{"id":"05c36fb7-9866-42e1-9f40-2144fc371801","arxiv_id":"2606.22981","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Adaptive distance learning via scoring rules for out-of-sample performance improves ABC and localized NPE-PFN on misspecified time series and connects to linear pooling of forecasts.","lead":"This paper extends adaptive distance learning for approximate Bayesian computation to misspecified time series and applies similar ideas to localized neural posterior estimation with prior-data fitted networks. A smart generalist might read it to see how optimizing distances via scoring rules can improve simulation-based forecasting without manual tuning.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.3","headline":"Optimizing distance via predictive scoring rule may decouple forecast gains from posterior quality under misspecification","rationale":"The reader’s weakest_assumption is precisely the load-bearing step required by the abstract’s claim. No stronger internal inconsistency is visible from the given material, and the lack of full-text access already produced an UNVERDICTED verdict; the identified assumption remains the decisive unverified link.","tokens_in":1646,"tokens_out":281,"duration_ms":17222,"concrete_test":"On the simulated misspecified examples, recompute the posterior predictive coverage (or PIT uniformity) for the learned distance versus the fixed-distance baseline; if forecast scores improve while coverage deviates further from nominal, the claimed posterior improvement does not follow from the predictive optimization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that distance adaptation (via out-of-sample scoring-rule optimization) produces better posterior approximations, not merely better point forecasts. In misspecified time-series settings this link is not automatic: a distance can be tuned to improve predictive scores while distorting the implicit weighting or acceptance regions that define the ABC or localized NPE posterior. The abstract states the adaptation “optimizes out-of-sample predictive performance” and then asserts improved “posterior estimation methods,” but supplies no argument that the two objectives coincide when the model is wrong.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript extends adaptive distance learning for ABC to misspecified time series and applies similar ideas to localized NPE-PFN. The distance is adapted by optimizing an out-of-sample scoring rule for predictive performance. A connection is drawn between randomized-distance posteriors and linear pooling for forecast combination. Empirical results on simulated and real examples are reported to show improved forecasting performance for both ABC and NPE-PFN methods.","tokens_in":1734,"tokens_out":354,"duration_ms":23742,"significance":"If the empirical gains are robust and the adapted distance demonstrably improves posterior quality (rather than only point forecasts), the approach could be useful for likelihood-free inference under misspecification in time series. The explicit link to linear pooling is a constructive observation that may aid interpretability.","major_comments":[{"comment":"Abstract: the central claim that adaptive distance learning improves 'posterior estimation methods' rests on the unargued assumption that optimizing the distance for out-of-sample predictive scoring rules also improves the quality of the ABC or localized NPE posterior. Under misspecification this link is not automatic; the manuscript supplies no theorem, diagnostic, or simulation showing that the resulting acceptance regions or implicit weights yield better-calibrated or more accurate posteriors rather than merely better point forecasts.","section":"Abstract"},{"comment":"Abstract: no quantitative results, baseline comparisons, metrics, or implementation details are supplied to support the statement that 'adaptive distance learning improves forecasting performance in simulated and real examples.' Without these, the magnitude and statistical reliability of the claimed gains cannot be assessed.","section":"Abstract"}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive report. We address the two major comments on the abstract below, focusing on clarifying claims and strengthening support for the stated improvements in forecasting performance.","responses":[{"response":"We agree that no theorem or direct diagnostic is provided establishing improved posterior calibration or accuracy from the adapted distances. The manuscript targets improved out-of-sample predictive performance under misspecification via scoring-rule optimization, which is the relevant objective when the model is misspecified. The explicit connection to linear pooling of forecasts reinforces the predictive focus. We will revise the abstract to remove any phrasing suggesting direct gains in posterior quality and instead state the improvements in forecasting performance.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the central claim that adaptive distance learning improves 'posterior estimation methods' rests on the unargued assumption that optimizing the distance for out-of-sample predictive scoring rules also improves the quality of the ABC or localized NPE posterior. Under misspecification this link is not automatic; the manuscript supplies no theorem, diagnostic, or simulation showing that the resulting acceptance regions or implicit weights yield better-calibrated or more accurate posteriors rather than merely better point forecasts."},{"response":"The abstract is a high-level summary; the full quantitative results (including scoring-rule values, baseline comparisons to standard ABC and NPE-PFN, and implementation details) appear in Sections 4–5 with accompanying figures and tables. To address the concern directly in the abstract, we will add a concise statement summarizing the observed average improvements in predictive scores across the simulated and real examples.","revision_made":"partial","referee_comment":"[Abstract] Abstract: no quantitative results, baseline comparisons, metrics, or implementation details are supplied to support the statement that 'adaptive distance learning improves forecasting performance in simulated and real examples.' Without these, the magnitude and statistical reliability of the claimed gains cannot be assessed."}],"tokens_in":1281,"tokens_out":416,"duration_ms":29233,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The core contribution is taking existing adaptive distance methods for ABC and applying them to misspecified time series, while adding a reinterpretation of linear pooling weights as randomized distance learning. They also test the same idea inside localized NPE-PFN. That is a clean, incremental step within the simulation-based inference literature.\n\nWhat works is the practical framing: they optimize the distance directly on out-of-sample scoring rules and then show forecasting gains on both simulated and real data. The pooling connection is a useful observation that ties two separate literatures together without forcing a new framework.\n\nThe soft spot is the missing link between forecast improvement and posterior quality. Under misspecification, a distance tuned for predictive scores can shift acceptance regions or weighting in ways that help point forecasts while leaving the approximate posterior distorted. The abstract asserts better posterior estimation methods but gives no argument or diagnostic that the two objectives stay aligned when the model is wrong. Without quantitative results, baseline comparisons, or checks on posterior calibration in the provided text, it is hard to judge how large or reliable the gains are.\n\nThis is for researchers already working on likelihood-free methods for time series, especially in applied fields that rely on simulation. A reader looking for a modest practical tweak would find it useful; someone expecting a deeper fix for misspecification would not.\n\nI would send it to peer review. The extension is legitimate and the empirical direction is reasonable, but referees need to see whether the posterior approximations actually improve or whether the method is mainly re-tuning for prediction.","headline":"The paper extends adaptive distance learning to misspecified time series ABC and NPE-PFN, with a link to linear pooling, but the claimed posterior gains rest on an unexamined assumption that predictive-score tuning improves inference rather than just forecasts.","tokens_in":2234,"tokens_out":399,"would_cite":false,"duration_ms":13379,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Adaptive distance learning via scoring-rule optimization improves forecasting performance in ABC and localized neural posterior estimation for misspecified time series.","keywords":["approximate Bayesian computation","neural posterior estimation","adaptive distance learning","scoring rules","likelihood-free inference","model misspecification","time series forecasting","forecast combination"],"falsifier":"An experiment on the paper's simulated or real time-series examples in which the learned adaptive distance yields equal or worse forecast scores than a fixed, non-adapted distance.","tokens_in":2529,"feed_emoji":"📊","tokens_out":643,"duration_ms":18282,"temperature":0.7,"pith_summary":"The paper extends adaptive distance learning, previously used in ABC, to handle misspecified time series by tuning the distance to maximize out-of-sample predictive performance with a scoring rule. The same adaptation is explored inside localized neural posterior estimation that uses prior-data fitted networks. A connection is established showing that empirical weights from linear pooling of forecasts can be viewed as another instance of randomized-distance adaptation. Experiments on simulated and real data demonstrate gains in forecasting accuracy for both families of methods when the distance is learned this way rather than fixed in advance.","feed_headline":"Adaptive distances boost forecasts in ABC and NPE","feed_subtitle":"Tuning distances to maximize out-of-sample scores improves posterior estimates and predictions for misspecified time series.","key_machinery":"Adaptive distance function whose parameters are chosen to maximize a scoring rule on out-of-sample predictive performance, then used inside rejection or weighting steps for ABC and inside localized neural posterior estimation.","core_discovery":"For both ABC algorithms and NPE-PFN methods with localization, adaptive distance learning improves forecasting performance in simulated and real examples. The adaptation optimizes out-of-sample predictive performance using a scoring rule, and empirical estimation of linear-pooling weights for forecast combination can be interpreted as another form of adaptive distance learning through randomized distances.","pith_inferences":["Distance choice in likelihood-free methods can be reframed as a tunable hyperparameter driven by predictive scoring rather than fixed by domain knowledge alone.","The same optimization principle may extend to other likelihood-free algorithms that rely on a distance or discrepancy measure.","Under stronger misspecification, the learned distance might reveal which data features remain informative for prediction even when the full model is wrong."],"forward_implications":["ABC particle approximations become more accurate for forecasting when the distance is tuned on held-out predictive performance.","Localized NPE-PFN posterior estimates likewise show improved forecast accuracy after the same distance adaptation.","Linear pooling weights estimated from data can be re-interpreted as the result of an adaptive-distance procedure that randomizes the distance.","The approach applies directly to both simulated and real misspecified time-series settings without requiring the model to be correctly specified."],"fun_headline_variants":["Adaptive distance learning for ABC and NPE-PFN forecasts","Optimizing distances via out-of-sample scores in ABC and NPE","Linear pooling as adaptive distance learning in posterior estimation","Adaptive distances aid ABC and localized neural posterior estimation"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"That choosing distances to maximize out-of-sample predictive scores will automatically produce better posterior approximations and forecasts when the underlying model is misspecified.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive distance learning for ABC and NPE-PFN forecasts","Optimizing distances via out-of-sample scores in ABC and NPE","Linear pooling as adaptive distance learning in posterior estimation","Adaptive distances aid ABC and localized neural posterior estimation"]},"model":"grok-4.3","cost_usd":0.005081,"raw_usage":{"total_tokens":2435,"prompt_tokens":590,"num_sources_used":0,"completion_tokens":63,"cost_in_usd_ticks":50812000,"prompt_tokens_details":{"text_tokens":590,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1782,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":590,"tokens_out":63,"duration_ms":13548,"temperature":1.0,"reasoning_tokens":1782,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T06:21:30.422839+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"An experiment on the paper's simulated or real time-series examples in which the learned adaptive distance yields equal or worse forecast scores than a fixed, non-adapted distance.","supporting_citations":[],"review_version":1}