Non-collective Calibrating Strategy for Time Series Forecasting

Bin Wang; Feng Hong; Junbo Zhang; Minbo Ma; Tianrui Li; Yanwei Yu; Yongqi Han

REVIEW 4 major objections 5 minor 38 references

Non-collective Calibrating Strategy for Time Series Forecasting

T0 review · 4 major / 5 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 Leave the trained forecaster frozen, attach a tiny per-target recalibration plug with its own optimizer and early-stopping rule, and forecasting error drops by up to roughly 22% across seven benchmark models and a weather-forecasting task.

desk verdict SoP is a practical calibration recipe that likely works, but the MTLC mechanism claim outruns the evidence. read the letter →

arxiv 2506.03176 v2 pith:4AG2BMT3 submitted 2025-05-29 cs.LG

Bin Wang , Yongqi Han , Minbo Ma , Tianrui Li , Junbo Zhang , Feng Hong , Yanwei Yu This is my paper

classification cs.LG MSC 68T0762M10

keywords timeseriesforecastingmodelcalibrationnon-collectivecalibratingmulti-targetlearningconflictearlystoppingSocket+Plugstrategymodel-agnosticenhancementspatio-temporal

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 cheapest reliable way to improve a trained deep forecasting model is not to design a new architecture but to freeze the model and attach a small, independent recalibration module to each prediction target. Its central discovery is a failure mode it names multi-target learning conflict: when one post-processing module is trained for all variables and horizons at once, the single early-stopping monitor stops the whole run as soon as the joint validation loss plateaus, which is too early for the targets that are harder to learn. The proposed Socket+Plug (SoP) strategy gives each target its own optimizer and its own early-stopping monitor, so each variable or horizon is calibrated for exactly as long as it needs. Across seven benchmark models and datasets, and on a spatio-temporal meteorological forecasting task, SoP lowers mean squared error consistently, by up to roughly 22%, with the largest gains on simple models and on datasets whose variables are weakly correlated. If the paper is right, the practical route to better forecasts is often a per-target post-processing layer over the model you already have, not a new model.

What carries the argument

The carrying object is the Socket+Plug decomposition. The Socket is the already-trained forecaster, kept frozen so that no weight of the original model changes. The Plug is a per-target calibrator: for variable-wise SoP, plug $i$ takes the Socket's $S$-length prediction for variable $i$, applies LayerNorm, projects through a two-hidden-layer MLP with hidden dimension 256 and GELU activation, and combines the result with the normalized prediction by element-wise multiplication, $\bar{Y}_{i,:}=\operatorname{MLP}_i(\hat{Y}_{i,:})\odot\operatorname{LayerNorm}(\hat{Y}_{i,:})$, so the plug is a learned multiplicative correction per series. What makes the strategy non-collective is that each plug $i$ carries its own optimizer $opt_i$ and its own early-stopping monitor on its own validation loss $L^{\mathrm{val}}_i$; because the total gradient is the sum of independent per-plug gradients, the plugs train in parallel and stop at different epochs, which the paper presents as the signature of resolved multi-target learning conflict. A plug-count hyperparameter $M$ interpolates between variable-wise SoP ($M=N$), step-wise SoP ($M=S$), and the collective strategy ($M=1$), and the ablation over $M$ is the experiment that ties the performance gap to the shared monitor.

What would settle it

Run one trained Socket through four calibrators on the same data: (1) full SoP; (2) collective training with a shared optimizer and shared monitor and matched total plug capacity; (3) plugs with a shared optimizer but a separate early-stopping monitor per target; and (4) per-target plugs with separate optimizers but a fixed, generous training budget and no early stopping at all. If MTLC is the mechanism, setup (3) should recover most of SoP's gain and the per-target stopping epochs should spread widely; if setups (3) and (4) show no advantage over collective training, the shared monitor is not the bottleneck. A sharper check: on a univariate, single-horizon forecasting task there is no multi-target conflict, so any SoP improvement there would show that something other than MTLC is driving the gains.

Watch

Extended reading notes

Core claim

The paper's claim, stated as a universal calibrating strategy, is that a fully trained forecasting backbone — the Socket — should be left frozen, and that its predictions should be corrected by many small per-target networks — the Plugs — each trained, validated, and early-stopped on its own target alone. The motivating observation is that the classical alternative, a single collective calibrator trained on the total loss over all variables and horizons, triggers early stopping prematurely (around epoch ten in the authors' illustration) because the easy variables dominate the joint validation curve, leaving hard variables undertrained; the paper calls this multi-target learning conflict. In SoP, each plug is a small MLP that reads one variable's predicted sequence, normalizes it, projects it through two hidden layers, and multiplies the result back onto the normalized prediction as a residual gate, so the plug learns a per-series multiplicative correction. The experimental claim is that non-collective calibration improves every Socket tested — DLinear, TSMixer, SOFTS, iTransformer, PatchTST, TimesNet, and FEDformer — with average gains about twice those of collective calibration, and with single gains up to roughly 22% (e.g., 22.9% MSE for iTransformer on Exchange), while collective calibration with a single plug can actively degrade a trained model. The same recipe, applied to a Unet Socket on the ERA5 weather dataset, reduces error on all five meteorological variables and across all twenty prediction horizons.

Load-bearing premise

The load-bearing premise is that the performance gap really comes from the collective early-stopping monitor stopping each target too soon, and not from the separate per-target optimizers, the LayerNorm-plus-residual-gate plug architecture, or other incidental differences between the two setups.

Editorial extensions

If this is right

Any trained deep forecaster can be upgraded by SoP without retraining or architecture changes: all seven Socket models improve on average, and the largest single gain is 22.9% MSE for iTransformer on Exchange.
Non-collective calibration roughly doubles the average improvement of collective calibration, and collective calibration (one plug for all targets) can push a model below its uncalibrated accuracy, which the paper reads as direct evidence for multi-target learning conflict.
Simpler architectures gain the most: DLinear+Plug beats TimesNet+Plug on most datasets, so calibration can substitute for architectural complexity.
SoP carries over to spatio-temporal forecasting: Unet+Plug lowers MSE on all five ERA5 variables and at every one of the 20 horizons, producing less over-smoothed fields.
With only a small MLP Plug, iTransformer+Plug matches or beats LLM-based forecasters on Traffic and Weather at a fraction of their cost.

Reading between the lines

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

If MTLC is the operative cause, the same non-collective principle should transfer to any multi-head learner that early-stops on a shared validation loss — multi-task networks, multi-class calibration layers, ensemble heads — making per-target monitors a cheap universal wrapper beyond forecasting.
Because plugs train only on frozen Socket outputs, plug fitting is supervised regression on fixed features; in a deployment one could fit plugs on a Socket's stored validation predictions, getting calibration without rerunning the training pipeline.
The per-plug early-stopping spread (some plugs stop within a few epochs while others train much longer, e.g., Var4 in the Exchange runs) doubles as a per-target learnability score, pointing at which variables the Socket models worst and where extra capacity should go.
The correlation finding — variable-wise SoP helps most on low-inter-variable-correlation datasets like Exchange — suggests that a cheap correlation statistic could pick between variable-wise, step-wise, and grouped plugs automatically, without validation sweeps over plug counts.

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.

Desk Editor's Note

read the letter

Quick take: this is a solid empirical paper with a mechanistic explanation that outruns its evidence. The recipe — freeze the trained forecaster, attach a small MLP per predicted target with its own optimizer and early-stopping monitor — is simple, model-agnostic, and backed by experiments across seven benchmarks, seven socket architectures, a spatio-temporal ERA5 case, and released code. I'd bet the performance gains are real: the improvements are consistent, sometimes large (up to ~22% MSE reduction), and the appendix includes a five-seed robustness check with small standard deviations. The ablation against collective calibration, the plug-count sweep, and the plug-transferability test are all genuinely useful. This paper deserves referee time.

Now the soft spots. The central claim, MTLC, is not directly tested. Section 4.4 varies per-target optimizers and per-target early stopping together, so you cannot attribute the gain to either factor. Appendix A compounds the problem: with independent plugs, the total gradient is a sum of independent per-plug gradients, so a shared optimizer does not create cross-variable gradient coupling in their independent-plug ablation. The collective condition differs mainly by using one global validation signal. The benefit could therefore come entirely from per-variable model selection — stopping each plug at its own best epoch — rather than from resolving a genuine learning conflict. Figure A4 only shows that different variables stop at different epochs, not that the collective stop point is suboptimal for each variable. Figure 5 adds another confound: changing plug counts changes the MLP connectivity, not just the grouping, even when total parameter count is matched. The MTLC label is an interpretation, not a demonstrated mechanism.

The main results table reports point estimates with no error bars or significance tests, and the plug-count exploration was done on the test set. The abstract also says SoP improves “any trained deep forecasting models,” which overreaches — seven architectures is a good start but not universality. None of this kills the empirical contribution, but the paper should either soften the mechanism language or run a cleaner test: for example, a collective optimizer with per-variable validation-based model selection, or per-target optimizers with a shared monitor.

Who is this for? Practitioners who want a cheap post-hoc improvement to a deployed forecasting model, and researchers working on calibration or multi-task training. A serious editor should send this to peer review, expecting a revision that decouples the mechanism and adds statistical rigor to the headline numbers.

Referee Report

4 major / 5 minor

Summary. The paper proposes Socket+Plug (SoP), a post-hoc calibration strategy for trained time series forecasting models. The Socket is a frozen pretrained forecasting model; each Plug is a small MLP that recalibrates predictions for a group of variables or horizons, with its own optimizer and early-stopping monitor. The authors argue that the conventional collective approach, in which all targets are calibrated jointly under a single loss, suffers from a multi-target learning conflict (MTLC) that leads to premature early stopping and under-training; SoP mitigates this by keeping per-target optimizers and monitors. The central claims are that SoP is model-agnostic, improves a wide range of deep forecasting models on seven time series benchmarks and the ERA5 spatio-temporal dataset, and achieves up to roughly 22% MSE improvement with a simple MLP Plug. Code is provided, and the paper includes ablations on plug counts, comparisons against collective calibration, robustness checks with multiple seeds, and a transferability experiment.

Significance. If the empirical claims hold, SoP is a useful, lightweight, architecture-agnostic post-processing layer that can be applied to existing trained forecasting models, avoiding retraining from scratch. The paper's strengths include broad coverage across seven benchmark datasets and seven inference models, an explicit ablation against collective calibration, robustness and transferability experiments that give some confidence in the recipe, and the release of code. The conceptual contribution hinges on the MTLC mechanism, which is plausible but not directly tested. The paper's significance is therefore conditional on a cleaner isolation of the proposed non-collective mechanism from per-variable early-stopping and from architectural differences in the ablations.

major comments (4)

[Section 4.4 and Appendix B.7 (Table 2, Figure A3, Figure A4)] The ablation that is presented as supporting the MTLC mechanism compares non-collective calibrating (each Plug i has its own optimizer opt_i and its own monitor Monitor_i on Lval_i) against collective calibrating (all Plugs share a single optimizer and a single monitor on the total validation loss). These two conditions differ in two factors simultaneously: the optimizer grouping and the early-stopping/validation-signal grouping. The observed gains can therefore be fully explained by per-variable validation-based model selection (choosing each Plug's best epoch) rather than by a reduction of cross-target learning conflict. To attribute the gains to the proposed non-collective mechanism, the authors should include at least a 2x2 ablation: (a) shared optimizer with per-variable monitors, and (b) per-plug optimizers with a collective monitor. Without such a control, the central conceptual claim that MTLC, not per-target early stopping, drives the improvement is unsupported.
[Appendix A] The gradient analysis in Appendix A states that in the independent-plug setting the total gradient is ∇L = Σ_i ∇L_i(θ_i), i.e., each Plug's parameters are disjoint and the gradient for Plug i does not depend on any other Plug. This means that in the collective ablation of Section 4.4, where there are still N independent Plugs but a shared optimizer and a single monitor on the total loss, the shared optimizer does not create cross-variable gradient coupling. The collective condition differs from the non-collective condition mainly in the validation signal (one global monitor versus N per-variable monitors), not in gradient interference. The paper should acknowledge this limitation and provide an experiment that holds the optimizer type fixed while varying only the early-stopping monitors; otherwise the MTLC explanation remains a post-hoc interpretation of the observed different stopping epochs in Figure A4 rather than a demonstrated cause.
[Figure 5 and Appendix B.6] The plug-count ablation in Section 4.3 and Figure 5 changes the architecture of the calibrating module: plug count 1 uses a single wide MLP whose hidden width is scaled to match the total parameter count of M independent smaller MLPs, while variable-wise SoP uses M independent MLPs. This confounds the grouping of targets with the connectivity and function class of the calibrating network. The conclusion that collective calibrating performs worst 'because of MTLC' could instead reflect the representational limitations of a single shared MLP compared with independent per-variable MLPs. To support the MTLC claim, the authors should compare collective and non-collective grouping under the same connectivity pattern (e.g., a single MLP with per-variable output heads that are optimized collectively versus independently, with the same architecture in both cases).
[Table 1 and Table A2] The main results table reports point estimates without error bars or significance tests. Several advertised improvements are near zero or negative (e.g., SOFTS on Exchange MAE is -0.820%, TimesNet on Weather MSE is 0.019%, PatchTST on ETTh1 MAE is -0.735%). Appendix Table A2 shows that repeated runs with different seeds produce non-negligible variation (e.g., MSE std up to 0.03 for the Exchange S=720 case). Given this variation, the claim that SoP reliably enhances a broad class of models requires either confidence intervals over at least 3-5 seeds for all Table 1 entries or paired statistical tests. This is a reliability issue for the headline 'up to 22% improvement' claim, not just a presentation detail.

minor comments (5)

[Section 3.1, Eq. (4)] Equation (4) defines Plugi(ˆYi,:) = MLPi(ˆYi,:) ⊙ Vi,:, where Vi,: = LayerNorm(ˆYi,:). The text preceding the equation says the normalized values are 'processed by a Plug', but the equation shows the MLP operating on the un-normalized prediction and then multiplying by the normalized values. Please clarify whether the MLP input is the raw Socket output or the layer-normalized output, since this affects the interpretation of the calibration transformation.
[Section 4.2] The text says 'DLinear underperforms compared to TimeNet'; this should read 'TimesNet' (the model name used everywhere else).
[Algorithm 1] The algorithm appears to describe sequential training of N Plugs in a single for-loop, each with a fresh pass over the training data. Since the Plugs are independent (as stated in Appendix A), the algorithm should clarify that the N Plugs can be trained in parallel, and that the loop is only for exposition.
[Figure A4] The panels in Figure A4 use bars for 'Collective' and 'Non-collective' but do not include a legend or distinct labels in each panel; please add a shared legend or a distinguishing pattern so that the two conditions are readable in grayscale.
[Figure 2] The banner text 'MTLC leads to premature early stops making training inadequate' is stated as fact. Since the paper later treats this as a hypothesis to be tested by the ablation, the banner should be phrased as a hypothesis or a proposed explanation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SoP is an empirical calibration recipe, not a derived prediction; MTLC is an interpretive hypothesis.

full rationale

The paper does not present a formal derivation in which an output is defined in terms of fitted parameters or in which a prediction reduces by construction to its inputs. SoP is an empirical post-processing strategy: a frozen Socket feeds per-target Plugs trained with MSE and early-stopping monitors on validation losses, and the reported improvements are benchmark comparisons against the uncalibrated Socket on held-out test sets. Model selection by validation loss is standard practice, not a circular ‘prediction’ of the fitted quantity. The MTLC phenomenon is introduced as a heuristic hypothesis motivated by Figure 2 and Appendix Figure A4, and the ablation in Section 4.4 compares non-collective versus collective deployment; whether the observed gains come from per-target optimizers, per-target monitors, or the LayerNorm/residual gate is an experimental attribution question, not a circularity. No load-bearing self-citations appear: the cited inference models, libraries, and normalization works are external. The paper explicitly labels the MTLC account as a hypothesis and supports it with experiments, so the central empirical claim remains independent of the interpretive label. Therefore no circular step is exhibited, and the appropriate score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The paper is an empirical calibration study, so the reader pays for a set of hand-set hyperparameters and domain assumptions. The central claim (SoP improves forecasts) depends on the Socket being informative, on the validation protocol being trustworthy, and on the causal interpretation of the comparison (MTLC) being accepted. There are no physical entities or hidden free constants beyond these modeling choices.

free parameters (4)

Plug hidden dimension d = 256
Chosen for all Plug MLPs and reported in Appendix B.3; no sensitivity analysis is provided.
Early-stopping patience = 5 epochs (10 for Unet)
Set once and applied to every Plug monitor; affects when snapshots are taken and therefore the final result.
Plug optimizer learning rate = 1e-4
Adam learning rate used for all Plugs on time series benchmarks; Appendix C.2 reports a different value for Unet but no tuning study.
Plug counts M in ablation sweeps = Variable view {21,7,3,1}; step view varying per horizon
Figures 5 and 6 explore several plug counts and identify the best-performing value on the test set; the paper then recommends variable-wise and step-wise SoP as quick starts, which is a post-hoc selection.

assumptions (4)

domain assumption MSE is the appropriate loss for calibration in both collective and non-collective settings
MSE is used throughout Algorithm 1 and all experiments, but the paper does not justify it against other losses such as MAE or distributional losses.
domain assumption The frozen Socket outputs contain sufficient information for a small MLP to improve forecasts
All Plugs operate only on the Socket's predictions; if the Socket features are poor, calibration cannot help. This is implicit in Section 3.1.
domain assumption Validation-based early stopping with fixed patience transfers to test performance
Each Plug is snapshotted on a single validation split with patience 5, and the paper uses that snapshot on the test set without cross-validation (Algorithm 1, Appendix B.3).
standard math Standard backpropagation and independent loss gradients are sufficient for the claimed non-interference
Appendix A uses the independence of gradients across Plugs to justify parallel training and the absence of conflicting updates, which relies on standard automatic differentiation.

invented entities (1)

Multi-target learning conflict (MTLC)
purpose: Hypothesized mechanism explaining why collective calibration underperforms non-collective calibration
MTLC is inferred from early-stopping epoch distributions (Figure 2 and Figure A4); no direct measurement or controlled manipulation is provided, and it is not given a falsifiable quantitative signature beyond the observed performance gap.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Non-collective Calibrating Strategy for Time Series Forecasting." pith.science (2026). https://pith.science/paper/4AG2BMT3

@misc{pith2026250603176,
  author       = {Pith},
  title        = {Pith review of: Non-collective Calibrating Strategy for Time Series Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4AG2BMT3}},
  note         = {Machine review of arXiv:2506.03176}
}

read the original abstract

Deep learning-based approaches have demonstrated significant advancements in time series forecasting. Despite these ongoing developments, the complex dynamics of time series make it challenging to establish the rule of thumb for designing the golden model architecture. In this study, we argue that refining existing advanced models through a universal calibrating strategy can deliver substantial benefits with minimal resource costs, as opposed to elaborating and training a new model from scratch. We first identify a multi-target learning conflict in the calibrating process, which arises when optimizing variables across time steps, leading to the underutilization of the model's learning capabilities. To address this issue, we propose an innovative calibrating strategy called Socket+Plug (SoP). This approach retains an exclusive optimizer and early-stopping monitor for each predicted target within each Plug while keeping the fully trained Socket backbone frozen. The model-agnostic nature of SoP allows it to directly calibrate the performance of any trained deep forecasting models, regardless of their specific architectures. Extensive experiments on various time series benchmarks and a spatio-temporal meteorological ERA5 dataset demonstrate the effectiveness of SoP, achieving up to a 22% improvement even when employing a simple MLP as the Plug (highlighted in Figure 1). Code is available at https://github.com/hanyuki23/SoP.

Figures

Figures reproduced from arXiv: 2506.03176 by the authors.

**Figure 2.** This banner highlights the distinct effects between SoP and the classical collective calibrating strategy. The MTLC phenomenon [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** The framework of the proposed SoP. On the left are Socket, such as fully trained models like iTransformer, DLinear, etc., which [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Performance of SoP at different prediction horizons. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: Effect of plug counts on the performance of SoP from the [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Effect of plug counts on the performance of SoP from the [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

**Figure 8.** Figure 8: Visualization of a spatio-temporal forecasting case study [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

**Figure 7.** Figure 7: (a) The averaged MSE for each of five variables. (b) The [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 32 canonical work pages

[1]

The First World Sci- ence and Intelligence Competition: Atmospheric Science Track - East China AI Medium-Range Weather Forecast- ing Competition

[Alibaba Group, 2023] Alibaba Group. The First World Sci- ence and Intelligence Competition: Atmospheric Science Track - East China AI Medium-Range Weather Forecast- ing Competition. https://tianchi.aliyun.com/competition/ entrance/532111/information,

work page 2023
[5]

Conditional time series fore- casting with convolutional neural networks

[Borovykh et al., 2017] Anastasia Borovykh, Sander Bohte, and Cornelis W Oosterlee. Conditional time series fore- casting with convolutional neural networks. arXiv,

work page 2017
[6]

Arik, Nathanael C

[Chen et al., 2023b] Si-An Chen, Chun-Liang Li, Sercan ¨O. Arik, Nathanael C. Yoder, and Tomas Pfister. TSMixer: An all-mlp architecture for time series forecast-ing. Trans. Mach. Learn. Res., 2023,

work page 2023
[7]

A multi-view multi-task learning framework for multi-variate time series forecast- ing

[Deng et al., 2022] Jinliang Deng, Xiusi Chen, Renhe Jiang, Xuan Song, and Ivor W Tsang. A multi-view multi-task learning framework for multi-variate time series forecast- ing. IEEE Transactions on Knowledge and Data Engineer- ing, 35(8):7665–7680,

work page 2022
[8]

SOFTS: efficient multivariate time series forecasting with series-core fusion

[Han et al., 2024] Lu Han, Xu-Yang Chen, Han-Jia Ye, and De-Chuan Zhan. SOFTS: efficient multivariate time series forecasting with series-core fusion. CoRR, abs/2404.14197,

arXiv 2024
[9]

Gaussian error linear units (gelus)

[Hendrycks and Gimpel, 2016] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415,

arXiv 2016
[10]

Prediction then cor- rection: An abductive prediction correction method for se- quential recommendation

[Huang et al., 2023] Yulong Huang, Yang Zhang, Qifan Wang, Chenxu Wang, and Fuli Feng. Prediction then cor- rection: An abductive prediction correction method for se- quential recommendation. In SIGIR, pages 2272–2276. ACM,

work page 2023
[11]

Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, and Qingsong Wen

[Jin et al., 2024] Ming Jin, Shiyu Wang, Lintao Ma, Zhix- uan Chu, James Y . Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, and Qingsong Wen. Time-LLM: Time series forecasting by reprogram- ming large language models. In ICLR. OpenReview.net,

work page 2024

Show all 38 references

[12]

Accurate uncertainties for deep learning using calibrated regression

[Kuleshov et al., 2018] V olodymyr Kuleshov, Nathan Fen- ner, and Stefano Ermon. Accurate uncertainties for deep learning using calibrated regression. In ICML, volume 80 of Proceedings of Machine Learning Research , pages 2801–2809. PMLR,

2018
[14]

Evaluation of near- surface wind speed climatology and long-term trend over china’s mainland region based on ERA5 reanalysis

[Liu et al., 2021] H Liu, L Dong, R Yan, X Zhang, C Guo, S Liang, J Tu, X Feng, and X Wang. Evaluation of near- surface wind speed climatology and long-term trend over china’s mainland region based on ERA5 reanalysis. Cli- matic and Environmental Research, 26(3):299–311,

2021
[16]

Unitime: A language-empowered unified model for cross-domain time series forecasting

[Liu et al., 2024b] Xu Liu, Junfeng Hu, Yuan Li, Shizhe Diao, Yuxuan Liang, Bryan Hooi, and Roger Zimmer- mann. Unitime: A language-empowered unified model for cross-domain time series forecasting. In Proceedings of the ACM Web Conference 2024, pages 4095–4106,

2024
[17]

Structural property-aware multilayer net- work embedding for latent factor analysis

[Lu et al., 2018] Jie Lu, Junyu Xuan, Guangquan Zhang, and Xiangfeng Luo. Structural property-aware multilayer net- work embedding for latent factor analysis. Pattern Recog- nition, 76:228–241,

2018
[18]

Fuzzy multiple-source transfer learning

[Lu et al., 2019] Jie Lu, Hua Zuo, and Guangquan Zhang. Fuzzy multiple-source transfer learning. IEEE Transac- tions on Fuzzy Systems, 28(12):3418–3431,

2019
[20]

Probabilistic outputs for support vector machines and comparisons to regular- ized likelihood methods

[Platt and others, 1999] John Platt et al. Probabilistic outputs for support vector machines and comparisons to regular- ized likelihood methods. Advances in large margin classi- fiers, 10(3):61–74,

1999
[23]

U-Net: Convolutional networks for biomedical image segmentation

[Ronneberger et al., 2015] Olaf Ronneberger, Philipp Fis- cher, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, volume 9351 of Lecture Notes in Computer Science , pages 234–

2015
[25]

Are lan- guage models actually useful for time series forecasting? In NeurIPS,

[Tan et al., 2024] Mingtian Tan, Mike A Merrill, Vinayak Gupta, Tim Althoff, and Thomas Hartvigsen. Are lan- guage models actually useful for time series forecasting? In NeurIPS,

2024
[26]

Deep time series models: A comprehensive survey and benchmark

[Wang et al., 2024] Yuxuan Wang, Haixu Wu, Jiaxiang Dong, Yong Liu, Mingsheng Long, and Jianmin Wang. Deep time series models: A comprehensive survey and benchmark. CoRR, abs/2407.13278,

2024 arXiv
[27]

Diffstg: Probabilistic spatio-temporal graph forecasting with denoising diffusion models

[Wen et al., 2023] Haomin Wen, Youfang Lin, Yutong Xia, Huaiyu Wan, Qingsong Wen, Roger Zimmermann, and Yuxuan Liang. Diffstg: Probabilistic spatio-temporal graph forecasting with denoising diffusion models. InPro- ceedings of the 31st ACM International Conference on Ad- vance...

2023
[28]

TimesNet: Temporal 2d-variation modeling for general time series analysis

[Wu et al., 2023] Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Mingsheng Long. TimesNet: Temporal 2d-variation modeling for general time series analysis. In ICLR. OpenReview.net,

2023
[29]

Are transformers effective for time series fore- casting? Proceedings of the AAAI Conference on Artificial Intelligence, 37(9):11121–11128, Jun

[Zeng et al., 2023] Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. Are transformers effective for time series fore- casting? Proceedings of the AAAI Conference on Artificial Intelligence, 37(9):11121–11128, Jun

2023
[30]

MAP-FCRNN: multi-step ahead prediction model using forecasting correction and RNN model with memory functions.Inf

[Zhang et al., 2023] Rongtao Zhang, Xueling Ma, Weiping Ding, and Jianming Zhan. MAP-FCRNN: multi-step ahead prediction model using forecasting correction and RNN model with memory functions.Inf. Sci., 646:119382,

2023
[31]

ProbTS: Benchmarking point and distributional forecasting across diverse prediction horizons

[Zhang et al., 2024] Jiawen Zhang, Xumeng Wen, Zhenwei Zhang, Shun Zheng, Jia Li, and Jiang Bian. ProbTS: Benchmarking point and distributional forecasting across diverse prediction horizons. In NeurIPS,

2024
[32]

Time-vlm: Exploring multimodal vision-language models for aug- mented time series forecasting

[Zhong et al., 2025] Siru Zhong, Weilin Ruan, Ming Jin, Huan Li, Qingsong Wen, and Yuxuan Liang. Time-vlm: Exploring multimodal vision-language models for aug- mented time series forecasting. InProceedings of the 42nd International Conference on Machine Learning,

2025
[34]

Appendix A Analysis of Efficient Parallel Training in SoP In practice, the computational cost and training time of the non-collective, variable-wise SoP strategy remain efficient compared to the collective calibration. As each of theN Plugs calibrates its corresponding target ...

2016
[35]

• Electricity: the electricity dataset contains the hourly electricity consumption of 321 customers from 2012 to

2012
[37]

• Solar-energy: the solar-energy dataset captures the solar power production of 137 photovoltaic (PV) plants in 2006, sam- pled at 10-minute intervals. B.2 Inference models Below is a brief overview of the inference models selected for our study: • SOFTS: it employs the centra...

2006
[720]

All experiments are conducted on a single NVIDIA GeForce RTX 4090 GPU with 24 GB VRAM. B.4 Performance of SoP at different prediction horizons To provide more detailed experimental results, we also evaluated the Weather and Traffic datasets, demonstrating that a Plug with cons...
[1999]

[Platt, 2000] John C. Platt. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in Large Margin Classifiers,

2000
[2000]

Jensen, Zhenli Sheng, and Bin Yang

[Qiu et al., 2024] Xiangfei Qiu, Jilin Hu, Lekui Zhou, Xingjian Wu, Junyang Du, Buang Zhang, Chenjuan Guo, Aoying Zhou, Christian S. Jensen, Zhenli Sheng, and Bin Yang. TFB: towards comprehensive and fair bench- marking of time series forecasting methods. VLDB, 17(9):2363–2377,

2024
[2015]

Predict-then-Calibrate: A New Perspective of Robust Contextual LP

[Sun et al., 2023] Chunlin Sun, Linyu Liu, and Xiaocheng Li. Predict-then-Calibrate: A New Perspective of Robust Contextual LP. In NeurIPS,

2023
[2016]

Classifier calibration with roc-regularized isotonic regression

[Berta et al., 2024] Eugene Berta, Francis Bach, and Michael Jordan. Classifier calibration with roc-regularized isotonic regression. In International Conference on Artifi- cial Intelligence and Statistics, pages 1972–1980. PMLR,

2024
[2018]

Modeling long- and short-term temporal patterns with deep neural networks

[Lai et al., 2018] Guokun Lai, Wei-Cheng Chang, Yiming Yang, and Hanxiao Liu. Modeling long- and short-term temporal patterns with deep neural networks. In SIGIR, pages 95–104. ACM,

2018
[2019]

Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam

[Nie et al., 2023] Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. A time series is worth 64 words: Long-term forecasting with transformers. In ICLR,

2023
[2020]

• Exchange: the exchange dataset contains panel data of daily exchange rates from 8 countries, spanning the period from 1990 to

1990
[2021]

Calf: Aligning llms for time series forecasting via cross-modal fine-tuning

[Liu et al., 2024a] P Liu, H Guo, T Dai, N Li, J Bao, X Ren, Y Jiang, and ST Xia. Calf: Aligning llms for time series forecasting via cross-modal fine-tuning. arXiv preprint arXiv:2403.07300,

arXiv
[2023]

[Ba et al., 2016] Lei Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. CoRR, abs/1607.06450,

2016 arXiv
[2024]

Accurate medium- range global weather forecasting with 3d neural networks

[Bi et al., 2023] Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian. Accurate medium- range global weather forecasting with 3d neural networks. Nat., 619(7970):533–538,

2023
[2025]

FEDformer: Fre- quency enhanced decomposed transformer for long-term series forecasting

[Zhou et al., 2022] Tian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin. FEDformer: Fre- quency enhanced decomposed transformer for long-term series forecasting. In ICML, volume 162 ofProceedings of Machine Learning Research, pages 27268–27286. PMLR,

2022

Pith tools

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

[1] [1]

The First World Sci- ence and Intelligence Competition: Atmospheric Science Track - East China AI Medium-Range Weather Forecast- ing Competition

[Alibaba Group, 2023] Alibaba Group. The First World Sci- ence and Intelligence Competition: Atmospheric Science Track - East China AI Medium-Range Weather Forecast- ing Competition. https://tianchi.aliyun.com/competition/ entrance/532111/information,

work page 2023

[2] [5]

Conditional time series fore- casting with convolutional neural networks

[Borovykh et al., 2017] Anastasia Borovykh, Sander Bohte, and Cornelis W Oosterlee. Conditional time series fore- casting with convolutional neural networks. arXiv,

work page 2017

[3] [6]

Arik, Nathanael C

[Chen et al., 2023b] Si-An Chen, Chun-Liang Li, Sercan ¨O. Arik, Nathanael C. Yoder, and Tomas Pfister. TSMixer: An all-mlp architecture for time series forecast-ing. Trans. Mach. Learn. Res., 2023,

work page 2023

[4] [7]

A multi-view multi-task learning framework for multi-variate time series forecast- ing

[Deng et al., 2022] Jinliang Deng, Xiusi Chen, Renhe Jiang, Xuan Song, and Ivor W Tsang. A multi-view multi-task learning framework for multi-variate time series forecast- ing. IEEE Transactions on Knowledge and Data Engineer- ing, 35(8):7665–7680,

work page 2022

[5] [8]

SOFTS: efficient multivariate time series forecasting with series-core fusion

[Han et al., 2024] Lu Han, Xu-Yang Chen, Han-Jia Ye, and De-Chuan Zhan. SOFTS: efficient multivariate time series forecasting with series-core fusion. CoRR, abs/2404.14197,

arXiv 2024

[6] [9]

Gaussian error linear units (gelus)

[Hendrycks and Gimpel, 2016] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415,

arXiv 2016

[7] [10]

Prediction then cor- rection: An abductive prediction correction method for se- quential recommendation

[Huang et al., 2023] Yulong Huang, Yang Zhang, Qifan Wang, Chenxu Wang, and Fuli Feng. Prediction then cor- rection: An abductive prediction correction method for se- quential recommendation. In SIGIR, pages 2272–2276. ACM,

work page 2023

[8] [11]

Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, and Qingsong Wen

[Jin et al., 2024] Ming Jin, Shiyu Wang, Lintao Ma, Zhix- uan Chu, James Y . Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, and Qingsong Wen. Time-LLM: Time series forecasting by reprogram- ming large language models. In ICLR. OpenReview.net,

work page 2024