Pith. sign in

REVIEW 4 major objections 4 minor 35 references

BinConv: A Neural Architecture for Ordinal Encoding in Time-Series Forecasting

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read 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.

desk verdict 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. read the letter →

arxiv 2505.24595 v3 pith:B5WEMCKQ submitted 2025-05-30 cs.LG cs.AIstat.ML

Andrei Chernov , Vitaliy Pozdnyakov , Ilya Makarov This is my paper
classification cs.LGcs.AIstat.ML
keywords probabilistictimeseriesforecastingcumulativebinaryencodingordinaldiscretizationconvolutionalneuralnetworkautoregressivepointquantization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

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.

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 (4)
  1. [Section 3, Eq. (1)-(2)]
  2. [Section 5.3, Tables 3 and 5]
  3. [Section 6.1, Table 8]
  4. [Section 6.4, Table 8]
minor comments (4)
  1. [Section 6.3]
  2. [Table 1 and Table 2]
  3. [Section 6.2]
  4. [General]

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: BinConv's central claim is evaluated against external baselines on public benchmarks with fixed, untuned hyperparameters.

full rationale

The paper's main claim is empirical: BinConv with CBE outperforms DLinear, PatchTST, GRU-NVP, and TimeGrad on ProbTS benchmarks. Hyperparameters are fixed (Section 5.1, Table 2), not fitted to the test set, and baselines use independently published configurations. The encoding (Section 3) is defined via explicit thresholds and an inverse transform; the inverse transform's midpoint rule is a definition, not a prediction smuggled from data. The ablation claims (CBE+Transformer, one-hot+BinConv) are in-house empirical comparisons, not self-citations or definitional reductions. The paper cites Gorishniy et al. (2022) as prior work for piecewise linear encoding and explicitly states CBE is a simplified version; this is an acknowledgment, not a load-bearing self-citation. There are no author self-citations in the reference list. The only notable weakness is an empirical assumption, not circularity: the inverse transform is unspecified for all-ones/all-zeros encodings, so values outside [-5,5] after mean scaling saturate; Section 6.1 acknowledges this class of issue ('most targets are well below the maximum bin value bD'). This is a correctness/generalization risk, not a case where the derivation or prediction reduces to an input by construction.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The central claims rest on a hand-picked quantization grid and a set of architectural hyperparameters; the grid is the most consequential because it bounds the forecastable range. The generative distribution relies on a Bernoulli-independence assumption. No new physical or conceptual entities are introduced.

free parameters (6)
  • Number of bins D = 1000 univariate, 500 multivariate
    Hyperparameter chosen by the authors; D=500 for multivariate to accelerate runtime; not fitted to data, but the ledger includes hand-chosen numbers that affect the resolution and range of the encoding.
  • Bin range b0, bD = -5 and 5
    Fixed symmetric range after scaling; no per-dataset tuning; this range determines saturation behavior at the edges.
  • Context length C = 3H for univariate, H for multivariate
    Set to 3H and H respectively; chosen by the authors and critical for the receptive field of the 2D convolution.
  • Number of blocks M = 3
    Depth of the residual stack; chosen by the authors without ablation.
  • Final conv kernel size s3 = 51
    Large kernel over the bin dimension; chosen by the authors without sensitivity analysis.
  • Dropout = 0.35
    Regularization hyperparameter; fixed across datasets.
assumptions (3)
  • domain assumption Sigmoid outputs at each bin position are independent Bernoulli trials given the context
    Equation (2) defines the probability of a valid sequence as a product of independent Bernoulli probabilities over components; this conditional independence is an assumption about the model's predictive distribution.
  • ad hoc to paper The fixed bin grid b0=-5, bD=5, D=1000 after mean scaling covers all target values encountered
    Used throughout Section 3 and the inverse transform; no handling for out-of-range values is specified, so the grid is assumed to be sufficient for the benchmark data.
  • domain assumption Mean scaling with s = mean absolute value over the context window is an adequate normalization
    Borrowed from DeepAR/Chronos; relies on context mean being positive and representative, which may not hold for series with many zeros or negative values.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BinConv: A Neural Architecture for Ordinal Encoding in Time-Series Forecasting." pith.science (2026). https://pith.science/paper/B5WEMCKQ

@misc{pith2026250524595,
  author       = {Pith},
  title        = {Pith review of: BinConv: A Neural Architecture for Ordinal Encoding in Time-Series Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B5WEMCKQ}},
  note         = {Machine review of arXiv:2505.24595}
}
read the original abstract

Recent work in time series forecasting has explored reformulating regression as a classification task. By discretizing the continuous target space into bins and predicting over a fixed set of classes, these approaches benefit from more stable training, improved uncertainty modeling, and compatibility with modern deep learning architectures. However, most existing methods rely on one-hot encoding, which ignores the inherent ordinal structure of the target values. As a result, they fail to convey information about the relative distance between predicted and true values during training. In this paper, we address this limitation by applying \textbf{Cumulative Binary Encoding} (CBE), a monotonic binary representation that transforms both model inputs and outputs. CBE implicitly preserves ordinal and magnitude information, allowing models to learn distance aware representations while operating within a classification framework. To leverage CBE effectively, we propose \textbf{BinConv}, a fully convolutional neural network architecture designed for probabilistic forecasting. We demonstrate that standard fully connected layers are not only less computationally efficient than convolutional layers when used with CBE, but also degrade forecasting performance. Our experiments on standard benchmark datasets show that BinConv achieves superior performance compared to widely used baselines in both point and probabilistic forecasting, while requiring fewer parameters and enabling faster training.

Figures

Figures reproduced from arXiv: 2505.24595 by the authors.

Figure 2
Figure 2. BinConv architecture. (a) First, we preprocess the [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Converting the neural network output [0.4, 0.9, 0.2] into a CBE, the most probable sequence is [1, 1, 0]. 5 Experiments 5.1 Setup We evaluate performance on three univariate datasets: M4 Daily (daily), M4 Weekly (weekly), and Tourism Monthly (tourism) (Makridakis, Spiliotis, and Assimakopoulos 2020; Athanasopoulos et al. 2011). The prediction horizon H for each dataset is set to 14, 13, and 24, respectively. The con… view at source ↗
Figure 4
Figure 4. Forecast comparison of BinConv and BinConv w/ [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 23 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Akita, R.; Yoshihara, A.; Matsubara, T.; and Uehara, K. 2016. Deep learning for stock prediction using numerical and textual information. In 2016 IEEE/ACIS 15th International Conference on Computer and Information Science (ICIS), 1--6. IEEE

    work page 2016

  4. [4]

    F.; Stella, L.; Turkmen, A

    Ansari, A. F.; Stella, L.; Turkmen, A. C.; Zhang, X.; Mercado, P.; Shen, H.; Shchur, O.; Rangapuram, S. S.; Arango, S. P.; Kapoor, S.; Zschiegner, J.; Maddix, D. C.; Wang, H.; Mahoney, M. W.; Torkkola, K.; Wilson, A. G.; Bohlke-Schneider, M.; and Wang, B. 2024. Chronos: Learning the Language of Time Series. Transactions on Machine Learning Research. Exper...

    work page 2024

  5. [5]

    J.; Song, H.; and Wu, D

    Athanasopoulos, G.; Hyndman, R. J.; Song, H.; and Wu, D. C. 2011. The tourism forecasting competition. International Journal of Forecasting, 27(3): 822--844

    work page 2011

  6. [6]

    Z.; and Koltun, V

    Bai, S.; Kolter, J. Z.; and Koltun, V. 2018. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv preprint arXiv:1803.01271

    arXiv 2018

  7. [7]

    Bandara, K.; Shi, P.; Bergmeir, C.; Hewamalage, H.; Tran, Q.; and Seaman, B. 2019. Sales demand forecast in e-commerce using a long short-term memory neural network methodology. In Neural Information Processing: 26th International Conference, ICONIP 2019, Sydney, NSW, Australia, December 12--15, 2019, Proceedings, Part III 26, 462--474. Springer

    work page 2019

  8. [8]

    Che, Z.; Purushotham, S.; Cho, K.; Sontag, D.; and Liu, Y. 2018. Recurrent neural networks for multivariate time series with missing values. Scientific reports, 8(1): 6085

    work page 2018

Show all 35 references
  1. [9]

    Chen, Y.; Kang, Y.; Chen, Y.; and Wang, Z. 2020. Probabilistic forecasting with temporal convolutional neural network. Neurocomputing, 399: 491--501

    2020

  2. [10]

    Furman, .; Duch, W.; Minati, L.; and To pa, K. 2023. Short-time Fourier transform and embedding method for recurrence quantification analysis of EEG time series. The European Physical Journal Special Topics, 232(1): 135--149

    2023

  3. [11]

    Gorishniy, Y.; Rubachev, I.; and Babenko, A. 2022. On embeddings for numerical features in tabular deep learning. Advances in Neural Information Processing Systems, 35: 24991--25004

    2022

  4. [12]

    Gruver, N.; Finzi, M.; Qiu, S.; and Wilson, A. G. 2023. Large language models are zero-shot time series forecasters. Advances in Neural Information Processing Systems, 36: 19622--19635

    2023

  5. [13]

    Hochreiter, S.; and Schmidhuber, J. 1997. Long short-term memory. Neural computation, 9(8): 1735--1780

    1997

  6. [14]

    J.; and Athanasopoulos, G

    Hyndman, R. J.; and Athanasopoulos, G. 2021. Forecasting: principles and practice. OTexts. OTexts: Melbourne, Australia. OTexts.com/fpp3. Accessed on May 2025

    2021

  7. [15]

    Kingma, D. 2015. Adam: a method for stochastic optimization. arXiv: 1412.6980

    2015 arXiv

  8. [16]

    Li, W.; and Law, K. E. 2024. Deep learning models for time series forecasting: a review. IEEE Access

    2024

  9. [17]

    \"O .; Loeff, N.; and Pfister, T

    Lim, B.; Ar k, S. \"O .; Loeff, N.; and Pfister, T. 2021. Temporal fusion transformers for interpretable multi-horizon time series forecasting. International Journal of Forecasting, 37(4): 1748--1764

    2021

  10. [18]

    Lim, B.; and Zohren, S. 2021. Time-series forecasting with deep learning: a survey. Philosophical Transactions of the Royal Society A, 379(2194): 20200209

    2021

  11. [19]

    Liu, Y.; Zhang, H.; Li, C.; Huang, X.; Wang, J.; and Long, M. 2024. Timer: generative pre-trained transformers are large time series models. In Proceedings of the 41st International Conference on Machine Learning, 32369--32399

    2024

  12. [20]

    Liu, Z.; Mao, H.; Wu, C.-Y.; Feichtenhofer, C.; Darrell, T.; and Xie, S. 2022. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 11976--11986

    2022

  13. [21]

    Luo, D.; and Wang, X. 2024. Moderntcn: A modern pure convolution structure for general time series analysis. In The twelfth international conference on learning representations, 1--43

    2024

  14. [22]

    Makridakis, S.; Spiliotis, E.; and Assimakopoulos, V. 2020. The M4 Competition: 100,000 time series and 61 forecasting methods. International Journal of Forecasting, 36(1): 54--74

    2020

  15. [23]

    H.; Sinthong, P.; and Kalagnanam, J

    Nie, Y.; Nguyen, N. H.; Sinthong, P.; and Kalagnanam, J. 2023. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. In The Eleventh International Conference on Learning Representations

    2023

  16. [24]

    Rasul, K.; Seward, C.; Schuster, I.; and Vollgraf, R. 2021 a . Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting. In International conference on machine learning, 8857--8868. PMLR

    2021

  17. [25]

    M.; and Vollgraf, R

    Rasul, K.; Sheikh, A.-S.; Schuster, I.; Bergmann, U. M.; and Vollgraf, R. 2021 b . Multivariate Probabilistic Time Series Forecasting via Conditioned Normalizing Flows. In International Conference on Learning Representations

    2021

  18. [26]

    Salinas, D.; Flunkert, V.; Gasthaus, J.; and Januschowski, T. 2020. DeepAR: Probabilistic forecasting with autoregressive recurrent networks. International journal of forecasting, 36(3): 1181--1191

    2020

  19. [27]

    Song, Y.; Gao, S.; Li, Y.; Jia, L.; Li, Q.; and Pang, F. 2020. Distributed attention-based temporal convolutional network for remaining useful life prediction. IEEE Internet of Things Journal, 8(12): 9594--9602

    2020

  20. [28]

    Tay, Y.; Dehghani, M.; Bahri, D.; and Metzler, D. 2022. Efficient transformers: A survey. ACM Computing Surveys, 55(6): 1--28

    2022

  21. [29]

    L.; Joosen, A.; Singh, R.; and Asenov, M

    Toner, W.; Lee, T. L.; Joosen, A.; Singh, R.; and Asenov, M. 2025. Performance of zero-shot time series foundation models on cloud data. arXiv preprint arXiv:2502.12944

    2025 arXiv

  22. [30]

    N.; Kaiser, .; and Polosukhin, I

    Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, .; and Polosukhin, I. 2017. Attention is all you need. Advances in neural information processing systems, 30

    2017

  23. [31]

    W.; Ahmed, A

    Zaini, N.; Ean, L. W.; Ahmed, A. N.; and Malek, M. A. 2022. A systematic literature review of deep learning neural network for time series air quality forecasting. Environmental Science and Pollution Research, 1--33

    2022

  24. [32]

    Zeng, A.; Chen, M.; Zhang, L.; and Xu, Q. 2023. Are transformers effective for time series forecasting? In Proceedings of the AAAI conference on artificial intelligence, volume 37, 11121--11128

    2023

  25. [33]

    Zhang, J.; Wen, X.; Zhang, Z.; Zheng, S.; Li, J.; and Bian, J. 2024. ProbTS: Benchmarking point and distributional forecasting across diverse prediction horizons. Advances in Neural Information Processing Systems, 37: 48045--48082

    2024

  26. [34]

    Zhou, H.; Zhang, S.; Peng, J.; Zhang, S.; Li, J.; Xiong, H.; and Zhang, W. 2021. Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting. In The Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Virtual Conference , volume 35, 11106--...

    2021

  27. [35]

    Zhu, J.; Chen, X.; He, K.; LeCun, Y.; and Liu, Z. 2025. Transformers without normalization. arXiv preprint arXiv:2503.10622

    2025 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.