Pith. sign in

REVIEW 6 major objections 5 minor 2 references

QuantFlow claims that bidirectional Mamba layers over inverted variable embeddings can replace attention in a forecaster, holding R2 above 0.90 on eight datasets and transferring zero-shot.

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

2026-08-02 09:01 UTC pith:AOFPRYHH

load-bearing objection A plausible Mamba+inverted-embedding integration, but the paper's internal contradictions and missing rollout methodology make its central claims unsupported. the 6 major comments →

arxiv 2607.02632 v2 pith:AOFPRYHH submitted 2026-07-02 cs.LG cs.AI

QuantFlow: A Federated Mamba-Based Post-Transformer Foundation Model for Time-Series Forecasting

Shah Nawaz Haider , Steve Austin , Arnab Barua , Sarowar Morshed Shawon , Hadaate Ullah This is my paper
classification cs.LG cs.AI MSC 68T0762M1062M20
keywords time-series forecastingstate-space modelsMambafederated learningquantile regressionprobabilistic forecastingzero-shot transfernon-IID data
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.

QuantFlow tries to prove that a state-space architecture — bidirectional Mamba layers over inverted variable embeddings — can replace Transformer attention in a foundation model for time-series forecasting without losing accuracy, while also producing probabilistic outputs and supporting federated training. If correct, it means practitioners can forecast long, high-dimensional, privacy-sensitive series at linear cost in sequence length, without centralizing raw data, and with explicit uncertainty intervals. The paper reports R2 above 0.90 on all eight centralized datasets, retained accuracy after three federated rounds with 20 non-IID clients, and zero-shot transfer to solar, cryptocurrency, and weather series. The central promise is that selective state-space modeling is a viable backbone for scalable, uncertainty-aware, privacy-preserving forecasting.

Core claim

On its own terms, QuantFlow's discovery is that a decoder built from six bidirectional Mamba state-space layers, operating on inverted embeddings where each variable is a token spanning the full 100-step look-back, can match or beat Transformer-based forecasters across 96-to-720-step horizons. The authors show this empirically on eight datasets for centralized training (R2 from 0.912 on traffic to 0.972 on ETTm2), on a 20-client non-IID federated setup after only three sample-weighted averaging rounds (R2 still above 0.89), and zero-shot on five unseen domains, with best average R2 of 0.760 on solar energy and 0.752 on Bitcoin. The architecture's forward/reverse scanning is presented as the

What carries the argument

The central mechanism is the bidirectional Mamba decoder block combined with inverted embedding. Inverted embedding treats each variable as a single token representing its entire observation window, permuting the input to B×V×100, so the sequence model processes variables rather than time positions. Each of six layers runs the Mamba state-space model forward and backward over these variable tokens, flips the reverse pass back, sums the two directions, and pushes through a convolutional feed-forward block, projecting to five conditional quantiles under pinball loss. This yields linear complexity in sequence length and captures dependencies in both temporal directions; sample-weighted federate

Load-bearing premise

The multi-horizon claims assume that a model trained one step ahead can be rolled out autoregressively for 720 steps without error accumulation; the paper offers no analysis or mitigation of compounding rollout errors.

What would settle it

Train QuantFlow on the same data with a direct multi-step objective — predict all 96/192/336/720 steps in one forward pass — and compare to the autoregressive rollout at 720 steps; if direct training is clearly better, the long-horizon advantage is a rollout artifact. Separately, compute empirical coverage of the 50% and 80% quantile intervals on the test split; coverage well below nominal rates would falsify the probabilistic-calibration claim.

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

If this is right

  • Long-horizon forecasting becomes computationally feasible for high-dimensional series, since cost grows linearly rather than quadratically with context length.
  • Organizations can collaboratively train a shared forecaster without sharing raw records; after three rounds, the model retains R2 above 0.89 even on non-IID client data.
  • Probabilistic forecasts with 50% and 80% intervals come from the quantile head, giving decision-makers uncertainty bands alongside point predictions.
  • Zero-shot results suggest a single pretrained QuantFlow could be deployed on new domains without per-dataset fine-tuning, with average R2 near 0.75 on solar and crypto series.
  • The ablation points to inverted embedding as the main contributor, implying that how tokens are formed matters more than the direction of the scan.

Where Pith is reading between the lines

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

  • A natural next test is whether QuantFlow's pretrained weights generalize to non-forecasting tasks like anomaly detection or imputation, since the inverted embeddings encode dense variable-level representations.
  • The federated experiment uses only three rounds; a stress test with more rounds and stronger client drift would reveal whether the retained accuracy is stable or an artifact of a short run.
  • Because long-horizon results come from autoregressive rollout of a one-step model, a direct multi-step training variant would isolate how much of the reported 720-step performance is due to the architecture versus rollout behavior.
  • Quantile intervals are displayed but empirical coverage is not measured; a coverage check on held-out windows would confirm whether the reported 50%/80% intervals are statistically honest.

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

6 major / 5 minor

Summary. The paper proposes QuantFlow, a federated time-series forecasting framework that combines inverted sequence embedding, bidirectional Mamba state-space layers, quantile regression, TSMixup augmentation, and federated averaging. The claimed contributions are linear-complexity long-horizon forecasting, probabilistic predictions, privacy-preserving non-IID federated training, and zero-shot transfer across domains. Centralized experiments report R2>0.90 on all eight training datasets, federated results retain accuracy after three rounds on 20 non-IID clients, zero-shot results show strong transfer to finance, energy, and epidemiology, and multi-horizon comparisons with Informer, FEDformer, TimesNet, TimeMixer, iTransformer, and N-HiTS are claimed to be comparable or better. The paper also includes an ablation study attributing performance to inverted embedding and bidirectional Mamba layers.

Significance. If the claims were supported, QuantFlow would be a useful addition to the literature: it targets linear-complexity probabilistic forecasting with a privacy-preserving training protocol, provides a public code link, and evaluates on multiple domains including a zero-shot setting. The architecture is simple and the combination of inverted variable embeddings with a state-space decoder is a plausible alternative to attention-based models. However, the manuscript as written contains multiple load-bearing inconsistencies: the abstract's numerical results conflict with Table 2; Section 4.4 attributes the method to a diffusion and flow-matching component that appears nowhere in the methodology; the long-horizon results rely on an undescribed autoregressive rollout of a one-step model; and Table 4 contradicts the text's claim of universally lower errors. These issues prevent the reported evidence from supporting the paper's headline claims, even though the underlying model may be salvageable with substantial corrections and re-evaluation.

major comments (6)
  1. [Abstract vs. Table 2] The abstract reports MSEs of 0.2834 on ETTm1 and 0.2218 on Weather, but Table 2 lists ETTm1 MSE=0.034, MAE=0.132 and Weather MSE=0.051, MAE=0.158. These differ by roughly an order of magnitude. Because the abstract is part of the public claims, this inconsistency must be resolved; it also undermines confidence in numerical reporting elsewhere in the paper.
  2. [Sections 4.4 and 5] The text states twice that QuantFlow uses 'a Mamba architecture combined with diffusion and flow matching' and that this 'effectively captures long-range dependencies.' No diffusion or flow-matching component is defined in the methodology (Section 3.4), in the model diagram, or in the loss (Eq. 2). Either this component must be added to the architecture and its training described, or the claim should be removed. As written, the experimental results cannot be attributed to an undefined mechanism.
  3. [Sections 3.3 and 4.4; Table 4] The training objective is one-step-ahead quantile prediction with look-back 100 and horizon 1 (Section 3.3). Table 4 reports horizons 96, 192, 336, and 720, and Section 4.4 says forecasting is performed 'with a 96-steps multi-step prediction as autoregressive.' No rollout procedure is specified: how predicted quantiles are fed back, whether all variables are predicted jointly, how instance normalization is applied at each step, and how the quantile heads are converted to multi-step point forecasts. No analysis of error accumulation or distribution shift is provided. The long-horizon numbers and the 'slower degradation' claim therefore rest on an unspecified procedure.
  4. [Table 4 vs. Section 4.4] The text claims QuantFlow provides 'lower prediction error, across all the datasets, than the competition.' The table does not support this. On Weather the QuantFlow average MSE is 0.300 versus TimeMixer 0.240, N-HiTS 0.248, and iTransformer 0.257; on Traffic the average is 0.450 versus iTransformer 0.428; on ETTm2 it is 0.280 versus iTransformer 0.275. In addition, many baseline cells are blank (e.g., Informer at 96/192 on several datasets, N-HiTS at 96/192 on many rows), so averages are not computed over the same set of runs. A valid comparison needs complete baseline entries and a clear statement of which entries are included in each average.
  5. [Section 3.4] The architecture permutes the input to B×V×100 and projects each variable's 100-step window to a 256-d embedding; the six Mamba layers then process the V variable embeddings. The sequence dimension for the state-space model is therefore the number of variables, not time. This is consistent with an inverted-embedding, cross-variable model, but the paper repeatedly claims the Mamba layers capture 'long-range temporal dependencies' (e.g., Sections 1 and 4.4). As described, temporal dependencies are summarized only by the linear projection over the fixed 100-step window, and the semantic meaning of forward/reverse direction over variables is unclear. The architecture description and the temporal-modeling claims need to be reconciled.
  6. [Section 4.1 and Section 5] The paper says the same eight datasets are used for 'pretraining, centralized and federated evaluations.' It does not state whether centralized evaluation is performed on a model pretrained on the same data, how the chronological split is constructed, or whether test windows overlap with pretraining windows. Without this protocol, the R2>0.90 claim could reflect in-distribution fitting. The paper also acknowledges that baseline comparisons are 'not a controlled leaderboard' because splits, horizons, and training budgets may not align. For the comparative claims in Table 4, baselines should be rerun under identical splits and budgets, or the claims should be softened accordingly.
minor comments (5)
  1. [Table 1] The configuration table omits optimizer, learning rate, batch size, and number of epochs, even though the text states 'A configuration is shown in Table 1.' These settings are necessary for reproducibility.
  2. [Section 3.5] The non-IID client partition is not described. Please specify how the 20 clients are constructed (e.g., per-domain, per-client temporal blocks, label/feature skew) and how local training is performed.
  3. [Equation (1)] TSMixup samples k sequences and forms a Dirichlet-weighted sum, but the paper does not state whether mixing is across time, across variables, or across client datasets. It should also explain how k and α are chosen and how the interpolation is guaranteed to preserve temporal dependencies rather than creating unrealistic mixtures.
  4. [Notation] The dataset names are inconsistent: Figure 4 and some text use 'ETTH1' while Tables 2 and 4 use 'ETTh1.' Please standardize the capitalization.
  5. [Section 4.5] The sentence 'inverted embedding is the foundation for the performance of QuantFlow, while the bidirectional Mamba is the foundation' is self-contradictory. Please rephrase to state which component contributes which effect.

Circularity Check

0 steps flagged

No significant circularity: QuantFlow's claims are empirical and do not reduce to fitted inputs or self-citations.

full rationale

QuantFlow is an empirical systems paper, not a derivation chain. The central components—inverted embedding, bidirectional Mamba layers, quantile loss (Eq. 2), FedAvg update (Eq. 3), and TSMixup (Eq. 1)—are defined independently of the reported results. No parameter is fit to a subset and then renamed as a prediction; the multi-horizon numbers in Table 4 are obtained by autoregressive rollout of the one-step model, which is a methodological risk (error accumulation) rather than a circular reduction. The comparison baselines are cited external methods, and the zero-shot results on Bitcoin, Ethereum, Solana, ILI, and Solar-Energy are out-of-distribution and therefore provide independent grounding. There are no self-citations, no imported uniqueness theorems, and no ansatz smuggled in through the authors' prior work. The apparent 'diffusion and flow matching' mentions in Section 4.4 are not in the methodology, but this is an internal consistency/correctness issue, not circularity. I therefore assign score 0.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The model rests on standard deep-learning components and benchmark assumptions. The main unexamined choices are the autoregressive rollout assumption, TSMixup's free concentration, and hand-picked hyperparameters. No new physical or mathematical entities are introduced.

free parameters (6)
  • Look-back window length = 100
    Chosen by hand in Section 3.3/Table 1; defines the embedding sequence length and is not swept.
  • Embedding and FFN dimensions = 256/1024
    Architecture hyperparameters from Table 1, chosen without reported search.
  • Decoder depth = 6
    Number of bidirectional Mamba layers in Table 1; no depth ablation is reported.
  • TSMixup Dirichlet concentration α = not reported
    Controls augmentation strength in Eq. (1); never specified or ablated.
  • Federated clients and rounds = 20/3
    Selected for the federated scenario in Table 1; no sensitivity analysis to client count or round count.
  • Optimizer, learning rate, batch size, epochs = not reported
    Training procedure is not specified anywhere; results cannot be reproduced from text alone.
axioms (6)
  • standard math Mamba selective state-space layers capture long-range dependencies with linear complexity.
    Taken from Gu and Dao (2024); not rederived, but a standard result in the cited literature.
  • standard math Minimizing pinball loss yields consistent conditional quantiles.
    Standard quantile-regression property behind Eq. (2).
  • domain assumption Sample-weighted FedAvg aggregation is effective for non-IID clients after three rounds.
    No convergence analysis or baseline comparison is provided in Section 3.5/Table 3.
  • domain assumption A one-step-ahead model can be rolled out autoregressively to 720 steps without error accumulation destroying accuracy.
    Critical for Table 4; stated in Section 4.4 as autoregressive prediction but never analyzed.
  • ad hoc to paper TSMixup interpolation preserves temporal dependencies.
    Introduced in Eq. (1) with no validation that synthetic sequences retain forecasting-relevant structure.
  • domain assumption Instance-wise min-max scaling and reverse normalization keep cross-dataset R² values comparable.
    Scaling details in Sections 3.3–3.4; the comparability of R² across datasets is assumed.

pith-pipeline@v1.3.0-alltime-deepseek · 122 in / 14941 out tokens · 236469 ms · 2026-08-02T09:01:13.598205+00:00 · methodology

0 comments
read the original abstract

Time-series forecasting supports decisions in finance, en-ergy, transportation, public health, and industrial monitoring. Recent foundation models improve transfer across forecast-ing tasks, but many depend on centralized data and Trans-former attention, which restricts their use for long, high-di-mensional, and privacy-sensitive signals. This paper presents QuantFlow, a probabilistic forecasting framework that com-bines inverted sequence embedding, bidirectional Mamba state-space decoders, quantile regression, and federated learning. Each variable is embedded over the complete ob-servation window, processed in forward and reverse direc-tions, and projected to five conditional quantiles. TSMixup expands temporal diversity through Dirichlet-weighted inter-polation while preserving sequence structure. Experiments cover cryptocurrency, traffic, electricity, Electricity Trans-former Temperature, influenza, and weather data. QuantFlow obtains mean squared errors of 0.2834 on ETTm1 and 0.2218 on Weather, and a 20-client non-IID deployment retains use-ful accuracy after three communication rounds without cen-tralizing raw records. The results indicate that selective state-space modelling is a promising basis for scalable, uncer-tainty-aware, and privacy-conscious time-series prediction, while also revealing limitations on irregular epidemiological signals and long-horizon generalization.

Figures

Figures reproduced from arXiv: 2607.02632 by Arnab Barua, Hadaate Ullah, Sarowar Morshed Shawon, Shah Nawaz Haider, Steve Austin.

Figure 1
Figure 1. Figure 1: Federated training workflow, model structure, and strategies for pre-training and fine-tuning [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: TSMixup Augmentation with different samples The sampling procedure, shown in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 1
Figure 1. Figure 1: Federated training workflow, model structure, and strategies for pre-training and fine-tuning. 3.2 Data and TSMixup We evaluate the framework on datasets from six application areas: Electricity (15 and 60-minute, 321 clients), Weather (21 meteorological variables, 60-minute resolution), Elec￾tricity Transformer (1-hour (h1, h2); 15 minutes (m1, m2) for oil temperature and power load, respectively) broken i… view at source ↗

Review history (2 revisions) →

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

2 extracted references · 2 linked inside Pith

  1. [1191]

    Wang, S.; Wu, H.; Shi, X.; Hu, T.; Luo, H.; Ma, L.; Zhang, J

    doi.org/10.1016/j.ijforecast.2019.07.001. Wang, S.; Wu, H.; Shi, X.; Hu, T.; Luo, H.; Ma, L.; Zhang, J. Y.; and Zhou, J. 2024. TimeMixer: Decomposable Multiscale Mix- ing for Time Series Forecasting. In Proceedings of the Twelfth International Conference on Learning Representations. https://doi.org/10.48550/arXiv.2405.14616 Woo, G.; Liu, C.; Kumar, A.; Xi...

    Pith/arXiv arXiv doi:10.48550/arxiv.2405.14616 2019

  2. [1764]

    Liu, Y.; Hu, T.; Zhang, H.; Wu, H.; Wang, S.; Ma, L.; and Long, M

    doi.org/10.1016/j.ijforecast.2021.03.012. Liu, Y.; Hu, T.; Zhang, H.; Wu, H.; Wang, S.; Ma, L.; and Long, M. 2024. iTransformer: Inverted Transformers Are Effective for Time Series Forecasting. In Proceedings of the Twelfth Interna- tional Conferen ce on Learning Representations. https://doi.org/10.48550/arXiv.2310.06625 McMahan, H. B.; Moore, E.; Ramage,...

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.06625 2021