Pith. sign in

REVIEW 4 major objections 4 minor 13 references

CaReTS claims that splitting multi-step forecasts into a trend direction plus a deviation magnitude, learned jointly, lowers error and gives a stable trend readout.

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

2026-08-03 22:32 UTC pith:NQAUFNQB

load-bearing objection Useful multi-task framework for short-horizon forecasting, but the SOTA claim is not supported by the paper's own table; the trend-accuracy comparison lacks trivial baselines. the 4 major comments →

arxiv 2511.09789 v2 pith:NQAUFNQB submitted 2025-11-12 cs.LG

CaReTS: A Multi-Task Framework Unifying Classification and Regression for Time Series Forecasting

Fulong Yao , Wanqing Zhao , Chao Zheng , Xiaofei Han This is my paper
classification cs.LG
keywords multi-task learningtime series forecastingtrend classificationdeviation regressionresidual forecast decompositionuncertainty-aware loss weightingdual-stream architecturemulti-step prediction
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.

CaReTS proposes that multi-step time series forecasting is better framed as two complementary tasks rather than one pure regression: a classification branch decides, step by step, whether the next value moves up or down relative to the last observed value, and a regression branch estimates the size of that move. Predictions are assembled in residual form — the latest observation plus a signed deviation — so macro trend and micro deviation are explicitly separated. The paper argues this dual-stream decomposition, trained jointly with an uncertainty-weighted loss, improves numerical accuracy while providing an interpretable trend readout, and reports competitive or improved RMSE compared with ten reference forecasting algorithms on two one-year hourly energy datasets while keeping stepwise trend accuracy above 90 percent. The claim matters because it offers a parameter-efficient way to make deep forecasters more interpretable without sacrificing accuracy.

Core claim

The central discovery claimed is that replacing the usual end-to-end regression forecast with a residual trend-deviation structure works: for each future step k, the model predicts a binary or probabilistic direction d̂(k) and a non-negative deviation δ̂(k), then forms the forecast as ŷ(k)=x_n + d̂(k)·δ̂(k) (or the soft probability-weighted analogue). Because the forecast is anchored to the most recent observation x_n, the classification branch carries the macro trend and the regression branch the micro adjustment; the two are trained together with a multi-task loss whose task weights are learned as inverse variances. The authors instantiate four variants over CNN, LSTM, and Transformer enco

What carries the argument

The load-bearing mechanism is the residual decomposition identity ŷ(k) = x_n + d̂(k)·δ̂(k) (and its soft/probabilistic forms), which anchors each multi-step forecast to the last observed value x_n. The classification branch (binary or softmax over up/down) supplies the sign d̂(k); the regression branch supplies a non-negative magnitude δ̂(k) (or direction-specific magnitudes, or a signed deviation); the multi-task loss combines classification, deviation, and output losses with uncertainty-based weights α_i = 1/(2σ_i^2) so task contributions self-balance during training. This decomposition is what turns a pure regressor into two interpretable streams.

Load-bearing premise

The accuracy advantage rests on the assumption that the binary direction of each future step relative to the latest observation is reliably learnable from a 15-dimensional input; because the forecast is built as the latest value plus a signed deviation, a wrong sign cannot be corrected by the regression branch.

What would settle it

Record the trend accuracy of a persistence baseline — predict that each future step moves in the same direction as the most recent step — on the same ten folds. If that baseline reaches or exceeds the reported roughly 91 percent, the classification branch's high trend accuracy does not demonstrate learning of macro trends, and the RMSE gains attributed to the decomposition would need re-examination.

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

If this is right

  • If the decomposition works as claimed, forecasters can expose a per-step up/down readout alongside point values, giving operators a direct signal for why the model is moving the forecast in a given direction.
  • The framework is encoder-agnostic: it can be attached to CNN, LSTM, or Transformer backbones, so the reported gains come from the output decomposition rather than a new temporal encoder.
  • Multi-task training with uncertainty-weighted losses improves RMSE relative to the same backbone trained only on output regression, indicating that the classification and deviation tasks complement rather than interfere with each other.
  • Trend classification accuracy stays roughly stable, even slightly increasing, as the prediction horizon grows, in contrast to RMSE which degrades — suggesting the macro-trend signal remains informative at longer horizons.
  • The hard sign fusion in CaReTS1 means any misclassification of direction directly distorts the forecast; the soft-probabilistic variant CaReTS3 is the natural hedge when trend confidence is low.

Where Pith is reading between the lines

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

  • The paper only tests a 15-to-6 setting on two hourly series, so the main open question is whether the sign of y(k)−x_n remains learnable at longer horizons or on noisier series; a natural extension is to test the decomposition on long-horizon and multivariate benchmarks.
  • The reported trend accuracy should be compared with trivial baselines such as 'same direction as the previous step' — on smooth hourly series this persistence baseline can already exceed 90 percent, which would put the classification branch's contribution in perspective.
  • Because trend accuracy stays stable while RMSE climbs with horizon, the optimal uncertainty-based task weighting likely differs across steps; making the weights horizon-dependent is a direct testable extension.
  • CaReTS3's soft blending already provides a mechanism to hedge uncertain trends, but its slightly worse RMSE in some settings suggests the right amount of softness may be data-dependent and worth tuning adaptively.

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 / 4 minor

Summary. The paper proposes CaReTS, a multi-task forecasting framework with a classification branch that predicts stepwise trend direction and a regression branch that predicts deviations from the latest observation. The final forecast is a residual combination of the latest observation, the predicted direction, and the predicted deviation. Four variants (CaReTS1--4) instantiate two dual-stream architectures with different fusion and loss designs, using CNN, LSTM, or Transformer encoders. The training objective combines classification, deviation, and output-prediction losses with uncertainty-based weighting. The authors evaluate the variants on two datasets (unmet power and electricity price) under a 15-input/6-output scheme using 10-fold cross-validation, comparing against three designed baselines and ten existing methods. The central claim is that CaReTS outperforms state-of-the-art algorithms in forecasting accuracy while also providing higher trend classification accuracy and improved interpretability.

Significance. If the empirical claims were rigorously established, the paper would offer a useful way to add directional interpretability to existing forecasting encoders while maintaining or improving accuracy. The residual decomposition into a sign and a magnitude is simple and plausible, and the four instantiations demonstrate compatibility with CNNs, LSTMs, and Transformers. The paper also ships an anonymous code link, which supports reproducibility. However, the significance is currently limited by the evidence: only two datasets are used, all from the same source and protocol; the headline SOTA claim is contradicted by the paper's own Table 5 on electricity price; and the trend-accuracy results are reported without comparison to persistence or majority-class baselines, leaving the interpretation gains not demonstrated. The technical derivation itself is internally coherent, so the main risk is evidential rather than mathematical.

major comments (4)
  1. [Abstract; Sec. 4.3; Table 5] The abstract states that CaReTS "outperforms state-of-the-art (SOTA) algorithms in forecasting accuracy," and Sec. 4.3 concludes that "CaReTS achieves state-of-the-art performance." Table 5 does not support this unqualified claim: for electricity price, TimeXer has the lower RMSE (0.0463±0.0013) than the best CaReTS variant, CaReTS2 (0.0465±0.0012). The difference is 0.0002, far smaller than the reported cross-validation standard deviations, and no paired significance test is provided anywhere. For unmet power, CaReTS2 is numerically lower than TimeXer (0.0691±0.0018 vs. 0.0700±0.0022), but again the gap is within one standard deviation. The load-bearing claim of SOTA accuracy is therefore not established by the reported evidence. The authors should either add proper paired significance tests (e.g., across the 10 folds) and correct the abstract/conclusion, or substantially weaken the cla
  2. [Sec. 3, Eqs. (9), (12); Table 3; Sec. 4.3] The trend-accuracy comparison is not controlled. The ground-truth trend labels t(k) are defined as sign(y(k)-x_n), and for CaReTS1/2 the classification branch is explicitly trained with BCE on these labels (Eq. 9). Reporting the training metric as the headline "trend classification performance" is circular in the sense that the metric is directly optimized, and the accuracy values in Table 3 are never compared with persistence (forecasting the previous direction), majority-class, or random baselines. If the sign of future changes is strongly autocorrelated, accuracy above 90% may be trivial. Figure 5 further shows stable accuracy across horizons, but without a persistence baseline this is not informative. To support the interpretability claim, the authors should report class balance, persistence/naive trend accuracy, and confusion or F1-type metrics.
  3. [Sec. 4.2, Table 4] The multi-task versus single-task ablation is not a clean comparison. The text states that in the single-task setting the trend branch was kept trainable by implementing the trend direction "in a continuous form," which changes the architecture and the classification stream. Thus the difference in Table 4 may be due to architectural or gradient-flow differences rather than the multi-task loss. The authors need to describe exactly how the single-task variant is constructed, what its continuous trend output means, and whether the reported trend accuracy is computed in the same way as for the multi-task model. Without this, the conclusion that multi-task learning causes the RMSE improvement is not established.
  4. [Sec. 4; Appendix A.6] The empirical evaluation is limited to two datasets, both hourly energy-related series from the same source, under a fixed 15-step input setting (with horizons 4, 6, and 8 in the appendix). The claim of broad superiority over SOTA and applicability to "multi-step time series forecasting" in general is over-generalized. At minimum, the authors should acknowledge this limitation in the main text and temper the SOTA claim to the two datasets, or add results on standard multivariate and long-horizon benchmarks (e.g., ETT, Weather, Electricity) to support the general claim.
minor comments (4)
  1. [Sec. 4.1] The name is inconsistently spelled as "CaReST" in several places (e.g., "CaReST 1--4" at the start of Sec. 4.1); should be CaReTS.
  2. [Sec. 2.2, Eq. (3)] The text says the task weight is modeled as the inverse of the predicted variance, but Eq. (3) has a factor 1/2. If the factor is intentional (following Kendall et al.), the text should say "half the inverse variance" or explain why the factor is present.
  3. [Sec. 4.1, Table 2] Table 2 reports train, validation, and test RMSE, but the surrounding text and the x-axis of Figures 2--3 refer to "test set." Clarify whether the figures show the test column or a per-step test-set breakdown, and make the table caption explicit.
  4. [Sec. 4.3, Table 5] Table 5 formatting is broken in the header row ("Approach Unmet power Electricity priceRMSE Trend Acc.") and should be fixed. Also, the time column for SOTA algorithms is reported but no hardware or measurement protocol is described for a fair runtime comparison.

Circularity Check

0 steps flagged

No significant circularity: CaReTS is trained end-to-end against external SOTA benchmarks; self-citations supply datasets and a baseline but are not load-bearing derivations.

full rationale

The derivation chain is self-contained. CaReTS constructs forecasts as the latest observation plus a learned signed magnitude (Eqs. 8, 12, 15, 17), with a classification head producing the sign and a regression head producing the magnitude. Both heads are supervised with ground-truth targets derived from future values (Eqs. 9, 10, 13, 16, 18), and the final forecast is additionally supervised by the output loss (Eqs. 11, 19). This is a reparameterized multi-task regression, not a quantity that reduces to its own inputs by construction: no fitted parameter is renamed as an independent prediction, and the trend labels are defined from the ground truth, not from the model's own output. The uncertainty-aware weights (Eqs. 3–5) are standard learnable scalars from Kendall et al. (2018), an external citation, and are not used to smuggle in the target result. Self-citations to Yao et al. (2025a,b) provide the two datasets and one baseline, but the main SOTA comparison in Table 5 uses external algorithms (Autoformer, FEDformer, TimeXer, etc.), so the central accuracy claim is independently benchmarked. The trend-accuracy comparison is asymmetric—CaReTS optimizes the classification loss directly, while SOTA trend accuracy is derived from real-valued forecasts—and Table 5 shows TimeXer has lower RMSE on electricity price; these are evidence-strength and benchmark-fairness concerns, not circular derivation. No uniqueness theorem or ansatz is imported via self-citation. Therefore no step meets the required bar of exhibiting Eq. X = Eq. Y by construction.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 1 invented entities

The central claim depends on the sign/magnitude decomposition, the self-derived trend labels, and the borrowed uncertainty weighting. No new physical entities are introduced. The learned uncertainty scalars and hand-set hyperparameters are the main fitting choices.

free parameters (3)
  • Learned uncertainty scalars log σ_i² = Not reported (per-fold learned scalars)
    Eqs. (4)-(5): task weights are learned during training rather than set by an external principle; no sensitivity analysis is given.
  • Uncertainty-loss regularizer coefficient (1/2) and clamping range [-10, 10] = 1/2 and ±10
    Sec. 2.2: chosen by hand to prevent degenerate uncertainty; no ablation or justification beyond stability.
  • Architecture hyperparameters (hidden 64, N_l=2, lr=1e-3, batch 64, patience 50, seed 2025) = See App. A.4
    Hand-set standard values; no tuning study, yet they affect all reported results.
axioms (4)
  • domain assumption Each future value can be decomposed as y(k) = x_n + sign(y(k)-x_n)·|y(k)-x_n|, and modeling the sign as classification and the magnitude as regression is beneficial.
    Eqs. (8), (12), (15), (17): this decomposition is the core inductive bias and is not derived or externally validated.
  • domain assumption Ground-truth trend labels t(k)=1[y(k) ≥ x_n] are a valid supervision signal.
    Eqs. (9)/(16): labels are computed from the target being predicted, with no independent or annotated trend definition.
  • domain assumption Kendall et al. (2018) uncertainty weighting transfers from scene geometry and semantics to time-series forecasting and improves joint training.
    Sec. 2.2: the loss weighting is borrowed from computer vision; its benefit is asserted from the multi-task vs single-task comparison.
  • domain assumption The two datasets from Yao et al. (2025b) and the ten SOTA implementations are representative and faithful.
    Sec. 4: only two one-year series from the authors' prior work are used; external comparisons are not independently reproduced.
invented entities (1)
  • None no independent evidence
    purpose: No new physical entities or independent constructs are postulated.
    The trend and deviation heads are model components, not invented entities.

pith-pipeline@v1.3.0-alltime-deepseek · 15617 in / 14139 out tokens · 154617 ms · 2026-08-03T22:32:23.579624+00:00 · methodology

0 comments
read the original abstract

Recent advances in deep forecasting models have achieved remarkable performance, yet most approaches still struggle to provide both accurate predictions and interpretable insights into temporal dynamics. This paper proposes CaReTS, a novel multi-task learning framework that combines classification and regression tasks for multi-step time series forecasting problems. The framework adopts a dual-stream architecture, where a classification branch learns the stepwise trend into the future, while a regression branch estimates the corresponding deviations from the latest observation of the target variable. The dual-stream design provides more interpretable predictions by disentangling macro-level trends from micro-level deviations in the target variable. To enable effective learning in output prediction, deviation estimation, and trend classification, we design a multi-task loss with uncertainty-aware weighting to adaptively balance the contribution of each task. Furthermore, four variants (CaReTS1--4) are instantiated under this framework to incorporate mainstream temporal modelling encoders, including convolutional neural networks (CNNs), long short-term memory networks (LSTMs), and Transformers. Experiments on real-world datasets demonstrate that CaReTS outperforms state-of-the-art (SOTA) algorithms in forecasting accuracy, while achieving higher trend classification performance.

Figures

Figures reproduced from arXiv: 2511.09789 by Chao Zheng, Fulong Yao, Wanqing Zhao, Xiaofei Han.

Figure 1
Figure 1. Figure 1: Two types of dual-stream CaReTS architectures [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: RMSE on power across approaches [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 5
Figure 5. Figure 5: Trend accuracy across forecasting steps using CaReTS2-Transformer 4.2 MULTI-TASK LEARNING EVALUATION We then took the Transformer encoder as a representative case to further evaluate the effectiveness of the multi-task learning mechanism in CaReTS1-4 [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

13 extracted references · 7 linked inside Pith

  1. [2]

    Tsmixer: An all-mlp architecture for time series forecasting.arXiv preprint arXiv:2303.06053,

    Si-An Chen, Chun-Liang Li, Nate Yoder, Sercan O Arik, and Tomas Pfister. Tsmixer: An all-mlp architecture for time series forecasting.arXiv preprint arXiv:2303.06053,

    Pith/arXiv arXiv

  2. [4]

    itransformer: Inverted transformers are effective for time series forecasting.arXiv preprint arXiv:2310.06625,

    Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. itransformer: Inverted transformers are effective for time series forecasting.arXiv preprint arXiv:2310.06625,

    Pith/arXiv arXiv

  3. [5]

    A time series is worth 64 words: Long-term forecasting with transformers.arXiv preprint arXiv:2211.14730,

    Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. A time series is worth 64 words: Long-term forecasting with transformers.arXiv preprint arXiv:2211.14730,

    Pith/arXiv arXiv

  4. [6]

    Etsformer: Exponential smoothing transformers for time-series forecasting.arXiv preprint arXiv:2202.01381,

    Gerald Woo, Chenghao Liu, Doyen Sahoo, Akshat Kumar, and Steven Hoi. Etsformer: Exponential smoothing transformers for time-series forecasting.arXiv preprint arXiv:2202.01381,

    Pith/arXiv arXiv

  5. [8]

    CNN-based methods (Durairaj & Mohan, 2022; Sadouk,

    A APPENDIX A.1 RELATEDWORK Deep learning has substantially advanced time series forecasting, with models ranging from CNNs and RNNs to Transformers. CNN-based methods (Durairaj & Mohan, 2022; Sadouk,

    2022

  6. [11]

    Transformer variants including ETSformer (Woo et al., 2022), Autoformer (Wu et al., 2021), and FEDformer (Zhou et al.,

    demonstrate that simple linear trend–seasonal decomposition can outperform complex architectures. Transformer variants including ETSformer (Woo et al., 2022), Autoformer (Wu et al., 2021), and FEDformer (Zhou et al.,

    2022

  7. [12]

    However, these methods largely operate at the input or representation level, without directly disentangling the prediction targets

    explicitly model trend, seasonal, or frequency components, thereby enhancing interpretability. However, these methods largely operate at the input or representation level, without directly disentangling the prediction targets. In contrast, CaReTS in- troduces an output-level decomposition, separating macro-level trends via classification from micro- level...

    2022

  8. [2017]

    have emerged as the dominant backbone in time series forecasting. Representative ex- amples include Autoformer (Wu et al., 2021), which incorporates trend-seasonal decomposition and autocorrelation mechanisms; FEDformer (Zhou et al., 2022), which introduces frequency-domain decomposition; TimesNet (Wu et al., 2023), which captures temporal variations in a...

    2021

  9. [2019]

    are ef- fective in extracting local spatial or temporal patterns, while RNN variants such as LSTM and GRU (Elsworth & G¨uttel, 2020; Zhang et al.,

    2020

  10. [2022]

    Time series forecasting using lstm networks: A symbolic ap- proach.arXiv preprint arXiv:2003.05672,

    Steven Elsworth and Stefan G ¨uttel. Time series forecasting using lstm networks: A symbolic ap- proach.arXiv preprint arXiv:2003.05672,

    Pith/arXiv arXiv 2003

  11. [2023]

    Less is more: Fast multivariate time series forecasting with light sampling-oriented mlp structures

    Tianping Zhang, Yizhuo Zhang Zhang, Wei Cao, Jiang Bian, Xiaohan Yi, Shun Zheng, and Jian Li. Less is more: Fast multivariate time series forecasting with light sampling-oriented mlp structures. arXiv preprint arXiv:2207.01186,

    Pith/arXiv arXiv

  12. [2024]

    Edformer: Embedded decom- position transformer for interpretable multivariate time series predictions.arXiv preprint arXiv:2412.12227,

    Sanjay Chakraborty, Ibrahim Delibasoglu, and Fredrik Heintz. Edformer: Embedded decom- position transformer for interpretable multivariate time series predictions.arXiv preprint arXiv:2412.12227,

    Pith/arXiv arXiv

  13. [2025]

    For the encoder design,N l = 2layers with 64 hidden units were adopted in three encoder variants

    All datasets were preprocessed using Min-Max normalization, and ReLU was applied as the activation function. For the encoder design,N l = 2layers with 64 hidden units were adopted in three encoder variants. Specifically, a kernel size of 3 with padding of 1 was used in the CNN encoder, while all Transformer encoders were configured with 4 attention heads....

    arXiv