{"id":"a35a3649-944f-4063-b633-73ff963183d8","arxiv_id":"2606.26774","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"e2eTD forecasts a small set of aggregate series and disaggregates them via copula-based historical proportions, producing coherent probabilistic forecasts for huge retail hierarchies in minutes.","lead":"This paper introduces e2eTD, a forecasting method that predicts only a few smooth aggregate sales series and then randomly splits those forecasts down to individual products using historical sales patterns. On two large retail datasets it reports better probabilistic accuracy than the compared methods while running in minutes on a laptop.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Top-down split relies on pooled stationary joint distributions; weekday/covariate shifts could break bottom-level calibration.","rationale":"The reader's weakest assumption is the stationarity of the historical joint distributions used for disaggregation. I agree that this is the most load-bearing concern in the paper's argument. The central claim is empirical—lowest WSPL on M5 and Favorita and a simulated M5 rank of 11th—but the method's mechanism for generating bottom-level and intermediate-level distributions is exactly this historical joint distribution. If the true disaggregation proportions shift over the forecast horizon (due to trend, promotions, assortment changes, or even deterministic weekday seasonality), the conditional draws from p_{A,B} are mis-specified, and the resulting bottom samples, and all aggregates above them, are miscalibrated regardless of how good the upper-level forecasts are. The paper's own future-work section acknowledges this limitation, and the ablation study does not address it because all variants retain the pooled, time-invariant estimation of the split distributions. The empirical success on the two test splits is encouraging, but it does not establish robustness to this assumption; a single split test cannot distinguish between a method that genuinely tracks the split distribution and one that happens to do well on two particular 28-day windows. I considered other concerns—rounding/clipping distortion, lack of confidence intervals, recursive copula approximation—but these are either smaller in impact, addressable as implementation details, or do not attack the core mechanism as directly. The concrete test proposed (weekday-conditional estimation) would directly probe the stationarity assumption and settle whether the concern is real on these datasets. Since the reader already reached CONDITIONAL, my analysis does not change the verdict; it sharpens the reason why conditionality is warranted.","tokens_in":20694,"tokens_out":13559,"duration_ms":156433,"concrete_test":"On M5, for every internal node used in the top-down splits, fit the marginal/copula joint distributions separately for each weekday (7 versions per node) instead of one pooled version; keep all other e2eTD steps unchanged. Re-run the forecast and compare bottom-level WSPL (L12) and mean WSPL. Also run a heterogeneity test (e.g., chi-square on the empirical split proportions across weekdays for a sample of nodes). If the weekday-conditional variant improves mean WSPL by more than ~1% (or the heterogeneity test is significant), the pooled stationarity assumption is measurably violated and the method's bottom-level accuracy is not robust; if not, this concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"e2eTD's central mechanism is the probabilistic top-down algorithm (Sect. 4.1), which conditions each disaggregation on a joint distribution p_{A,B} estimated once in-sample (Sect. 4.2). The paper explicitly assumes approximate stationarity: 'We model each pair of series as approximately stationary, summarizing it by a single joint distribution estimated in-sample.' The half-life weighting (28 days) only handles slow drift; it does not condition on weekday, holiday, promotion, or SNAP state. Yet the upper-level ETS forecasts do condition on these covariates (Sect. 5, 'Forecasting model'). Retail disaggregation proportions are likely to vary with these factors (e.g., weekend vs weekday product mix, promotion-driven shifts). If the conditional split probabilities w_j are wrong for a given day, every bottom sample is drawn from the wrong conditional distribution, and all levels below the lowest reconciled level (L10–L12 on M5; L10–L16 on Favorita) inherit the miscalibration even if the reconciled upper totals are perfect. This directly threatens the claim of 'lowest WSPL across aggregation levels.' The paper's ablations do not test this: the 'indep. copula' and 'no reconc.' variants keep the same pooled marginals, so they cannot reveal seasonality/covariate misspecification. The paper itself lists 'accounting for trend, seasonality, or time-varying parameters' as future work (Sect. 6), acknowledging the gap. This is a robustness/generalizability concern rather than a demonstrated error on the two test splits, but it is the least secure link in the central argument.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes e2eTD, a scalable method for coherent probabilistic forecasting of large hierarchical/grouped retail time series. Instead of forecasting all bottom-level series, e2eTD fits univariate ETS models (with exogenous regressors) to a small selected subhierarchy of smooth upper series, reconciles the resulting Gaussian predictive distributions, and propagates integer samples down the hierarchy through a recursive probabilistic top-down algorithm. The split probabilities are obtained from in-sample joint distributions of aggregated sibling series, modeled with Poisson/NB marginals and a Plackett copula. Final coherent forecasts for every level are obtained by bottom-up summation of joint bottom-level samples. On M5 and Favorita, the method reports the lowest mean WSPL among five scalable baselines, a hypothetical 11th place in the M5 Uncertainty competition, and runtimes under 20 minutes on a standard laptop.","tokens_in":20970,"tokens_out":6251,"duration_ms":70994,"significance":"If robust, this is a meaningful practical contribution: e2eTD avoids forecasting intermittent bottom series, is simple and fast, and produces coherent sample paths across very large hierarchies. The manuscript is careful in specifying the algorithm, reports an ablation study, a runtime breakdown, and level-wise evaluation, and compares against several credible baselines. Its principal strengths are transparency and scalability rather than statistical novelty; the top-down split mechanism is new, but its core components (Gaussian reconciliation, count conditioning) build directly on the authors' prior work. The main caveats are the strong stationarity assumption on split distributions and the lack of uncertainty quantification in the empirical comparisons, both of which affect the central 'lowest WSPL across aggregation levels' claim.","major_comments":[{"comment":"The disaggregation engine conditions every bottom sample on a single joint distribution per binary split, estimated once in-sample with a 28-day decay half-life (§4.2). The upper-level ETS forecasts condition on weekday/holiday/SNAP/promotion regressors, but the split distributions do not. Retail disaggregation proportions are likely to vary with weekday and promotions; if they do, bottom-level samples (L10–L12 M5; L10–L16 Favorita) are drawn conditionally on the wrong distribution, and all levels below the lowest reconciled upper series inherit the error even when the reconciled totals are correct. The ablations in §5.4 change copula, reconciliation, subhierarchy, and regressors, but never condition the split on covariates, so they cannot detect this misspecification. Section 6 acknowledges time-varying proportions only as future work. I would like a diagnostic (e.g., holdout split-prop","section":"§4.2, §5.4, Table 7"},{"comment":"All WSPL comparisons are on a single 28-day test window with no interval estimate. On Favorita the differences between e2eTD and S-empD are small at several levels (e.g., 0.112 vs 0.114 at L1; 0.225 vs 0.241 at L7) and the mean difference is 0.013. Given that WSPL is an average over tens of thousands of series, a paired bootstrap over series or a multi-window evaluation would establish whether the reported ordering is stable. The hypothetical M5 rank of 11 is likewise a point estimate from one window. This is the main empirical evidence for the paper's central claim, so it needs an uncertainty assessment.","section":"§5.1, Tables 2–4"},{"comment":"Step III draws samples from a Gaussian reconciled distribution and then 'round[s] and clip[s] to zero' the lowest-upper samples before passing them to the top-down sampler. Rounding and clipping alter the sample totals, so the final coherent forecasts are not exactly the reconciled upper forecasts; the magnitude and effect of this distortion is not reported. Since many lowest-upper totals are small counts, the clipping/rounding can be a substantial fraction of the total. Please quantify the distortion (e.g., mean absolute change before/after) and justify that it does not bias the WSPL comparisons, or use a count-consistent reconciliation.","section":"§4, step III"}],"minor_comments":[{"comment":"The number of Monte Carlo samples N per horizon is never stated; this matters for both reproducibility and quantile stability.","section":"§5"},{"comment":"Typo: 'they they produced runtime errors' should be 'they produced runtime errors'.","section":"§5, competing methods"},{"comment":"HINT uses a GPU while all other methods are CPU-only; the paper notes this, but the headline runtime comparison still mixes hardware. Consider reporting a CPU-only time or making the hardware distinction more prominent in the main text.","section":"§5.3, Table 5"},{"comment":"The simulated benchmark for moment matching vs ML says 'near-identical estimates' but no details or table are provided; please report the setup and results.","section":"§4.2"},{"comment":"The Favorita preprocessing (removing classes/items with non-integer sales, clamping returns) is sensible for a count-based method, but it transforms the dataset; the text calls Favorita 'the largest publicly available retail dataset' after this filtration, which should be stated more carefully.","section":"§5, Favorita preprocessing"}],"recommendation":"major_revision","confidential_remarks":"The core proposal is sound and the computational upside is real. The two risks I would ask the authors to address before publication are the stationarity of the split distributions and the absence of uncertainty quantification in the empirical claims. I would also check whether the Favorita preprocessing changes the ranking. The self-citation pattern is acceptable, but the distinct contribution relative to Zambon et al. 2024c could be sharper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: e2eTD is a genuine new combination — forecast a small subhierarchy, reconcile, then recursively disaggregate via copula-conditioned sampling — and the empirical case is stronger than I expected. The M5 rank-11 claim is credible against the published competition standings, and the runtime numbers (about 5 min on M5, 20 on Favorita on a laptop) are the kind of concrete claim practitioners will actually care about.\n\nWhat's new: the probabilistic top-down step. Instead of fixed historical proportions, each split conditions on a bivariate Plackett copula with Poisson/NB marginals fit in-sample with recency decay. The idea of forecasting only ~0.3% of the hierarchy and then reconstructing bottom samples that sum correctly is clean. The ablation study is honest: it shows the copula matters on Favorita but not much on M5, reconciliation helps consistently, and regressors help where they exist. That transparency earns credit.\n\nSoft spots, in order of seriousness:\n\n1. The stationary split-proportion assumption is the load-bearing simplification. The joint distribution for each pair is pooled over the training window (with a 28-day half-life), while the upper ETS forecasts condition on weekday seasonality, promotions, holidays, and payday effects. If the true disaggregation mix shifts by weekday or promotion state, every bottom sample is drawn from the wrong conditional law and the error propagates up. The authors list \"accounting for trend, seasonality, or time-varying parameters\" as future work, so they know. The two test splits don't settle it — Favorita is short, M5 is long but the method wins anyway. This is a generalizability concern, not a demonstrated error, but it's the least secure link in the central argument.\n\n2. The empirical comparison lacks uncertainty quantification. Tables 2–4 report a single split of WSPL; no CIs, no multiple evaluation windows. The differences on Favorita to S-empD are modest (mean 0.183 vs 0.196), and without error bars I'd want replication before quoting those numbers. The runtime table does report mean±sd over 7 runs, which is good.\n\n3. The rounding/clipping of reconciled Gaussian samples to integer totals is unanalyzed. It's probably minor — totals are large enough that the distortion is small — but a sensitivity check would close it.\n\n4. No code release yet. The method is described in enough detail to reimplement, but I'd want the R package before signing off on the exact WSPL figures.\n\nThe math is sound as far as I can tell. The self-citations are to actual components (Gaussian reconciliation, conditioning), and they're clearly flagged. I don't see circularity.\n\nWho this is for: anyone working on hierarchical or retail forecasting, especially with an eye to scale. It deserves a serious referee — the editor should send it out, not desk reject. I'd push the authors to release code, add uncertainty quantification, and at least run a weekday-conditional split-proportion variant as a robustness check. Conditional accept is the right call.\n\nRecommendation: yes, send to peer review; expect heavy-but-addressable revision.","headline":"A solid, well-specified methods paper with a compelling accuracy/compute trade-off; the main soft spot is the stationary split-proportion assumption, which is acknowledged but not stress-tested.","tokens_in":21593,"tokens_out":2292,"would_cite":true,"duration_ms":24000,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M20","62H05"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that coherent probabilistic forecasts for hierarchies of hundreds of thousands of retail series can be produced by forecasting only a small set of smooth aggregate series and then probabilistically splitting those forecast","keywords":["hierarchical forecasting","probabilistic forecasting","top-down sampling","coherent forecasts","intermittent demand","copula","retail demand","scalable forecasting"],"falsifier":"On a simulated hierarchy with two bottom series whose in-sample split proportion is 50/50 but whose true proportion shifts to 90/10 during the forecast window, run e2eTD and measure the coverage of the 90% predictive interval for each bottom series; if coverage falls well below nominal levels while the aggregate-level intervals remain calibrated, the stationarity of split proportions is the point of failure. Alternatively, use a public dataset with a known assortment change in the test period and compare e2eTD's bottom-level WSPL against a variant that re-estimates proportions on the most rece","tokens_in":20467,"feed_emoji":"📊","tokens_out":5494,"duration_ms":56122,"temperature":0.7,"pith_summary":"This paper claims that coherent probabilistic forecasts for hierarchies of hundreds of thousands of retail series can be produced by forecasting only a small set of smooth aggregate series and then probabilistically splitting those forecasts down to the bottom level, instead of forecasting every bottom series. The split is driven by historical joint distributions of sibling series—Poisson or negative-binomial marginals coupled through a bivariate Plackett copula—so the resulting bottom-level samples preserve cross-series dependence and sum exactly to the aggregate samples. On the two largest public retail datasets the method reports the lowest weighted scaled pinball loss among the compared methods at essentially every aggregation level, and would have ranked 11th of 892 teams in the M5 uncertainty competition. Because only about 0.3% of series are directly forecast and the splitting step is batched and parallel, the entire pipeline runs in about five minutes on roughly 40K series and twenty minutes on roughly 300K series on a standard laptop. A reader should care because forecasting cost at retail scale is a first-order operational concern, and coherence across levels is what lets replenishment, capacity, and strategic decisions rest on the same view of future demand.","feed_headline":"Forecasting 0.3% of series tops M5 and Favorita hierarchies","feed_subtitle":"Coherent probabilistic forecasts for the whole hierarchy follow from a fast top-down split of few aggregate forecasts.","key_machinery":"The load-bearing object is the probabilistic top-down sampling algorithm. For a parent total z, the set of feasible splits (a,b) with a+b=z is assigned probability proportional to the historical joint distribution p_{A,B}(a,b), estimated in-sample with Poisson or negative-binomial marginals and a bivariate Plackett copula whose dependence parameter is recovered from Spearman's rank correlation. A hierarchy is traversed by recursively applying this bivariate split along a binary partition of each group's descendants, and the procedure is made fast by computing conditional split weights once per distinct total value, reusing the same joint distributions across all forecast horizons, and parall","core_discovery":"The central claim is that a top-down strategy can be made fully probabilistic and end-to-end: instead of forecasting the noisy bottom series, e2eTD fits univariate probabilistic models to a small subhierarchy of aggregate series, reconciles those upper forecasts, and then disaggregates every sample through a sequence of bivariate splits. At each split the count total is divided between two sibling groups by sampling from the in-sample joint distribution of those groups conditioned on their sum. Because the same sample index is carried through all splits, the bottom samples remain jointly dependent, so summing them gives coherent forecasts for every aggregation level. The empirical claim is t","pith_inferences":["Editor's inference: the paper's stationarity assumption suggests a natural stress test—if assortment changes, promotions, or trend shift sibling split proportions within the 28-day horizon, bottom-level calibration should degrade; a time-varying proportion model (e.g., a shorter half-life or a trend in the copula parameter) is the obvious extension the authors leave open.","Editor's inference: because the method only needs aggregate forecasts, it could be combined with any upper-level forecaster that handles exogenous regressors; one testable extension is plugging a quantile-regression or gradient-boosted forecaster at the top and checking whether the accuracy gap widens or narrows on datasets with strong promotional dynamics.","Editor's inference: the binary-split reduction means the number of copula fits is roughly twice the number of bottom series, so for hierarchies with millions of series the per-node Plackett fit is the scaling bottleneck to watch; a shared or amortized estimate of the dependence structure would be the natural next step.","Editor's inference: the paper evaluates only 28-day horizons, but the method's cost structure suggests it would also suit rolling weekly refreshes; a testable claim is that re-estimating the joint distributions every week with a 28-day half-life preserves accuracy while staying within the reported time budget."],"forward_implications":["At retail scale, coherent probabilistic forecasts no longer require fitting models to the majority of series: cost scales with the number of selected upper series plus the splitting work, so a hierarchy of hundreds of thousands of series can be refreshed on a laptop in minutes.","Because every sample satisfies the aggregation constraints, decisions at SKU, store, and national levels can be drawn from one consistent predictive distribution, enabling quantile-based replenishment and safety-stock setting at every level.","Bottom-level intermittency emerges naturally from splitting smooth aggregate forecasts, so the method avoids modeling intermittent bottom series directly while still delivering competitive upper-tail quantiles at the SKU level.","Ablation results indicate the accuracy gains are concentrated in the probabilistic top-down mechanism rather than in a specific upper-level forecaster: replacing ETS with ARIMA or shrinking the upper subhierarchy degrades accuracy only mildly, whereas removing reconciliation or omitting regressors hurts more.","If the historical joint distributions remain approximately stationary over the forecast horizon, the paper's results imply that static in-sample distributional information suffices to outperform much heavier end-to-end neural models on major retail benchmarks."],"fun_headline_variants":["Forecast 0.3% of series, win M5 and Favorita","Probabilistic top-down: 0.3% of series, full coherence","End-to-end top-down beats M5 and Favorita baselines","Fast coherent forecasts from 0.3% of hierarchy","Top-down sampling: coherent forecasts at scale"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the historical joint distribution of each pair of sibling series, estimated once from in-sample data with a 28-day half-life, remains a valid description of how aggregate counts will split across those siblings over the 28-day forecast horizon; if true split proportions shift, the bottom-level samples and every aggregate built from them will be miscalibrated even when the upper forecasts are perfect.","fun_headline_variants_meta":{"raw":{"variants":["Forecast 0.3% of series, win M5 and Favorita","Probabilistic top-down: 0.3% of series, full coherence","End-to-end top-down beats M5 and Favorita baselines","Fast coherent forecasts from 0.3% of hierarchy","Top-down sampling: coherent forecasts at scale"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000429,"raw_usage":{"total_tokens":2067,"prompt_tokens":821,"completion_tokens":1246,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":1168}},"tokens_in":565,"tokens_out":1246,"duration_ms":7786,"temperature":1.0,"reasoning_tokens":1168,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T10:01:33.121025+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a simulated hierarchy with two bottom series whose in-sample split proportion is 50/50 but whose true proportion shifts to 90/10 during the forecast window, run e2eTD and measure the coverage of the 90% predictive interval for each bottom series; if coverage falls well below nominal levels while the aggregate-level intervals remain calibrated, the stationarity of split proportions is the point of failure. Alternatively, use a public dataset with a known assortment change in the test period and compare e2eTD's bottom-level WSPL against a variant that re-estimates proportions on the most rece","supporting_citations":[],"review_version":2}