Pith. sign in

REVIEW 4 major objections 7 minor 2 cited by

This paper claims that forking-sequences—jointly encoding all forecast creation dates and decoding them in parallel—reduces gradient and forecast-revision variance at rate O(1/T), yielding more stable, more accurate forecasts.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 11:27 UTC pith:LMVV7VND

load-bearing objection A useful formalization of a known averaging trick, with an empirical comparison that needs a fair baseline before the accuracy claims can be trusted. the 4 major comments →

arxiv 2510.04487 v5 pith:LMVV7VND submitted 2025-10-06 cs.LG

Forking-Sequences: Statistically and Computationally Efficient Multi-Horizon Forecasting with Reduced Volatility

Willa Potosnak , Malcolm Wolff , Mengfei Cao , Ruijun Ma , Tatiana Konstantinova , Dmitry Efimov , Michael W. Mahoney , Boris Oreshkin
show 1 more author
Kin G. Olivares
This is my paper
classification cs.LG
keywords time series forecastingforecast revisionsforecast stabilityforking-sequencesgradient variance reductionprobabilistic forecastingtemporal cross-validationmulti-horizon forecasting
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Forking-sequences is a training and inference scheme for neural forecasters that treats all forecast creation dates jointly: one encoder pass over the observed series, then parallel decoders for every date-and-horizon cell of the forecast grid. The paper's central claim is that this architectural choice reduces the variance of stochastic gradients during training at rate O(1/T) under short-range dependence, so optimization converges faster, and that averaging the several forecasts pointing at the same target date reduces forecast-revision volatility at rate O(1/|H|). It also shows that reusing encoder outputs turns temporal cross-validation inference from quadratic into linear cost. Across 16 M-competition and Tourism datasets, keeping the architecture fixed and switching from window-sampling to forking-sequences improves scaled CRPS on average for all five encoder types, up to 43.2% for LSTM, and reduces a new symmetric quantile percentage change metric on average, up to 37.9% for LSTM. A sympathetic reader would care because erratic forecast revisions between creation dates are a real operational cost, and this work offers one mechanism that addresses accuracy, stability, and compute together.

Core claim

The central discovery is that the forking-sequences design—already present in some production forecasters but never formalized—is not only a pragmatic trick but a variance-reduction mechanism with provable rates. The paper formalizes the gradient estimator that sums losses over all FCDs in a batch, proves under M-dependence (gradient samples whose correlations vanish beyond a finite lag) that its variance decays as O(1/T), and proves an analogous ensemble variance reduction O(1/|H|) for averaging forecasts that share a target date. It further shows that the encoder can be computed once for the entire series and reused across all FCDs, dropping cross-validation inference from O(T^2) to O(T).

What carries the argument

The forking-sequences architecture itself: a single encoder consumes the full observed series and produces one hidden representation, and separate decoder heads emit forecasts for every forecast creation date and horizon simultaneously. The variance theorems ride on M-dependence—correlation between gradient samples or forecast errors vanishes beyond a finite lag—which lets the average over T dependent samples behave like an average of weakly correlated quantities and yields O(1/T) variance. The compute gain comes from reusing the encoder's activations across all FCDs, so the encoder runs once instead of once per FCD, changing inference cost from O(T^2) to O(T).

Load-bearing premise

The load-bearing premise is that gradient samples (and forecast errors) across forecast creation dates are M-dependent, with correlations that vanish after a finite lag; if real gradients are long-range dependent, the O(1/T) variance reductions do not follow, and the empirical comparison also rests on the window-sampling baseline being a fair match—the paper's own appendix shows the WS MLP sees a context of length 1 instead of 2H, WS models are not given the inference-time en

What would settle it

Record per-FCD gradient vectors during training and compute their autocorrelation at increasing lags; if correlations remain substantial at long lags, the O(1/T) bound in Theorem 1 does not hold. Equally decisive: retrain the window-sampling baseline with input size 2H, the same ensembling at inference, and the same number of FCDs per optimizer step (or matched total compute); if the reported sCRPS and sQPC gaps shrink substantially, the headline gains are due to underspecification of the baseline rather than to forking-sequences itself.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Under M-dependent gradient samples, training with forking-sequences reduces gradient variance at O(1/T), which accelerates convergence; the paper demonstrates faster train-loss convergence across five deep encoders and in a convex autoregressive ablation.
  • Averaging the forecasts aimed at a single target date from earlier forecast creation dates reduces forecast variance at O(1/|H|), yielding more stable revisions with negligible extra encoder cost because the encoder runs once.
  • Cross-validated inference over many forecast creation dates drops from O(T^2) to O(T) for convolutional and recurrent encoders, making large-scale temporal backtesting practical.
  • On 16 M-competition and Tourism datasets, switching a fixed architecture from window-sampling to forking-sequences improves average sCRPS for all five encoder types (up to 43.2% for LSTM) and reduces average median-forecast sQPC (up to 37.9% for LSTM).
  • Ensembling during inference cuts forecast volatility by roughly 10-13% across architectures while maintaining accuracy, suggesting that stability can be improved without an accuracy penalty.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the O(1/T) argument is the whole story, forking-sequences is essentially a cheap way to enlarge the effective batch over correlated forecast dates; a window-sampling baseline matched on context length (2H), on ensembling, and on total FCDs processed per optimizer step would reveal how much of the reported gain is architectural reuse versus raw sample count.
  • The sQPC metric—symmetric quantile percentage change between consecutive forecast creation dates—could become a standard report for operational forecasting, where revision volatility is a direct cost in planning and inventory systems.
  • The encoder-reuse pattern (one context, many time-offset targets) appears beyond forecasting, for example in survival analysis and streaming anomaly detection, so the variance-reduction argument may transfer to those settings.
  • Averaging forecasts across FCDs weights all revisions equally; a testable extension is horizon-weighted or recency-weighted ensembling to tune the accuracy-stability tradeoff further.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper formalizes forking-sequences, a training and inference scheme for multi-horizon neural forecasting in which a shared encoder processes all FCDs jointly and the decoder produces forecasts for every FCD in parallel. The authors claim three benefits: (i) gradient variance reduction at O(1/T) under M-dependence (Theorem 1), (ii) forecast variance reduction via ensembling at O(1/|H|) (Theorem 2), and (iii) inference complexity O(T) instead of O(T^2) by reusing encoder computations. They validate these claims on 16 M1/M3/M4/Tourism datasets with MLP, RNN, LSTM, CNN, and Transformer encoders, reporting sCRPS and sQPC gains over window-sampling.

Significance. If the empirical claims hold, the paper addresses an important and under-studied property of forecasting systems: revision volatility across forecast creation dates. The formalization of forking-sequences and the computational-complexity argument for encoder reuse are useful contributions, and the paper is honest about the simplicity of the variance-reduction mechanism. The theoretical results are elementary but correct in spirit; the main value of the paper is therefore empirical. However, the current experiments do not provide a controlled comparison between forking-sequences and window-sampling, so the headline improvements are not yet established. The paper also contains inconsistent headline numbers between the abstract and the body. The central idea is defensible, but the empirical evidence needs substantial strengthening.

major comments (4)
  1. [§3, Appendix C, Table 4] The FS-vs-WS comparison is not controlled. The WS MLP uses input size 1 while FS uses 2H (Table 4), so the WS MLP is a strictly different, weaker model. Section 2.4 states that FS models receive a moving-average forecast ensemble at inference, while WS models do not; the Discussion admits that 'almost all hyperparameter tuning' was skipped for WS; and the WS window length L is never reported. Because both schemes are trained for the same 30K steps while FS processes all FCDs per step, FS also receives far more gradient evaluations per step. Table 1/2 and the headline percentages therefore measure a combination of training scheme, model capacity, ensembling, and compute, not forking-sequences alone. Please rerun with matched context length, matched capacity, matched ensembling (WS forecasts can also be averaged via Eq. (6)), and matched total gradient evaluations or compute budget.
  2. [Abstract vs §3.1/§3.2, Tables 1–2] The headline numbers are internally inconsistent. The abstract reports median sCRPS improvements of 46.2%, 49.3%, 28.6%, 24.7%, and 6.4% for RNN, LSTM, CNN, Transformer, and State Space architectures; the body (Section 3.1) reports average improvements of 16.9% (RNN), 43.2% (LSTM), 19.2% (CNN), 1.8% (Transformer), and includes MLP (14.3%) while no State Space architecture appears in the experiments. The abstract also omits MLP from the list. The manuscript must present one consistent set of results, with average/median labeling explicit and every value traceable to Tables 1 and 2.
  3. [§2.4, Theorem 2, Appendix D] Theorem 2 is not proved. 'The proof is analogous to Theorem 1' is insufficient because the ensemble averages forecasts for the same target date made at different FCDs and horizons; these are not identically distributed, and the unbiasedness assumption on every member of H is strong and unverified. With horizon-dependent means, averaging can introduce bias, so convergence to the 'true value' requires a proof or a precise condition under which the average is unbiased. The same appendix defines Sigma ambiguously in the M-dependence definition; the Markov bound in Theorem 1 should use a long-run covariance rather than a per-lag covariance.
  4. [§3.2, Fig. 10] The text around Figs. 10a–10d is confusing. It says 'Ensembling during inference results in a marginal reduction in sCRPS across models, with substantial improvements in sQPC compared to no ensembling during inference, as shown in Figs. 10a and 10b' — but Figs. 10a/10b are labeled as 'without ensembling' versus window-sampling. Please report separately: (i) FS without ensembling vs WS (isolates the training scheme), (ii) FS with ensembling vs WS without ensembling (the current headline), and (iii) FS with ensembling vs WS with ensembling (isolates ensembling). The current presentation does not let the reader separate these effects.
minor comments (7)
  1. [Appendix D] Reference [43] is 'Shumway and Stoffer', not 'Stuffer'.
  2. [§2.4, Eq. (8)] Eq. (8) introduces an undefined function f(...) after the ensemble definition; either define it or remove it.
  3. [Appendix C, Tables 4–8] The window length L for the window-sampling baseline is never reported, despite being a key hyperparameter. Please list it for every architecture and frequency.
  4. [§2.2, Eq. (5)] The notation T is overloaded: it denotes the set of FCDs, its cardinality, and the total series length in Section 2.3. Use distinct symbols (e.g., cal T vs T).
  5. [Tables 1–2] Several WS standard deviations are very large relative to FS (e.g., LSTM M1 Y in Table 1), and no significance tests or confidence intervals are reported across the five runs. Consider adding paired tests or at least error bars on the percentage improvements.
  6. [Contributions (Section 1)] The bullet says accuracy is improved 'up to 43.2%', but Section 3.1 reports an average of 43.2% for LSTM. 'Up to' is inconsistent with an average.
  7. [Fig. 10] The y-axis on panels (b) and (d) ranges from -300 to 100, which is unusual for a percentage scale; check whether this is a typo and ensure all panels use the same scale.

Circularity Check

0 steps flagged

No significant circularity; variance theorems are self-contained averaging results, and the empirical baseline inequities are validity confounds rather than circular steps.

full rationale

The formal claims are not circular. Theorem 1 proves an O(1/T) variance bound for the averaged gradient estimator under M-dependence (Appendix D), which is a textbook weak-law-of-large-numbers calculation; the paper itself says it 'echo[es] results from the weak law of large numbers.' Theorem 2 is explicitly 'analogous to Theorem 1' and is just averaging of unbiased M-dependent forecasts. Neither theorem defines the target metric in terms of the estimator, fits a parameter and calls it a prediction, or imports a uniqueness theorem from the authors' prior work. The self-citations to MQCNN/MQT/SPADE are background identification of an existing design, not load-bearing evidence for the theoretical or empirical conclusions. The empirical comparison does contain material confounds: the paper says FS models 'also compute a moving average forecast ensemble during inference' while WS models do not; Appendix C Table 4 gives the WS MLP an input size of 1 versus 2H for FS; both schemes are trained for 30K steps even though FS uses all FCDs per step; and the Discussion admits 'we skipped almost all hyperparameter tuning.' These are experiment-design/validity concerns (the WS baseline may not be representative), not circular derivation. Because the central improvement claims are not forced by construction, the circularity score is low: 2.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central empirical claim rests on unequal baselines: unreported WS window length, asymmetric MLP inputs, ensemble applied only to FS, and unmatched compute. The theoretical results assume M-dependence/unbiasedness but are otherwise standard.

free parameters (4)
  • WS MLP input size = 1
    Window-sampling MLP encoder receives a single input feature while FS receives 2H (Appendix C, Table 4); this hand-chosen asymmetry handicaps the baseline and inflates FS gains.
  • WS window length L = not reported
    The context length for window-sampling baselines is never specified, so the comparison is underdetermined; if L is small, WS is unfairly weak.
  • Training steps = 30,000
    Both schemes stop at 30K steps, but FS computes losses over all FCDs per step, giving it more effective gradient information per step; a matched-compute comparison is missing.
  • Ensemble for FS only = moving average
    FS models average forecasts across FCDs at inference (Section 2.4), which directly smooths revisions; WS models get no ensemble, so part of the sQPC gain is mechanical.
axioms (4)
  • domain assumption Gradient samples across FCDs are M-dependent
    Theorem 1 assumes M-dependence of ∇L_t (Appendix D); for long-range dependent series this may fail, and no empirical verification is given.
  • domain assumption Ensembled forecasts are unbiased and M-dependent
    Theorem 2 requires unbiased forecasts; neural network quantile forecasts are generally biased, so the variance-reduction rate is conditional on an unverified premise.
  • domain assumption Data preprocessing follows Chronos/NBEATS protocols
    Section 3 cites [1,2,39] for preprocessing; details are not fully specified in this paper, so exact replication depends on external descriptions.
  • standard math Standard M-dependence covariance bound (Markov/Chebyshev)
    The proof of Theorem 1 uses the standard bound Cov(mean) = (1/T) Σ(1-|b|/T)γ(b) and Chebyshev's inequality; this is standard.

pith-pipeline@v1.3.0-alltime-deepseek · 21170 in / 17265 out tokens · 113799 ms · 2026-08-04T11:27:38.309765+00:00 · methodology

0 comments
read the original abstract

While accuracy is a critical requirement for time series forecasting, an equally important desideratum is reasonable forecast volatility across forecast creation dates (FCDs). Even highly accurate models can produce erratic revisions between FCDs, undermining trust and disrupting downstream decision-making. To improve the volatility of forecast revisions, state-of-the-art models like MQCNN, MQT, and SPADE employ a powerful yet underexplored neural network architectural design: forking-sequences. This architectural design jointly encodes and decodes the entire time series across all FCDs, producing an entire multi-horizon forecast grid in a single forward pass. This approach contrasts with conventional neural forecasting methods that process FCDs independently, generating only a single multi-horizon forecast per forward pass. In this work, we formalize the forking-sequences design and motivate its broader adoption by introducing a metric for quantifying excess volatility in forecast revisions and by providing theoretical and empirical analysis. We theoretically motivate three key benefits of forking-sequences: (i) reduced forecast volatility through ensembling; (ii) gradient variance reduction, improving the statistical efficiency of the training procedure; and (iii) improved inference computational efficiency. We validate the benefits of forking-sequences compared to baseline window-sampling on the M-series benchmark, using 16 datasets from the M1, M3, M4, and Tourism competitions. We observe median sCRPS improvements across datasets of 46.2%, 49.3%, 28.6%, 24.7%, and 6.4% for RNN, LSTM, CNN, Transformer, and State Space-based architectures, respectively. We then show that forecast ensembling during inference can reduce median forecast volatility by 13.2%, 13.0%, 10.9%, 10.2%, and 11.2% for these respective models trained with forking-sequences, while maintaining accuracy.

Figures

Figures reproduced from arXiv: 2510.04487 by Boris Oreshkin, Dmitry Efimov, Kin G. Olivares, Malcolm Wolff, Mengfei Cao, Michael W. Mahoney, Ruijun Ma, Tatiana Konstantinova, Willa Potosnak.

Figure 1
Figure 1. Figure 1: Comparison of forecast distributions on a series from the M1 dataset [31]. a) Forecasts generated without the forking-sequences ensemble. b) Forecasts with the forking-sequences ensemble applied. Augmenting the model with the forking-sequences ensemble significantly reduces forecast variability across forecast creation dates, resulting in more stable and consistent forecast distributions. The lines show P5… view at source ↗
Figure 2
Figure 2. Figure 2: a) A neural forecasting model trained using the window-sampling training scheme. Blue rectangles indicate encoder windows, while the green rectangles denotes decoder, the red circle denotes the prediction target. b) A neural forecasting model trained using the forking-sequences training-scheme. Forking-sequences collects multi-horizon errors from all intermediate forecast creation times. Additionally decod… view at source ↗
Figure 3
Figure 3. Figure 3: Specifically, increasing the batch size by [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Example time series from the M1 competition dataset. The train and validation sets consist of all observations preceding the first dotted line. The cross validation test set is comprised of a set of forecast creation dates with lengths defined in [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: a) Visualization of the covariance of M-dependent random variables. Green sections indicate correlated samples, while blue indicate uncorrelated variables. b) Mean estimator variance reduction as a function of the samples of the forking-sequences training scheme for different levels of M-dependence. In this Appendix, we provide a proof of the gradient variance reduction guarantees for the forking￾sequences… view at source ↗
Figure 6
Figure 6. Figure 6: Forking-sequences enables: a) faster convergence in loss on the train set because it b) reduces the variance of the stochastic gradient at a rate of O (1/|T |), with T the number of FCDS. In this example we report the training trajectories for a simple univariate autorregresive model on the M1 dataset. More details in Appendix E. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Convergence of quantile loss computed on the train set for Deep Learning models using either forking-sequences (solid) or window-sampling (dashed) techniques. Figures a, b, c, d, e, f show quantile loss versus train step for models trained across frequency-specific datasets. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: a) We adapt the forking-sequences inference into an ensembling technique by averaging forecasts generated prior to the prediction time. b) As shown in this sub-figure, the forking sequences ensembling approach reduces forecast variance, with a linear convergence rate analogous to the weak law of large numbers. G.1 Forecast Ensembling Techniques1970-01-01 1970-01-01 1970-01-01 1970-01-01 1970-01-01 2000 400… view at source ↗
Figure 9
Figure 9. Figure 9: Forecast outputs for a) no ensembling technique and various ensembling techniques applied during inference, including b) moving median, c) moving average, and d) cumulative average. We propose the use of ensembling techniques for model inference with the forking-sequences scheme which aggregate predictions for each target date across FCDs to reduce forecast variance. Fig. 9a shows model forecast examples f… view at source ↗
Figure 10
Figure 10. Figure 10: Percentage improvement in sCRPS and sQPC metrics for models with the forking-sequences scheme without ensembling during inference compared with the window-sampling scheme, averaged across datasets (a, b). Percentage improvement in sCRPS and sQPC metrics for models with the forking-sequences scheme with ensembling during inference compared with the window-sampling scheme, averaged across datasets (c, d). R… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. MICA: Multivariate Infini Compressive Attention for Time Series Forecasting

    cs.LG 2026-04 unverdicted novelty 6.0

    MICA adds linearly scaling compressive cross-channel attention to Transformers, cutting average forecast error by 5.4% and ranking first among multivariate baselines.

  2. MICA: Multivariate Infini Compressive Attention for Time Series Forecasting

    cs.LG 2026-04 unverdicted novelty 6.0

    MICA adapts infini compressive attention to the channel dimension, enabling scalable cross-channel dependencies in Transformers and cutting forecast error by 5.4% on average versus channel-independent baselines.

Reference graph

Works this paper leans on

52 extracted references · 4 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Maddix, Syama Rangapuram, David Salinas, Jasper Schulz, Lorenzo Stella, Ali Caner Tarkmen, and Yuyang Wang

    Alexander Alexandrov, Konstantinos Benidis, Michael Bohlke-Schneider, Valentin Flunkert, Jan Gasthaus, Tim Januschowski, Danielle C. Maddix, Syama Rangapuram, David Salinas, Jasper Schulz, Lorenzo Stella, Ali Caner Tarkmen, and Yuyang Wang. GluonTS: Probabilistic and neural time series modeling in python.Journal of Machine Learning Research, 21(116):1–6, 2020

    2020

  2. [2]

    Maddix, Michael W

    Abdul Fatir Ansari, Lorenzo Stella, Caner Turkmen, Xiyuan Zhang, Pedro Mercado, Huibin Shen, Oleksandr Shchur, Syama Sundar Rangapuram, Sebastian Pineda Arango, Shubham Kapoor, Jasper Zschiegner, Danielle C. Maddix, Michael W. Mahoney, Kari Torkkola, An- drew Gordon Wilson, Michael Bohlke-Schneider, and Yuyang Wang. Chronos: Learning the language of time ...

    2024

  3. [3]

    Hyndman, Haiyan Song, and Doris C

    George Athanasopoulos, Rob J. Hyndman, Haiyan Song, and Doris C. Wu. The Tourism forecasting competition.International Journal of Forecasting, 27(3):822–844, 2011. Special Section 1: Forecasting with Artificial Neural Networks and Computational Intelligence. Special Section 2: Tourism Forecasting

    2011

  4. [4]

    Neural machine translation by jointly learning to align and translate

    Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. In3rd International Conference on Learning Representations, ICLR 2015, 2016

    2015

  5. [5]

    PyTorchForecasting: Forecasting with neural networks made simple

    Jan Beitner. PyTorchForecasting: Forecasting with neural networks made simple. GitHub Repository, 2020

    2020

  6. [6]

    A neural probabilistic language model.Journal of Machine Learning Research, 3:1137–1155, 2003

    Yoshua Bengio, Réjean Ducharme, Pascal Vincent, and Christian Jauvin. A neural probabilistic language model.Journal of Machine Learning Research, 3:1137–1155, 2003

    2003

  7. [7]

    On the use of cross-validation for time series predictor evaluation.Information Sciences, 191:192–213, 2012

    Christoph Bergmeir and José M Benítez. On the use of cross-validation for time series predictor evaluation.Information Sciences, 191:192–213, 2012

    2012

  8. [8]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-V oss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott G...

    1901

  9. [9]

    Olivares, Boris N

    Cristian Challu, Kin G. Olivares, Boris N. Oreshkin, Federico Garza Ramirez, Max Mergenthaler-Canseco, and Artur Dubrawski. NHITS: neural hierarchical interpolation for time series forecasting. InProceedings of the Thirty-Seventh AAAI Conference on Arti- ficial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial In- telligenc...

    2023

  10. [10]

    Dilated recurrent neural networks

    Shiyu Chang, Yang Zhang, Wei Han, Mo Yu, Xiaoxiao Guo, Wei Tan, and et al. Dilated recurrent neural networks. In31st Conference on Neural Information Processing Systems, 2017

    2017

  11. [11]

    Chen, Lee Dicker, Carson Eisenach, and Dhruv Madeka

    Kevin C. Chen, Lee Dicker, Carson Eisenach, and Dhruv Madeka. MQTransformer: Multi- Horizon Forecasts with Context Dependent Attention and Optimal Bregman V olatility. In Maria Florina Balcan and Marina Meila, editors,In: Proceedings of 8th SIGKDD Interna- tional Workshop on Mining and Learning From Time Series - Deep Forecasting: Models, Interpretability...

    2022

  12. [12]

    Dai and Quoc V

    Andrew M. Dai and Quoc V . Le. Semi-supervised sequence learning. InAdvances in Neural Information Processing Systems, NIPS, 2015

    2015

  13. [13]

    A decoder-only foundation model for time-series forecasting, 2024

    Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou. A decoder-only foundation model for time-series forecasting, 2024

    2024

  14. [14]

    MQTransformer: Multi-Horizon Forecasts with Context Dependent and Feedback-Aware Attention

    Carson Eisenach, Yagna Patel, and Dhruv Madeka. MQTransformer: Multi-Horizon Forecasts with Context Dependent and Feedback-Aware Attention. In Maria Florina Balcan and Marina Meila, editors,Submitted to Proceedings of the 38th International Conference on Machine Learning. PMLR. Working Paper version available at arXiv:2009.14799, 8 2021

    Pith/arXiv arXiv 2009

  15. [15]

    Foster and Robert A

    Dean P. Foster and Robert A. Stine. Threshold martingales and the evolution of forecasts, 2021

    2021

  16. [16]

    Timegpt, 2023

    Azul Garza and Max Mergenthaler-Canseco. Timegpt, 2023

    2023

  17. [17]

    Olivares

    Federico Garza, Max Mergenthaler Canseco, Cristian Challú, and Kin G. Olivares. StatsForecast: Lightning fast forecasting with statistical and econometric models. PyCon Salt Lake City, Utah, US 2022, 2022

    2022

  18. [18]

    Strictly proper scoring rules, prediction, and estimation

    Tilmann Gneiting and Adrian E Raftery. Strictly proper scoring rules, prediction, and estimation. Journal of the American statistical Association, 102(477):359–378, 2007

    2007

  19. [19]

    Jeffrey D. Hart. Automated kernel smoothing of dependent data by using time series cross- validation.Journal of the Royal Statistical Society. Series B (Methodological), 56(3):529–542, 1994

    1994

  20. [20]

    HEATH and PETER L

    DA VID C. HEATH and PETER L. JACKSON. Modeling the evolution of demand forecasts ith application to safety stock analysis in production/distribution systems.IIE Transactions, 26(3):17–30, 1994

    1994

  21. [21]

    Darts: User-friendly modern machine learning for time series.Journal of Machine Learning Research, 23(124):1–6, 2022

    Julien Herzen, Francesco Lässig, Samuele Giuliano Piazzetta, Thomas Neuer, Léo Tafti, Guillaume Raille, Tomas Van Pottelbergh, Marek Pasieka, Andrzej Skrodzki, Nicolas Huguenin, Maxime Dumonal, Jan Kościsz, Dennis Bader, Frédérick Gusset, Mounir Benheddi, Camila Williamson, Michal Kosinski, Matej Petrik, and Gaël Grosch. Darts: User-friendly modern ...

    2022

  22. [22]

    Forecasting seasonals and trends by exponentially weighted moving averages

    Charles C Holt. Forecasting seasonals and trends by exponentially weighted moving averages. (O.N.R. Memorandum No. 52), 1957

    1957

  23. [23]

    Olivares.Forecasting: Principles and Practice, the Pythonic Way

    Rob J Hyndman, George Athanasopoulos, Azul Garza, Cristian Challu, Max Mergenthaler, and Kin G. Olivares.Forecasting: Principles and Practice, the Pythonic Way. OTexts, Melbourne, Australia, 2025. available at https://otexts.com/fpppy/. 8

    2025

  24. [24]

    Hyndman and Baki Billah

    Rob J. Hyndman and Baki Billah. Unmasking the theta method.International Journal of Forecasting, 19(2):287–290, 2003

    2003

  25. [25]

    Hyndman and Yeasmin Khandakar

    Rob J. Hyndman and Yeasmin Khandakar. Automatic time series forecasting: The forecast package for r.Journal of Statistical Software, Articles, 27(3):1–22, 2008

    2008

  26. [26]

    Hyndman and Anne B

    Rob J. Hyndman and Anne B. Koehler. Another look at measures of forecast accuracy.Interna- tional Journal of Forecasting, 22(4):679 – 688, 2006

    2006

  27. [27]

    Regression quantiles.Econometrica, 46(1):33–50, 1978

    Roger Koenker and Gilbert Bassett. Regression quantiles.Econometrica, 46(1):33–50, 1978

    1978

  28. [28]

    Deeply- Supervised Nets

    Chen-Yu Lee, Saining Xie, Patrick Gallagher, Zhengyou Zhang, and Zhuowen Tu. Deeply- Supervised Nets. In Guy Lebanon and S. V . N. Vishwanathan, editors,Proceedings of the Eighteenth International Conference on Artificial Intelligence and Statistics, volume 38 of Proceedings of Machine Learning Research, pages 562–570, San Diego, California, USA, 09–12 Ma...

    2015

  29. [29]

    Arık, Nicolas Loeff, and Tomas Pfister

    Bryan Lim, Sercan Ö. Arık, Nicolas Loeff, and Tomas Pfister. Temporal fusion transformers for interpretable multi-horizon time series forecasting.International Journal of Forecasting, 37(4):1748–1764, 2021

    2021

  30. [30]

    Kale, Charles Elkan, and Randall C

    Zachary Chase Lipton, David C. Kale, Charles Elkan, and Randall C. Wetzel. Learning to diagnose with LSTM recurrent neural networks. In Yoshua Bengio and Yann LeCun, editors, 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings, 2016

    2016

  31. [31]

    Makridakis, A

    S. Makridakis, A. Andersen, R. Carbone, R. Fildes, M. Hibon, R. Lewandowski, J. Newton, E. Parzen, and R. Winkler. The accuracy of extrapolation (time series) methods: Results of a forecasting competition.Journal of Forecasting, 1(2):111–153, 1982

    1982

  32. [32]

    The M3-competition: results, conclusions and implica- tions.International Journal of Forecasting, 16(4):451–476, 2000

    Spyros Makridakis and Michèle Hibon. The M3-competition: results, conclusions and implica- tions.International Journal of Forecasting, 16(4):451–476, 2000. The M3- Competition

    2000

  33. [33]

    The M4 competition: 100,000 time series and 61 forecasting methods.International Journal of Forecasting, 36(1):54– 74, 2020

    Spyros Makridakis, Evangelos Spiliotis, and Vassilios Assimakopoulos. The M4 competition: 100,000 time series and 61 forecasting methods.International Journal of Forecasting, 36(1):54– 74, 2020. M4 Competition

    2020

  34. [34]

    Marshall and Ingram Olkin

    Albert W. Marshall and Ingram Olkin. Multivariate Chebyshev Inequalities.The Annals of Mathematical Statistics, 31(4):1001 – 1014, 1960

    1960

  35. [35]

    Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam

    Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. A time series is worth 64 words: Long-term forecasting with transformers. InInternational Conference on Learning Representations, 2023

    2023

  36. [36]

    Olivares, Cristian Challú, Federico Garza, Max Mergenthaler Canseco, and Artur Dubrawski

    Kin G. Olivares, Cristian Challú, Federico Garza, Max Mergenthaler Canseco, and Artur Dubrawski. NeuralForecast: User friendly state-of-the-art neural forecasting models. PyCon Salt Lake City, Utah, US 2022, 2022

    2022

  37. [37]

    Olivares, Nganba Meetei, Ruijun Ma, Rohan Reddy, Mengfei Cao, and Lee Dicker

    Kin G. Olivares, Nganba Meetei, Ruijun Ma, Rohan Reddy, Mengfei Cao, and Lee Dicker. Probabilistic hierarchical forecasting with deep poisson mixtures.International Journal of Forecasting, accepted, Preprint version available at arXiv:2110.13179, 2023

    Pith/arXiv arXiv 2023

  38. [38]

    Olivares, Malcolm Wolff, Tatiana Konstantinova, Shankar Ramasubramanian, Boris Ore- shkin, Andrew Gordon Wilson, Andres Potapczynski, Willa Potosnak, Mengfei Cao, Michael W

    Kin G. Olivares, Malcolm Wolff, Tatiana Konstantinova, Shankar Ramasubramanian, Boris Ore- shkin, Andrew Gordon Wilson, Andres Potapczynski, Willa Potosnak, Mengfei Cao, Michael W. Mahoney, and Dmitry Efimov. A realistic evaluation of cross-frequency transfer learning and foundation forecasting models. InThirty-Ninth Annual Conference on Neural Informatio...

    2025

  39. [39]

    Oreshkin, Dmitri Carpov, Nicolas Chapados, and Yoshua Bengio

    Boris N. Oreshkin, Dmitri Carpov, Nicolas Chapados, and Yoshua Bengio. N-BEATS: neural ba- sis expansion analysis for interpretable time series forecasting. In8th International Conference on Learning Representations, ICLR 2020, 2020. 9

    2020

  40. [40]

    Improving language understanding by generative pre-training

    Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. https://cdn.openai.com/research-covers/ language-unsupervised/language_understanding_paper.pdf, 2018

    2018

  41. [41]

    DeepAR: Probabilistic forecasting with autoregressive recurrent networks.International Journal of Forecasting, 36(3):1181–1191, 2020

    David Salinas, Valentin Flunkert, Jan Gasthaus, and Tim Januschowski. DeepAR: Probabilistic forecasting with autoregressive recurrent networks.International Journal of Forecasting, 36(3):1181–1191, 2020

    2020

  42. [42]

    Investigating the accuracy of cross-learning time series forecasting methods

    Artemios-Anargyros Semenoglou, Evangelos Spiliotis, Spyros Makridakis, and Vassilios As- simakopoulos. Investigating the accuracy of cross-learning time series forecasting methods. International Journal of Forecasting, 37(3):1072–1084, 2021

    2021

  43. [43]

    Shumway and D.S

    R.H. Shumway and D.S. Stoffer.Time Series Analysis and Its Applications: With R Examples. Springer Texts in Statistics. Springer International Publishing, 2017

    2017

  44. [44]

    A hybrid method of exponential smoothing and recurrent neural networks for time series forecasting.International Journal of Forecasting, 07 2019

    Slawek Smyl. A hybrid method of exponential smoothing and recurrent neural networks for time series forecasting.International Journal of Forecasting, 07 2019

    2019

  45. [45]

    On the categorization of demand patterns.Journal of the Operational Research Society, 56, 05 2005

    M Syntetos, John Boylan, and JD Croston. On the categorization of demand patterns.Journal of the Operational Research Society, 56, 05 2005

    2005

  46. [46]

    Beril Toktay and Lawrence M

    L. Beril Toktay and Lawrence M. Wein. Analysis of a forecasting-production-inventory system with stationary demand.Management Science, 47(9):1268–1281, 2001

    2001

  47. [47]

    Senior, and Koray Kavukcuoglu

    Aäron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew W. Senior, and Koray Kavukcuoglu. WaveNet: A generative model for raw audio.Computer Research Repository, abs/1609.03499, 2016

    Pith/arXiv arXiv 2016

  48. [48]

    A Multi- horizon Quantile Recurrent Forecaster

    Ruofeng Wen, Kari Torkkola, Balakrishnan Narayanaswamy, and Dhruv Madeka. A Multi- horizon Quantile Recurrent Forecaster. In31st Conference on Neural Information Processing Systems NIPS 2017, Time Series Workshop, 2017

    2017

  49. [49]

    Olivares, Boris Oreshkin, Sunny Ruan, Sitan Yang, Abhinav Katoch, Shankar Ramasubramanian, Youxin Zhang, Michael W

    Malcolm Wolff, Kin G. Olivares, Boris Oreshkin, Sunny Ruan, Sitan Yang, Abhinav Katoch, Shankar Ramasubramanian, Youxin Zhang, Michael W. Mahoney, Dmitry Efimov, and Vincent Quenneville-Bélair. ♠ SPADE ♠: Split Peak Attention DEcomposition. InThirty-Eighth Annual Conference on Neural Information Processing Systems NeurIPS 2024, volume Time Series in the A...

    2024

  50. [50]

    Unified training of universal time series forecasting transformers, 2024

    Gerald Woo, Chenghao Liu, Akshat Kumar, Caiming Xiong, Silvio Savarese, and Doyen Sahoo. Unified training of universal time series forecasting transformers, 2024

    2024

  51. [51]

    Time Series Library (TSLib)

    Haixu Wu, Yong Liu, Huikun Weng, Yuxuan Wang, Tengge Hu, Haoran Zhang, and Jiawei Guo. Time Series Library (TSLib). GitHub Repository, 2023

    2023

  52. [52]

    Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting

    Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting. The Association for the Advancement of Artificial Intelligence Conference 2021 (AAAI 2021)., abs/2012.07436, 2020. 10 A Related Work Similar ideas to forking-sequences have been in...

    Pith/arXiv arXiv 2021