agent surplus loss to a total welfare bound is informal; please provide the explicit aggregation formula, such as ∑_i L_i · d_W(P, Q_i) when agents have different Lipschitz constants and scenario distributions.","section":"Section 3, welfare aggregation"},{"comment":"Please specify whether the profit percentages in Figures 4-6 are averages over the same 100 sampled days for all curves, and add error bars or a measure of dispersion, since the differences between curves are often small.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The false total-unimodularity claim in Proposition 2 is likely to be the most contentious point in review. The paper's reliance on the author's own preprint (Hübner 2025) for the equilibrium assumption is acceptable but should be flagged as self-citation. If the LP integrality question is resolved positively, the paper could become a solid contribution with revised scope claims; if not, the main algorithmic contribution would need to be repositioned as a heuristic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The LP reformulation of the XOR package-bid selection problem is the real substance here. For a finite package set, the total-unimodularity argument is clean and gives polynomial-time solvability, a genuine improvement over the MILP in Karasavvidis et al. (2024). That alone is worth citing, and the code being public makes it reproducible. The case studies (generator, battery, flexible demand) are sensible, and the discussion of OR bids and convex package bids is thoughtful. The paper is also honest about its Walrasian-equilibrium assumption, acknowledging it in the introduction and conclusion, and provides some empirical context from EUPHEMIA data that makes that assumption less outlandish than it sounds.\n\nBut the stress-test note is correct and it matters. Theorem 1 bounds the loss from approximating the true price distribution P by a discrete scenario measure Q—under strategy (5), which requires B ≥ S. That is not the regime the paper cares about. In the simulations, the cap binds (B=24, S=24–400), and the recommended algorithm (9) is a heuristic with no optimality guarantee. So the abstract's claim that the paper derives 'an upper bound on the profit loss caused by the limit on the number of bids' is not supported for the case where that limit actually binds. The bound only covers the loss caused by using a finite scenario set when you can submit one bid per scenario. That is a different, much easier problem. The simulations are suggestive, but they are averages over 100 days with no variance or confidence intervals, so I cannot judge whether the apparent differences between bid-count settings are real. The theoretical gap also means the simulation results are not backed by the Wasserstein bound, even in spirit.\n\nThere is a smaller issue with Proposition 1: the proof is terse and compares feasible regions without spelling out why the no-truthful-biding program's lower-level constraint actually binds tighter than (4). It is probably fixable with a few lines, but as written it leans on intuition.\n\nNet: this is a useful decision-support paper with one strong, formally grounded contribution and one overstated result. A serious referee should see it, because the LP reformulation and the simulation setup deserve engagement, and the authors are clearly capable of addressing the gap. The revision should either prove a bound for the B<S case or significantly soften the language in the abstract and introduction. I would not desk-reject it, but I would insist on the fix before it is trusted for market design.\n\nRecommendation: send to peer review with major revision requested.","headline":"The LP reformulation is a solid, citable result, but the headline Wasserstein bound does not actually bound the loss from a binding bid cap—so the paper's central claim is overstated.","tokens_in":21042,"tokens_out":2954,"would_cite":true,"duration_ms":31166,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91B26","90C05","90C15","90C27"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that the welfare loss from limiting XOR package bids in day-ahead electricity auctions is bounded by a Wasserstein distance between forecast and true prices, and that optimal bid selection is a linear program.","keywords":["XOR package bids","combinatorial auctions","electricity market design","block bids","stochastic bilevel optimization","total unimodularity","Wasserstein distance","welfare loss"],"falsifier":"On historical German day-ahead data, run the paper's algorithm (9) with the scenario generator of Section 4.2 and compare realized profit with the perfect-information maximum for the generator, battery, and heating utility; if the average shortfall ever exceeds $L \\cdot d_W(P,Q)$, the Lipschitz argument behind Theorem 1 is false. Separately, solve the LP relaxation of program (8) on instances with many candidate packages; any fractional optimum strictly better than the best integral solution would contradict the total-unimodularity claim.","tokens_in":20055,"feed_emoji":"⚡","tokens_out":10797,"duration_ms":100328,"temperature":0.7,"pith_summary":"This paper claims that the welfare loss caused by limiting the number of XOR package bids in day-ahead electricity auctions is both computationally manageable and quantitatively bounded. The authors show that the stochastic bilevel program a bidder must solve to choose a limited set of package bids becomes a linear program when the bidder's candidate packages are finite, because the constraint matrix is totally unimodular. They then derive an upper bound on the expected profit loss: it is no more than a Wasserstein distance between the true price distribution and the scenario distribution used for bid selection, scaled by the size of feasible power profiles. This matters because European day-ahead auctions cap exclusive groups at 24 block bids, and the paper gives bidders a fast way to spend that budget and gives auction designers a way to price the cap's welfare cost.","feed_headline":"Package-bid limits cost little when price forecasts are sharp","feed_subtitle":"Expected profit loss is bounded by a Wasserstein distance, and a linear program finds the best limited bid set.","key_machinery":"The central object is the XOR package bid—called an exclusive group of block bids in European auctions—a collection of priced power profiles of which the auctioneer accepts at most one, or none. Two mechanisms carry the argument. First, the stochastic bilevel bid-selection program is shown to collapse to a single-level binary program under truthful bidding, and its constraint matrix is totally unimodular, so binary variables can be relaxed and the LP optimum is integral. Second, the profit-loss function is Lipschitz-continuous with constant $L = 2 \\max\\{\\|x\\|_2 : v(x) \\neq -\\infty\\}$, and the resulting Wasserstein distance $d_W(P,Q)$ between the true and scenario price distributions converts forecast error directly into an expected-profit-loss bound. The Wasserstein distance is doing the quantitative work: it is the minimal transport cost between $P$ and $Q$, and the theorem says that if scenarios approximate the true distribution well, the worst-case expected loss shrinks proportionally.","core_discovery":"On the paper's own terms, the central discovery is that the XOR bid selection problem is not inherently hard. Assuming a Walrasian equilibrium exists and no bidder has market power, truthful bidding is a dominant strategy even when the bidder may only submit a limited set of packages, so the bilevel program collapses to a single-level problem. For a finite set of candidate packages, the resulting binary program has a totally unimodular constraint matrix, its LP relaxation is exact, and optimal bid selection can be solved in polynomial time. The first welfare theorem then translates per-bidder profit loss into total welfare loss, and Theorem 1 bounds the expected loss by $L \\cdot d_W(P,Q)$, where $d_W(P,Q)$ is the Wasserstein-1 distance between the true price distribution $P$ and the scenario distribution $Q$, and $L$ is twice the largest feasible package norm. Simulations of a thermal generator, a battery, and a district heating utility on 2023 German data show profits rising with the number of bids and with forecast accuracy, approaching the perfect-information benchmark as the Wasserstein distance goes to zero.","pith_inferences":["The paper leaves implicit an operational rule for setting the bid cap: auctioneers could keep increasing the limit until the marginal Wasserstein-based welfare gain falls below a threshold, using the same bound as a monitoring metric.","A natural extension is distributionally robust bid selection, replacing the true price distribution with an ambiguity set and deriving a robust analogue of Theorem 1; the total unimodularity result suggests such an extension would likely stay tractable.","The LP machinery is not electricity-specific and could transfer to any Walrasian-style market with uniform prices and XOR package bids, such as spectrum or transportation auctions, when participants have usable price forecasts.","The empirical evidence covers one bidding zone and one year; repeating the profit-loss curves on intraday auctions with 96 periods or on other zones would test whether the 'more scenarios help only if fresh' pattern generalizes."],"forward_implications":["A bidder with a finite set of candidate packages can find its optimal limited XOR bid list in polynomial time; the paper shows the LP relaxation is exact, and the same formulation covers risk-averse objectives such as CVaR.","The welfare cost of a bid limit is governed by forecast quality rather than by the limit itself: with accurate scenarios, additional bids buy little, while poor forecasts make the cap expensive.","Auctioneers can loosen limits for convex package bids (minimum acceptance ratio 0) and keep tight limits only for nonconvex fill-or-kill bids, preserving tractability while reducing the missing-bids welfare loss.","The 24-bid cap on exclusive groups in European day-ahead auctions is an artifact of hourly flexi orders, not a computational necessity, so the paper's evidence supports raising it."],"supporting_citations":[{"why":"Supplies the strategy-proofness-in-the-large result that justifies truthful bidding as dominant, letting the bilevel selection problem collapse.","marker":"Azevedo and Budish (2019)"},{"why":"Provides the Walrasian equilibrium and first welfare theorem that translate each bidder's profit loss into total welfare loss.","marker":"Mas-Colell et al. (1995)"},{"why":"Formulates the XOR bid selection problem as a stochastic bilevel MILP; this paper's LP is a reformulation and the simulations use it as the baseline to beat.","marker":"Karasavvidis et al. (2024)"},{"why":"Supplies the total unimodularity and integrality results used to prove the LP relaxation of the bid-selection program is exact.","marker":"Nemhauser and Wolsey (1988)"},{"why":"Provides the optimal-transport existence theorem used in the proof that the Wasserstein distance bounds the expected profit loss.","marker":"Villani (2009)"},{"why":"Frames scenario selection as approximation of one probability measure by another, the interpretation that motivates the Wasserstein bound.","marker":"Rujeerapaiboon et al. (2022)"}],"fun_headline_variants":["Polynomial-time bid selection in XOR package auctions","Bid caps barely dent welfare when forecasts are sharp","Wasserstein bound: why bid limits don't cost much","Limited package bids: near-optimal with decent price foresight"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the day-ahead auction is Walrasian: uniform prices clear the market, no bidder can influence prices, and the auctioneer accepts each bidder's most profitable package at those prices; if any of those fail, the LP still solves an optimization problem but it no longer measures profit or welfare.","fun_headline_variants_meta":{"raw":{"variants":["Polynomial-time bid selection in XOR package auctions","Bid caps barely dent welfare when forecasts are sharp","Wasserstein bound: why bid limits don't cost much","Limited package bids: near-optimal with decent price foresight"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00025,"raw_usage":{"total_tokens":1565,"prompt_tokens":969,"completion_tokens":596,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":530}},"tokens_in":585,"tokens_out":596,"duration_ms":6151,"temperature":1.0,"reasoning_tokens":530,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T21:35:00.643430+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On historical German day-ahead data, run the paper's algorithm (9) with the scenario generator of Section 4.2 and compare realized profit with the perfect-information maximum for the generator, battery, and heating utility; if the average shortfall ever exceeds $L \\cdot d_W(P,Q)$, the Lipschitz argument behind Theorem 1 is false. Separately, solve the LP relaxation of program (8) on instances with many candidate packages; any fractional optimum strictly better than the best integral solution would contradict the total-unimodularity claim.","supporting_citations":[],"review_version":1}