{"id":"620c2606-ac57-4e23-a0aa-bd2a466bd8e3","arxiv_id":"2602.21869","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A Bayesian posterior-predictive framework with an empirically fitted prior predicts future interbank network links from past topology and current node strengths, and can run self-sustained without further observed links.","lead":"This paper builds a Bayesian method that learns from past network snapshots and then predicts future links between banks using only each bank's total activity. It is a way to reconstruct confidential interbank networks week by week, and the authors show the predictions can feed themselves forward without seeing new links.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The self-sustained claim rests on Eq. (39), which substitutes the continuous probability matrix Q_t for binary A_t in the likelihood without a derivable probabilistic model; this makes the recursive 'Bayesian' update a heuristic, not a posterior update.","rationale":"The reader's weakest assumption identifies exactly the load-bearing step: Eq. (39)'s substitution of Q_t for A_t in the likelihood. I agree that this is the central weakness. The one-step-ahead posterior predictive framework in Eqs. (7)–(9) is well-defined conditional on the observed A_t, and the empirical comparison against the in-sample dcGM is a legitimate control. However, the self-sustained extension, which is the paper's most striking claim, depends on treating a continuous probability matrix as binary data. The paper provides no derivation of this plug-in as a limit or approximation of a proper Bayesian update, and the empirical stability shown in Figs. 3 and F.1 does not resolve the absence of a likelihood for Q_t. The concern does not invalidate the one-step-ahead reconstruction, so the appropriate verdict remains CONDITIONAL rather than REJECT. The proposed test would quantify whether the plug-in is a harmless surrogate or a substantive departure from Bayesian inference.","tokens_in":22794,"tokens_out":11243,"duration_ms":118969,"concrete_test":"Using the released OR4CLE package, recompute the self-sustained R matrices for 2002–2012 with a proper Bayesian continuation: instead of Eq. (39)'s plug-in likelihood, propagate the full posterior over z given A_t (i.e., use the posterior predictive distribution from the previous step as the updated belief, rather than point estimates plus Q_t-as-data). Compare the resulting R matrices to those from Eq. (39). If the mean absolute difference in r_ij exceeds 0.01 or the TPR/JI series shift by more than 0.02, the plug-in is not a faithful Bayesian update and the headline self-sustained claim should be downgraded to a heuristic.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claim is that each predicted snapshot serves as a reliable prior for the next, enabling self-sustained reconstruction. That claim rests entirely on Eq. (39) and Appendix F, Eq. (E1), where the observed adjacency matrix A_t is replaced by the estimated probability matrix Q_t in the likelihood. The resulting object, P(Q_t|z) = ∏_{i<j} (z s_i s_j)^{q_ij} / (1+z s_i s_j), treats each q_ij—a real number in [0,1]—as if it were a binary adjacency entry a_ij. No sampling distribution for the continuous matrix Q_t is specified, and the fractional exponents do not correspond to a likelihood for q_ij under any stated model. The posterior P(z|Q_t) ∝ P(Q_t|z)π(z) is therefore not a coherent Bayesian update, and the recursion 'Q_t is a reliable prior for Q_{t+1}' has no formal justification. The empirical agreement in Figs. 3 and F.1 shows that the plug-in is stable on aggregate metrics, but it does not establish calibration, nor does it show that the approximation is a valid posterior continuation. This is a load-bearing weakness because the self-sustained regime, not the one-step-ahead predictor, is the headline contribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Bayesian framework for out-of-sample reconstruction of evolving binary networks. It instantiates the framework with two models: the Bayesian Erdős–Rényi Model (BERM), with a Beta prior, and the Bayesian Fitness Model (BFM), a single-parameter fitness model with an empirically calibrated Gamma prior. The authors derive posterior predictive link probabilities for time t+1 conditional on the snapshot at time t, and test the procedure on weekly eMID interbank snapshots over 1999–2012. In addition to one-step-ahead forecasts, they propose a \"self-sustained\" recursion in which the estimated probability matrix Q_t replaces the observed adjacency matrix A_t in the likelihood (Eq. 39), and they claim that each predicted snapshot then serves as a reliable prior for the next. The BFM recovers per-node degrees with TPR ≈ 0.40 and JI ≈ 0.20, and the self-sustained version is reported to closely match an in-sample dcGM that observes the true link count.","tokens_in":23137,"tokens_out":8888,"duration_ms":90945,"significance":"If the self-sustained recursion were given a principled Bayesian derivation, the paper would offer a novel and practically relevant temporal extension of entropy-based network reconstruction. The one-step-ahead posterior predictive construction is clean, the rolling-window empirical prior is carefully designed, the comparison against an in-sample dcGM is a sensible control, and the open-source OR4CLE implementation is a concrete asset. However, the headline claim about self-sustained reconstruction currently rests on a plug-in substitution in Eq. (39) that is not derived from a probability model for Q_t, so the central methodological advance is not yet rigorously established.","major_comments":[{"comment":"The self-sustained update replaces the observed binary adjacency matrix A_t with the estimated probability matrix Q_t inside the Bernoulli likelihood, writing P(Q_t|z)=\\prod (z s_i s_j)^{q_ij}/(1+ z s_i s_j). This expression is a valid likelihood only when the exponents are 0/1; for q_ij in [0,1] it is not a probability mass or density over the continuous matrix Q_t, and no sampling model for Q_t is specified. The resulting object P(z|Q_t) ∝ P(Q_t|z)π(z) is therefore a pseudo-posterior, and the recursion \"Q_t is a reliable prior for Q_{t+1}\" has no formal justification. The empirical stability in Figs. 3 and F.1 is suggestive but does not supply the missing derivation. This is load-bearing because the abstract's central claim about self-sustained reconstruction rests on Eq. (39). Please either derive the update from a proper probabilistic model (e.g., via expected sufficient statistics,","section":"Self-sustained inference, Eq. (39); Appendix F, Eq. (E1)"},{"comment":"The recursion is not self-contained in the sense of requiring only the initial adjacency matrix: Eq. (39) uses the future strengths s_i^{t+1}, s_j^{t+1}, which in the eMID application are computed from the future weekly weighted matrix (Appendix E). Thus each step receives new node-level information beyond the initial topology. This can be a legitimate reconstruction scenario if future aggregate strengths are assumed available, but the paper should state this boundary condition explicitly. As written, the phrases \"self-sustained\" and \"minimal amount of additional data\" overstate the degree to which the algorithm propagates with no new input.","section":"Self-sustained inference, Eq. (39); Appendix E"},{"comment":"The text says the Bayesian predictor \"frequently matches, and sometimes exceeds\" the in-sample dcGM, which is an appropriate characterization. However, the abstract's phrase \"outperforming probabilistic benchmarks\" is considerably stronger: the reported differences in TPR, PPV, TNR, and ACC are at the fourth decimal (0.2498 vs 0.2494, etc.), and the degree-level errors ARE_k and MRE_k are slightly in favor of the dcGM. In addition, the only external link-prediction benchmark cited is the DBCM result in [39]. The wording should be tempered or supplemented with error bars / significance tests and additional standard baselines.","section":"Section \"Out-of-sample versus in-sample reconstruction\", Fig. 5"}],"minor_comments":[{"comment":"The Kullback-Leibler divergences D_KL(A||Q) and D_KL(A||R) are defined with terms a_ij ln(a_ij/q_ij). This is fine when a_ij=0, but if any q_ij or r_ij is 0 while a_ij=1 the expression is infinite. It would be useful to state the convention used for such cases, even if none occur in practice.","section":"Eqs. (40)–(41)"},{"comment":"Typographical errors: \"toal\" should be \"total\" and \"constituing\" should be \"constituting\". Similar typos appear in Fig. B.1 captions.","section":"Appendix C, Fig. C.3 caption"},{"comment":"The text says the prior is updated by adding each new weekly value, but also says the prior rolls over a three-year window. Please clarify whether the window strictly drops observations older than three years or whether the prior distribution is refit each week on the trailing window. The current wording is ambiguous.","section":"Methods, Prior calibration"},{"comment":"The slice-sampling implementation uses M=3000 post-burn-in draws with no reported convergence diagnostics (e.g., effective sample size or multiple chains). A brief note on Monte Carlo error would strengthen the numerical results.","section":"Appendix D, Slice sampling"}],"recommendation":"major_revision","confidential_remarks":"The conditional reader's report and the stress-test note identify the same core weakness: Eq. (39) is a plug-in of Q_t into a Bernoulli likelihood without a probabilistic model, and the self-sustained claim depends on it. I agree with that assessment. The paper's one-step-ahead machinery is sound and publishable in principle, but the headline contribution needs either a formal derivation or a clear reframing as an approximate/heuristic procedure, plus an explicit statement about the use of future strengths. Major revision seems appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful core of this paper is the one-step-ahead posterior predictive scheme for a single-parameter fitness model. The derivation in Eqs. 7-9 and 25 is clean, the numerical integration (Gauss-Hermite and slice sampling) is handled carefully, and the eMID experiments are detailed and honestly reported. The comparison against an in-sample dcGM that observes the true link count is a good control, and the authors deserve credit for releasing the OR4CLE package. The conceptual novelty - carrying a prior across time in entropy-based reconstruction - is legitimate and will be useful to people working on financial network reconstruction.\n\nThe soft spot is exactly where the reader and the stress-test put it. The 'self-sustained' recursion, which is the abstract's headline claim, rests on Eq. (39)/Appendix F where the observed binary adjacency matrix A_t is replaced by the continuous probability matrix Q_t inside the likelihood. No sampling distribution for Q_t is specified, and fractional exponents (z s_i s_j)^{q_ij} do not correspond to a likelihood for a real-valued matrix. So the recursion is an ad hoc plug-in, not a Bayesian update. The empirical stability shown in Figs. 3 and F.1 is evidence that the heuristic is useful, but it is not evidence that the formal Bayesian story extends to the self-sustained regime. The paper should either derive the approximation (e.g., as a variational or empirical-Bayes step) or explicitly reframe the self-sustained regime as a heuristic and validate it as such.\n\nTwo minor points. First, the paper claims uncertainty quantification as a feature but never checks calibration of the predictive distribution - a coverage or proper-scoring assessment would be easy and would strengthen the paper. Second, the performance is modest (TPR around 0.4, JI around 0.2), and the abstract's 'accurately recovers' is overstated relative to those numbers, even if the in-sample comparison is favorable.\n\nThe citation pattern looks fine; the self-citations are used as external benchmarks, not inside the derivation. The prior calibration is genuinely out-of-sample with respect to the target snapshots, and the in-sample dcGM control rules out the most obvious circularity concern.\n\nI'd send this to peer review. The one-step-ahead method is solid, the application is relevant, and the self-sustained issue is addressable, but it needs to be confronted directly before publication. The right verdict for the current version is major revision, with the self-sustained claim either derived or demoted.","headline":"The one-step-ahead Bayesian reconstruction is clean and worth publishing, but the 'self-sustained' headline claim is a heuristic plug-in (Eq. 39/F.1) rather than a derived posterior update, so the paper needs revision before the self-sustained story can be taken at face value.","tokens_in":23605,"tokens_out":1870,"would_cite":true,"duration_ms":20206,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A Bayesian update using a single fitness parameter and a rolling empirical prior can reconstruct next-week network structure from past snapshots alone, matching an in-sample model that observes the true link count.","keywords":["network reconstruction","link prediction","Bayesian inference","posterior predictive distribution","exponential random graphs","fitness model","interbank networks","out-of-sample prediction"],"falsifier":"Take a network with a known structural break—e.g., a regulatory shock or a crisis—and run the self-sustained routine with no post-break adjacency data. If the relative error on the next snapshot's link count or degree sequence is no better than simply re-estimating link density from the last observed snapshot, the claim that predicted snapshots are reliable priors is falsified.","tokens_in":22618,"feed_emoji":"🏦","tokens_out":6164,"duration_ms":54518,"temperature":0.7,"pith_summary":"This paper tries to establish that out-of-sample network reconstruction can be carried out by carrying Bayesian priors across time rather than refitting a model to each snapshot. The proposed posterior predictive scheme, instantiated with a single-parameter fitness model, converts the observed network at time t into a posterior over the model parameter and then averages the link probabilities at t+1 over that posterior. Tested on eleven years of weekly eMID interbank snapshots, the method recovers the total number of links with relative error around 0.1 and captures per-node degree heterogeneity far better than a homogeneous Bernoulli version. Its stronger claim is that the procedure is self-sustaining: once seeded with one adjacency matrix, each predicted snapshot serves as the prior for the next, and the resulting forecasts match an in-sample model that is given the true link count. This matters for any setting where only aggregate node-level data are reliably available but future network structure is needed for risk or contagion analysis.","feed_headline":"Out-of-sample network reconstruction works from past snapshots alone","feed_subtitle":"A single fitness parameter plus a rolling prior matches an in-sample model that sees the true link count.","key_machinery":"The load-bearing object is the posterior predictive distribution, instantiated through the density-corrected Gravity Model's one-parameter link probability p_ij(z) = z s_i s_j / (1 + z s_i s_j), with node strengths s_i as exogenous fitnesses. An empirical prior π(z), fit as a Gamma distribution to point estimates of z from a rolling three-year calibration window, is updated with each new snapshot; Gauss-Hermite quadrature (with slice sampling as a check) evaluates the integral. For the self-sustained regime, the estimated probability matrix Q_t replaces the observed adjacency matrix A_t inside the likelihood, a plug-in device that lets the algorithm propagate without new topological data.","core_discovery":"The central claim is that a posterior predictive distribution P(A_{t+1}|A_t) = ∫ P(A_{t+1}|z) P(A_t|z) π(z) dz / P(A_t), with a single fitness parameter z and link probability p_ij = z s_i s_j / (1 + z s_i s_j), carries structural information forward in time. Fitting a Gamma distribution to past point estimates of z provides an empirical prior; the posterior is updated on a rolling three-year window; and the resulting q_ij are expected link probabilities for the next week. On the eMID data, the BFM recovers per-node degrees at the next weekly snapshot with TPR ≈ 0.40 and JI ≈ 0.20, and in self-sustained mode it ties the in-sample density-corrected Gravity Model that observes the true link co","pith_inferences":["The paper only tests the binary, undirected, weekly aggregation of one interbank market; an immediate untested extension is whether the same posterior-predictive recipe transfers to directed networks or to coarser monthly snapshots.","In the sparse regime the paper shows the link ranking reduces to products of node strengths (preferential attachment), independent of the prior; that suggests the ranking part of the method may transfer to other networks where only strengths are known, even when the prior is misspecified.","The paper's reported error bump around 2008 suggests a structural break stresses the prior; one could stress-test the self-sustained loop by deliberately withholding post-break data and measuring how quickly predictive accuracy degrades.","If the plug-in replacement of A_t by Q_t were replaced by a proper latent-variable treatment, predictive uncertainty would likely widen; the reported point estimates in self-sustained mode may understate the true uncertainty."],"forward_implications":["With only node strengths and past snapshots, the BFM produces next-week link probabilities, expected total link counts, and per-node degrees; on eMID, it recovers the degree sequence far more accurately than the homogeneous BERM.","A rolling empirical prior removes the need to recalibrate the model from scratch at each time step; the prior is simply enriched with each new weekly estimate.","After initialization with one observed adjacency matrix, the self-sustained mode continues predicting snapshots from its own estimates alone, matching the in-sample dcGM that is handed the true link count.","The framework yields uncertainty measures through the posterior predictive distribution—e.g., a beta-binomial predictive for the total number of links and expected confusion-matrix entries.","Ranking diagnostics (TPR about 0.40, JI about 0.20, AUROC above chance) show that the BFM induces a meaningful ordering of candidate links, something a homogeneous model cannot do."],"fun_headline_variants":["Bayesian prior from past snapshots forecasts future link probabilities","Self-sustained network reconstruction from a single-parameter fitness model","Past network snapshots drive accurate next-step link prediction","One rolling prior predicts network evolution without retraining","History-based Bayesian model outperforms link prediction benchmarks"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The self-sustained step treats the matrix of estimated link probabilities Q_t as though it were the observed binary adjacency matrix inside the likelihood; if this plug-in approximation is not a valid posterior update, the claim that the procedure can keep predicting from its own estimates lacks a rigorous grounding.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian prior from past snapshots forecasts future link probabilities","Self-sustained network reconstruction from a single-parameter fitness model","Past network snapshots drive accurate next-step link prediction","One rolling prior predicts network evolution without retraining","History-based Bayesian model outperforms link prediction benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00035,"raw_usage":{"total_tokens":1733,"prompt_tokens":718,"completion_tokens":1015,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":462,"completion_tokens_details":{"reasoning_tokens":937}},"tokens_in":462,"tokens_out":1015,"duration_ms":8920,"temperature":1.0,"reasoning_tokens":937,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T20:52:21.478335+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a network with a known structural break—e.g., a regulatory shock or a crisis—and run the self-sustained routine with no post-break adjacency data. If the relative error on the next snapshot's link count or degree sequence is no better than simply re-estimating link density from the last observed snapshot, the claim that predicted snapshots are reliable priors is falsified.","supporting_citations":[],"review_version":1}