{"id":"81de8d77-a278-4700-9466-912d54649dd0","arxiv_id":"2505.24595","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A paper combines a cumulative binary encoding of target values with a lightweight convolutional architecture to improve time series forecasting accuracy and speed.","lead":"The paper proposes a new way to encode time series values as ordered binary codes, together with a small convolutional network, and reports improved point and probabilistic forecasts on standard benchmarks. The idea matters because it gets better accuracy and faster training with far fewer parameters than current transformer baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Saturation at the fixed [-5,5] bin range is the load-bearing risk: the inverse transform in Section 3 is unspecified for all-ones/all-zeros encodings, and the paper never quantifies how often per-sample mean-scaled test targets fall outside the grid.","rationale":"I read the paper's central claim as the benchmark superiority of BinConv. For that claim to hold, the CBE quantization must be a faithful representation of the scaled target distribution. The fixed data-agnostic grid [-5,5] is the weakest point: Section 3 leaves saturated encodings undefined, and Section 6.1 gives only a casual, unquantified assertion that most targets lie below the upper edge. This is not an internal inconsistency in the math, but it is a missing empirical guarantee that directly bears on whether the reported metrics measure the intended method. A simple instrumentation test can settle it. The reader identified exactly this as the weakest assumption, so my read agrees. I also note the fully connected ablation contradiction in Table 8 (w/FC is better on two of three datasets), which undermines a secondary claim about FC layers degrading performance, but that does not affect the primary benchmark comparison. No verdict change is needed beyond the reader's CONDITIONAL.","tokens_in":12202,"tokens_out":11016,"duration_ms":148148,"concrete_test":"Instrument the evaluation (or a reimplementation) to count, for every dataset and test-horizon step, how many ground-truth targets fall outside [-5,5] after applying the per-sample mean scaling described in Section 3. If the count is nonzero, recompute Table 3 with a wider range, e.g., b0 = -20, bD = 20, keeping D fixed, and also define the all-ones and all-zeros inverse rules. If the NMAE/CRPS values change materially, the reported superiority is partly an artifact of clipping; if they do not change, the saturation concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central benchmark claim requires that the CBE quantization faithfully represents the scaled target distribution. Section 3 fixes b0 = -5, bD = 5 with D = 1000 (or 500) bins and uses per-sample mean scaling before quantization. The inverse transform is defined only as (bi + bi+1)/2 where i is the index of the last 1; the cases i = 0 (all bits zero, scaled value below b0) and i = D (all bits one, scaled value above bD) are never specified. Such values silently saturate at the edge bins. The only support offered is the statement in Section 6.1 that 'most targets are well below the maximum bin value,' which is unquantified, addresses only the upper bound, and says nothing about the lower bound. If a nontrivial fraction of test targets saturate after mean scaling, the reported NMAE and CRPS come from a capped forecaster that cannot produce extremes, so the comparison to baselines is not a test of the method as described. This is an empirical assumption, not a mathematical guarantee, and it is the least secure condition underlying the headline result.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Cumulative Binary Encoding (CBE), a monotonic binary quantization of mean-scaled time series values, and BinConv, a fully convolutional architecture that operates in this CBE space. BinConv is evaluated on three univariate benchmarks (M4 Daily, M4 Weekly, Tourism Monthly) and five multivariate datasets from the ProbTS benchmark. The authors report lower CRPS and NMAE on univariate datasets than DLinear, PatchTST, GRU-NVP, and TimeGrad, with competitive average ranks on multivariate datasets and lower parameter counts than most baselines. Ablation studies compare fully connected layers, transformers, one-hot encoding, and depthwise convolutions, and the paper claims that CBE and BinConv are effective only when used together.","tokens_in":12479,"tokens_out":6935,"duration_ms":84536,"significance":"If the empirical claims hold, the paper makes a modest but useful contribution: it adapts piecewise-linear-style ordinal encoding to time-series forecasting and shows that a convolutional architecture with strong weight sharing is a natural fit for cumulative binary representations. The five-seed univariate experiments, the use of public benchmarks, the efficiency measurements, and the ablation battery are strengths. However, the contribution is incremental relative to prior work on PLE and convolutional forecasting, and the verification is limited by missing code release, under-reported statistical methodology, unresolved boundary handling, and a few internal inconsistencies between the text and reported tables.","major_comments":[{"comment":"","section":"Section 3, Eq. (1)-(2)"},{"comment":"","section":"Section 5.3, Tables 3 and 5"},{"comment":"","section":"Section 6.1, Table 8"},{"comment":"","section":"Section 6.4, Table 8"}],"minor_comments":[{"comment":"","section":"Section 6.3"},{"comment":"","section":"Table 1 and Table 2"},{"comment":"","section":"Section 6.2"},{"comment":"","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable incremental contribution, but the empirical verification is not yet at the level that would justify the strength of the claims. The boundary handling in Section 3 is a genuine gap that affects the method as described, the FC ablation contradicts the abstract on two of three datasets, and the multivariate average ranks rest on single-seed runs. None of these issues appear to be unfixable within the scope of the paper, but they require substantive work rather than copy editing. I would also encourage the editor to require code availability for an empirical paper of this type."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe short version: this paper has a real empirical result and a few seams that need attention before the claims are clean. The central claim—that predicting directly in cumulative binary encoding space with a convolutional architecture gives better CRPS and NMAE than transformer baselines on the three univariate datasets—is supported by Table 3, and the parameter counts and training times are believable. The paper credits CBE as a variant of piecewise linear encoding (Gorishniy et al. 2022) and adds a genuinely new piece: forecasting directly in the encoding space, with proper normalization over the set of valid monotone sequences. That is a legitimate contribution to the discretized-forecasting line.\n\nWhat the paper does well: the ablations are informative. Comparing CBE+BinConv against one-hot+BinConv and CBE+transformer makes the case that the combination, not either ingredient alone, is what delivers. The synthetic linear-trend experiment in Section 6.1 is a nice sanity check for extrapolation, and reporting min/avg/max over five seeds on univariate data is better than most papers in this area.\n\nSoft spots, in order.\n\nFirst, the fixed bin range [-5,5] is a genuine risk handled too casually. Section 3 defines the inverse transform only as the midpoint of the crossed bin; all-zeros and all-ones encodings are not specified. The text says \"most targets are well below the maximum bin value\" without quantifying the lower tail. With per-sample mean scaling the risk of saturation is reduced, but the forecaster still cannot produce extremes if scaled values fall outside the grid. This should be quantified on the actual test sets, not asserted.\n\nSecond, Table 8 contradicts the prose in Section 6.1. The text says the fully connected variant degrades performance, but the table shows w/FC with lower NMAE than BinConv on weekly and tourism (0.0943 vs 0.0972 and 0.1910 vs 0.1955). Only on daily is it worse. The extrapolation argument is supported by the synthetic experiment, so the section's core point may survive, but the empirical claim as written is inaccurate.\n\nThird, the multivariate results are single-seed, p-values are reported only as p<0.05 without exact values or correction, and no code is released. All are fixable in revision.\n\nThe citation pattern is sound; the paper builds on the right literature and credits PLE properly. For a reader, someone working on discretization or efficient probabilistic forecasting gets real value. It deserves a serious referee. I would send it to review, with the bin-range saturation and Table 8 overclaim as required revisions.\n\nBest,\n[Your name]","headline":"Useful empirical combination of cumulative binary encoding with a convolutional forecaster; the main claims mostly hold on univariate benchmarks, but the fixed quantization range and a contradictory ablation need attention before the paper is clean.","tokens_in":12964,"tokens_out":3255,"would_cite":false,"duration_ms":36150,"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 cumulative binary encoding that preserves order in discretized time series lets a convolutional forecaster outpredict standard baselines on several benchmarks while using fewer parameters.","keywords":["probabilistic time series forecasting","cumulative binary encoding","ordinal encoding","discretization","convolutional neural network","autoregressive forecasting","point forecasting","quantization"],"falsifier":"Take a test series with a strong upward trend or an anomalous spike so that mean-scaled values exceed 5, and compare BinConv's forecast to a continuous-output baseline on those extreme steps. If BinConv systematically underestimates the extremes because all bits saturate at 1, while the baseline tracks them, that would contradict the paper's claim of consistently superior performance. A cheaper check: compute the fraction of validation samples whose mean-scaled values fall outside $[-5,5]$ on a heavy-tailed dataset and show that CRPS degradation grows with that fraction.","tokens_in":12007,"feed_emoji":"📈","tokens_out":6958,"duration_ms":75396,"temperature":0.7,"pith_summary":"This paper claims that the standard trick of turning time-series forecasting into classification by binning the target loses the order and distance information between bins, and that replacing one-hot codes with a cumulative binary code fixes this. The authors introduce BinConv, a convolutional network that reads and writes this code, and report that on M4 Daily, M4 Weekly, and Tourism Monthly it achieves lower average CRPS and NMAE than DLinear, PatchTST, GRU-NVP, and TimeGrad while using fewer parameters and training faster. The central claim is that the encoding and the architecture are mutually dependent: the cumulative code fails when paired with fully connected layers, and the convolutional network fails when paired with one-hot encoding. A sympathetic reader should come away with a concrete recipe for cheap probabilistic forecasting: quantize, encode monotonically, and let a weight-sharing convolution predict bits.","feed_headline":"Monotone binary code plus CNN beats four forecast baselines","feed_subtitle":"Better point and probabilistic scores with fewer parameters, and no per-dataset tuning.","key_machinery":"The central object is Cumulative Binary Encoding (CBE), the map $x \\mapsto [v_1,\\ldots,v_D]$ with $v_d = \\mathbf{1}\\{\\tilde{x} \\ge b_d\\}$, where $\\tilde{x}$ is the value divided by the mean absolute value over the context window and $b_0,\\ldots,b_D$ are a fixed uniform grid from $-5$ to $5$ (1000 bins on univariate data, 500 on multivariate). Its companion is the valid-sequence renormalization of Equation (2): sigmoid outputs are treated as independent Bernoulli probabilities, probabilities of monotone $11\\ldots10\\ldots0$ codes are renormalized by their total mass, and forecasting proceeds autoregressively by either sampling codes or taking the argmax. The BinConv block applies a 2D convolution over the context-by-bins matrix, two depthwise 1D convolutions with ReLU and dropout, DyTanh activations in place of normalization, and a residual connection, ending in a wide 1D convolution with sigmoid. These components do the work of the argument because the weight sharing across bit positions is what the paper claims enables extrapolation, and the monotone-code restriction is what turns a binary classifier into a probabilistic forecaster.","core_discovery":"The paper's central discovery is that a monotonic binary representation called Cumulative Binary Encoding (CBE) — a vector of $D$ bits where bit $d$ is 1 when the mean-scaled value $\\tilde{x}$ is at least the bin edge $b_d$, so the code always looks like $11\\ldots100\\ldots0$ — is an effective target space for autoregressive forecasting when paired with a convolutional architecture. The model outputs per-bit Bernoulli probabilities, restricts attention to valid monotone codes by renormalizing, and can therefore produce both a point forecast (argmax over valid codes) and a probabilistic forecast (sampled trajectories). On the paper's benchmarks, the average BinConv score beats every baseline average by more than one standard deviation on each univariate dataset for CRPS and NMAE, and BinConv attains the best average rank in CRPS and second-best in NMAE on multivariate datasets even though it forecasts each series independently. The same output layer thus serves both evaluation tasks, and the weight-sharing property of convolutions is what allows the model to extrapolate to values beyond those seen in training.","pith_inferences":["If the saturation risk at the grid edges is real, an adaptive variant that sets bin bounds from a robust scale estimate per series, or stretches the grid when forecasts hit the boundary, should improve extreme-quantile CRPS; I would test that before deploying on volatile series.","The valid-sequence renormalization is one way to enforce monotonicity; an alternative is to make monotonicity structural, for instance by predicting the bin index and feeding it through a cumulative-sum layer, which would remove the need to restrict the support.","Because CBE is a simplified piecewise linear encoding, replacing the bin-midpoint inverse transform with interpolation over the bin would reduce discretization bias and may close most of the remaining gap to continuous-output models.","The paper's explanation for why transformers fail with CBE is that they lack weight sharing across output positions; a transformer with tied weights or a convolutional attention pattern should be tested directly to pinpoint whether the failure is about weight sharing or about the input statistics."],"forward_implications":["Because forecasting is autoregressive on a fixed bin grid, the same trained model can be queried for any prediction horizon without retraining, which the non-autoregressive baselines cannot do.","Point and probabilistic forecasts come from the same output layer: the argmax of the renormalized valid-code distribution and the sampled trajectories, so uncertainty estimates cost almost no extra architecture.","The paper's ablations imply that weight sharing in convolutions is what allows extrapolation past the training distribution; fully connected heads saturate at the largest value seen during training.","The fixed default configuration, identical across datasets except for the number of bins, suggests that the method transfers without per-dataset hyperparameter search.","On multivariate data, forecasting each series independently with BinConv already matches or beats the baselines on average rank, implying a direct multivariate extension that models cross-series interactions is a natural next step."],"supporting_citations":[{"why":"Supplies the mean-scaling and discretization approach that CBE extends, and the Chronos-style token view of time series.","marker":"Ansari et al. 2024"},{"why":"Source of the piecewise linear encoding that CBE simplifies for tabular features.","marker":"Gorishniy, Rubachev, and Babenko 2022"},{"why":"Provides the benchmark datasets, baseline hyperparameters, and NMAE/CRPS evaluation setup used in all experiments.","marker":"Zhang et al. 2024"},{"why":"Defines the mean scaling operator $s = \\frac{1}{C}\\sum|x_i|$ and the autoregressive probabilistic forecasting paradigm.","marker":"Salinas et al. 2020"},{"why":"The PatchTST baseline, the strongest transformer competitor, which BinConv is compared against.","marker":"Nie et al. 2023"},{"why":"The DLinear baseline, used as an efficient non-probabilistic reference point.","marker":"Zeng et al. 2023"},{"why":"The TimeGrad diffusion baseline providing probabilistic forecasts.","marker":"Rasul et al. 2021a"},{"why":"The GRU-NVP normalizing-flow baseline providing probabilistic forecasts.","marker":"Rasul et al. 2021b"},{"why":"Supplies the DyTanh activation used inside BinConv blocks.","marker":"Zhu et al. 2025"},{"why":"Justifies the depthwise convolution design borrowed by BinConv blocks.","marker":"Luo and Wang 2024"}],"fun_headline_variants":["Binary encoding + CNN beats baselines in time-series forecasting","Cumulative binary encoding powers conv net to top forecast scores","Distance-aware binary targets give CNN edge in forecasting","BinConv: ordinal encoding with conv net beats forecast baselines","Fewer parameters, faster training: BinConv wins forecasting benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the fixed bin grid from $-5$ to $5$, applied after per-sample mean scaling, actually covers every test value; anything outside the grid is clamped at the edges, so the forecaster can never emit extremes beyond the grid.","fun_headline_variants_meta":{"raw":{"variants":["Binary encoding + CNN beats baselines in time-series forecasting","Cumulative binary encoding powers conv net to top forecast scores","Distance-aware binary targets give CNN edge in forecasting","BinConv: ordinal encoding with conv net beats forecast baselines","Fewer parameters, faster training: BinConv wins forecasting benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000212,"raw_usage":{"total_tokens":1435,"prompt_tokens":981,"completion_tokens":454,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":372}},"tokens_in":597,"tokens_out":454,"duration_ms":5490,"temperature":1.0,"reasoning_tokens":372,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:17:04.353293+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a test series with a strong upward trend or an anomalous spike so that mean-scaled values exceed 5, and compare BinConv's forecast to a continuous-output baseline on those extreme steps. If BinConv systematically underestimates the extremes because all bits saturate at 1, while the baseline tracks them, that would contradict the paper's claim of consistently superior performance. A cheaper check: compute the fraction of validation samples whose mean-scaled values fall outside $[-5,5]$ on a heavy-tailed dataset and show that CRPS degradation grows with that fraction.","supporting_citations":[{"cited_title":"F.; Stella, L.; Turkmen, A","cited_arxiv_id":null,"evidence_quote":"Supplies the mean-scaling and discretization approach that CBE extends, and the Chronos-style token view of time series."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the piecewise linear encoding that CBE simplifies for tabular features."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the benchmark datasets, baseline hyperparameters, and NMAE/CRPS evaluation setup used in all experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the mean scaling operator $s = \\frac{1}{C}\\sum|x_i|$ and the autoregressive probabilistic forecasting paradigm."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Justifies the depthwise convolution design borrowed by BinConv blocks."}],"review_version":1}