Pith. sign in

REVIEW 3 major objections 5 minor 62 references

A forecasting framework that reuses its learned predictive-state gate to localize conformal calibration can produce narrower prediction intervals without sacrificing coverage on nonstationary multivariate series.

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

2026-08-01 03:24 UTC pith:H56DWSAL

load-bearing objection Solid empirical paper on a genuinely new combination; the coverage theorem is conditional on an assumption close to the conclusion, so treat the guarantee as a design goal rather than a proven property. the 3 major comments →

arxiv 2607.23165 v1 pith:H56DWSAL submitted 2026-07-25 stat.ML cs.LGstat.APstat.ME

Adaptive Multi-Scale Forecasting and Gate-Localized Conformal Prediction for Multivariate Nonstationary Time Series

Ziling Ma , Junshu Jiang , \'Angel L\'opez-Oriona , Ying Sun , Hernando Ombao This is my paper
classification stat.ML cs.LGstat.APstat.ME MSC 62M1062G15
keywords nonstationary time seriesmultivariate forecastingconformal predictionuncertainty quantificationmulti-scale forecastingmixture of expertsrealized volatilitygate-localized calibration
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.

The paper tries to establish that a single learned predictive-state representation can serve both point forecasting and uncertainty quantification in nonstationary multivariate time series, so that prediction intervals adapt to the same regimes as the point forecasts. The forecasting side, ABF-T, keeps a validation-selected anchor expert and adds gated multi-scale corrections plus sparse cross-series transfer, avoiding the variance from blindly mixing weak experts. The uncertainty side, GLCP, weighs calibration residuals by temporal recency and gate-state similarity, producing sharper intervals while keeping empirical coverage near 90% on a 55-commodity high-frequency benchmark. A theorem gives approximate local coverage with error terms separating localization mismatch from finite-sample estimation. The authors also show the approach transfers to solar forecasting and to a neural backbone, arguing the framework is model-agnostic.

Core claim

The central claim is that the gate state — the normalized weight vector over temporal-scale experts, learned for forecasting — is a useful localization variable for conformal prediction under nonstationarity. GLCP defines weights w_g = λ^(t−g) exp(−||π_g − π_t||²/(2τ²)) and takes weighted quantiles of MAD-normalized residuals, optionally asymmetric for lower and upper tails. In the authors' experiments across horizons from 15 minutes to 8 hours, asymmetric GLCP achieves the lowest interval scores and widths among nine conformal baselines on the same ABF-T point forecasts, with coverage between 0.893 and 0.908 against a 0.90 nominal level, and effective sample size around 304 (between global

What carries the argument

The double-duty gate vector π_{t,j}. It is a softmax output of an asset-specific network over K lookback-scale experts; in ABF it weights corrections around a validation-anchored expert, and in GLCP it defines similarity between historical and current calibration points. The localization weight w_g = λ^(t−g) exp(−||π_g − π_t||²/(2τ²)) combines recency with gate proximity, and the weighted empirical quantile of MAD-normalized (absolute or signed) residuals sets the interval radius. Validation selection of anchor, correction strength ρ, transfer shrinkage γ, and bandwidth τ enforces the 'no negative transfer' safeguard.

Load-bearing premise

The guarantee rests on Assumption 1: that the gate-weighted historical residual distributions approximate the current test residual distribution up to a small error δ, and the paper gives no diagnostic or bound establishing that δ is small; if the gate is uninformative or unstable, localization can worsen coverage rather than improve it.

What would settle it

Run GLCP on a dataset where the gate state vectors are randomly permuted across time before computing similarity, keeping everything else identical; if interval scores still beat recency-only calibration, the gate-similarity mechanism is not the driver. Alternatively, on a labeled regime dataset, compute the empirical δ by comparing the weighted historical score CDF to the actual test score CDF; large δ when coverage falls below nominal would falsify the local-coverage premise.

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

If this is right

  • If the shared-representation claim holds, interval widths will track volatility regime changes without inflating coverage, since the gate shifts as market dynamics change.
  • GLCP generalizes existing conformal calibration: taking λ→1 with τ→∞ recovers split conformal, τ→∞ with λ<1 recovers recency weighting, and λ=1 with τ<∞ recovers gate-only localization.
  • The validation-anchored ABF design implies that the framework can only improve on the best single-scale expert; it never mixes in weaker experts unless validation shows gains, shifting the risk of mixture-of-experts to the validation split.
  • The gains grow with horizon (1.3% at h=3 to 15.0% at h=96 MSE), suggesting the anchor-plus-gated-correction scheme is most valuable where predictive uncertainty is large and no single scale is reliable.

Where Pith is reading between the lines

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

  • If the gate truly captures regimes, a synthetic experiment with known regime switches could quantify the δ localization error in Theorem 1; high δ under rapid switching would show when the guarantee degrades.
  • The recency parameter λ and gate bandwidth τ are tuned on calibration only; recursively updating the gate or re-estimating τ online could further improve nonstationary performance, an extension the paper does not explore.
  • The transfer matrices reveal sparse, horizon-dependent cross-series structure; beyond forecasting, these coefficients could serve as interpretable lead–lag or sector-coupling discovery tools.
  • The m_min fallback to unweighted conformal quantiles insures against over-localization, but the paper does not study how coverage behaves exactly at the fallback threshold; a dedicated study of the ESS threshold could guide practitioners.

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

3 major / 5 minor

Summary. The paper proposes ABF-T-GLCP, a two-stage framework for nonstationary multivariate time-series forecasting. The point-forecasting stage (ABF-T) trains horizon-specific experts on multiple lookback scales, selects a validation-best anchor expert, adds a gate-weighted correction with strength selected on validation data, and then applies a sparse cross-series residual transfer with validation-controlled shrinkage. The uncertainty stage (GLCP) constructs conformal prediction intervals by weighting calibration residuals according to both temporal recency and similarity in the learned gate state, with symmetric and asymmetric variants; a theoretical claim of approximate local coverage is stated in Theorem 1 under Assumption 1. Experiments on a 55-asset high-frequency Chinese commodity futures dataset report consistent MSE/MAE gains for ABF-T relative to single-scale and gated mixture baselines, and consistently lower interval scores and widths for GLCP relative to ACI, AgACI, SAOCP, RCQR, EnbPI, RLCP, and MoECP, with coverage close to 90%. Additional experiments on a solar dataset and with a PatchTST backbone are included, together with ablations and hyperparameter sensitivity analyses.

Significance. If the empirical results hold, the framework is a practically useful combination of validation-anchored multi-scale blending, sparse transfer, and regime-aware conformal calibration. The manuscript has notable strengths: strict chronological train/validation/calibration/test splits, a public high-frequency commodity dataset and a second solar dataset, multiple horizons, ablations separating the contributions of transfer, gate similarity, and recency, a neural-backbone check with PatchTST, and code availability. The reuse of the learned gate state for conformal localization is a sensible and potentially valuable idea, and the reported interval-score improvements are large, especially at longer horizons. However, the paper's central theoretical contribution is currently much weaker than the abstract suggests: Theorem 1 is conditional on an unquantified local-stability assumption that essentially contains the coverage error it claims to control, and the empirical coverage numbers are marginal rather than local. The paper is therefore promising but needs substantial revision before the theoretical claims can be accepted.

major comments (3)
  1. [§3.6, Theorem 1 / Assumption 1; Supp §5.1] The theorem's guarantee is conditional on δ^{-}_{t,j} and δ^{+}_{t,j} in Assumption 1, which are precisely the local-coverage errors the abstract claims to control. No estimator, upper bound, or diagnostic for δ is provided; the supplementary proof simply reuses the assumption. Thus 'approximate local coverage under mild stability conditions' is not established as a theorem, and contribution (2) is, as stated, an assumption about the data. The aggregate coverage in Table 3 does not fill this gap because it is marginal over test times and assets. Please either prove a nontrivial finite-sample case (e.g., known covariate-shift weights or split-conformal localized residuals), add a calibration diagnostic quantifying δ, or explicitly downgrade the theoretical claim to a heuristic.
  2. [Algorithm 1, lines 12-16; Supp §5.1] The proof postulates a joint weighted empirical concentration event holding with probability at least 1−η, but no finite-sample concentration inequality is established for sequentially updated, adaptively weighted residual scores. Because the calibration history includes observed test points and the weights depend on the gate state at time t, the residuals are neither independent nor exchangeable; the ε terms are therefore also unquantified. Please state the precise dependence/exchangeability assumption needed and provide a bound, or identify which existing localized-conformal result applies to this online sequence.
  3. [§4.6, Table 3, Figure 5] The empirical coverage column is a global average over the whole test period and all series. It cannot distinguish a method that achieves 90% coverage in every regime from one that over-covers in calm periods and under-covers in volatile periods — precisely the risk the authors acknowledge in §5 for uninformative gates. Report coverage stratified by gate regime or volatility tercile, or at least the dispersion of per-asset/per-period coverage, to substantiate the local-coverage claim.
minor comments (5)
  1. [§4.6] The sentence 'Consistent with the average ESS of 304, the minimum ESS threshold (m_min=100) was never reached' does not follow from the average being above the threshold. Report the empirical minimum of ESS over (t,j,h), or the fraction of points below m_min.
  2. [Table 2 / §4.5] At h=3 the MSE gain over the best baseline is 1.27%, which is small relative to the seed standard deviation reported in Table 6 (ABF-T MSE 0.851±0.009). No paired significance test is provided. The claim of consistent gains is well supported at longer horizons; at short horizons it should be qualified or tested.
  3. [Table 5] The column heading 'Best Baseline ABF-T GLCP IS Best IS Gain' is confusing; clarify which columns refer to point-forecast MSE and which to interval score. Also report standard deviations for the solar results.
  4. [Throughout] The main-text tables (Tables 2 and 3) omit standard deviations; these appear only in the supplementary. Please include them in the main tables or clearly refer to the supplementary table numbers.
  5. [Abstract / Code availability] The abstract says 'The full code is available at GitHub' but no URL or repository identifier is given. Please provide a working link or anonymized repository.

Circularity Check

0 steps flagged

No significant circularity: GLCP intervals are constructed from calibration residuals and gate weights without test information, and Theorem 1 is an explicitly conditional guarantee rather than a result that reduces to its inputs.

full rationale

The paper's empirical derivation chain is self-contained. ABF's anchor and correction strength are chosen on the validation set, the transfer model is selected on an inner validation split, and GLCP's weights in Eq. (3) depend only on F_t-measurable gate states and calendar time; the conformal quantile is computed from calibration residuals, not from test outcomes. The localization bandwidth tau is selected on the calibration set and lambda is fixed, so no fitted parameter is renamed as a test prediction. Theorem 1 is conditional on Assumption 1 and an explicit concentration event; the delta and epsilon terms appear as error allowances, and the Supplementary proof is a standard quantile/union-bound argument. The fact that delta is unquantified and no diagnostic is supplied is a genuine limitation and a correctness/validity risk, but it is not circular: the paper does not define the coverage probability as equal to the assumption, and the abstract qualifies the guarantee as 'approximate local coverage' under stability conditions. The self-citations (Jiang et al. 2025 for the public dataset; Ma et al. 2026 and Lopez-Oriona et al. 2025 for related work) are not load-bearing: the dataset is publicly available and the related-work claims are corroborated by the paper's own ablations, not by the self-citations alone. Section 5 also explicitly acknowledges that uninformative gates weaken localization, which further shows the empirical claims are contingent rather than definitionally forced. No specific step can be exhibited as reducing to its own inputs by construction.

Axiom & Free-Parameter Ledger

9 free parameters · 5 axioms · 0 invented entities

The empirical results rest mainly on a strict chronological split and validation-driven selection; the free parameters are selected on validation/calibration rather than test. The theory, however, rests on Assumption 1, which assumes the local score stability that Theorem 1 outputs; without a delta bound the guarantee is conditional. No new physical entities are introduced.

free parameters (9)
  • rho_h (ABF correction strength) = selected on validation grid {0,0.025,...,0.50} per horizon
    Controls how much gate-weighted expert deviations are added to the validation-best anchor; chosen by validation MSE.
  • gamma_h (transfer shrinkage) = selected on validation grid {0,0.25,0.50,0.75,1.0}; 0 if transfer fails 1% inner-validation improvement
    Scales the sparse transfer correction; validation-selected.
  • tau (GLCP localization bandwidth) = selected on Cal set from {0.10,0.25,0.50,0.75,1.00,1.50,2.00}
    Controls gate-similarity kernel width; chosen to minimize interval score with coverage within 0.5% of nominal on Cal.
  • lambda (temporal recency decay) = 0.995 (fixed; sensitivity range 0.99-1)
    Controls exponential recency weighting; fixed by authors, not cross-validated in main experiments.
  • Transfer regularization hyperparameters (Lasso lambda1, Ridge lambda2, ElasticNet r, intercept) = grids lambda1 in {0.001,...,0.10}, lambda2 in {0.1,...,100}, r in {0.15,0.50,0.85}
    Selected on an inner 70/30 validation split.
  • MAD residual scale sigma_hat_j = estimated from rolling calibration residuals per series
    Normalizes conformal scores; data-dependent.
  • Gate architecture hyperparameters = hidden 64, dropout 0.1, lr 1e-3, wd 1e-4, batch 512, epochs 200, temp 0.8, top-k 3, regularization coefficients
    Chosen by hand; not fitted to validation targets.
  • m_min (ESS fallback threshold) = 100
    If effective sample size falls below 100, GLCP falls back to unweighted conformal quantile; chosen by hand.
  • W (rolling calibration window) = 2000
    Size of calibration history used for conformal weights; fixed.
axioms (5)
  • domain assumption Assumption 1 (Predictable tail-wise localization): weighted historical tail-score distributions approximate test distributions up to delta
    Stated in §3.6; the proof uses exactly this to convert weighted empirical quantiles into coverage. No bound or diagnostic for delta is given.
  • standard math Weighted empirical CDFs concentrate around localized population CDFs (event with probability at least 1-eta)
    Theorem 1 relies on a joint concentration event with bounds epsilon(eta); the proof does not specify the concentration inequality.
  • domain assumption The learned gate pi is an informative and stable localization variable
    GLCP weights are built from gate-state similarity; the authors concede in §5 that if the gate fails to distinguish regimes, localization is less effective.
  • domain assumption Calibration residuals in the rolling window are representative under the chosen weights
    The empirical weighted quantile and its fallback assume the calibration history captures the current error distribution; nonstationarity makes this approximate.
  • domain assumption Target construction and preprocessing (5-min log RV with epsilon, missing-value filling, chronological split) do not distort coverage conclusions
    No sensitivity analysis is provided for epsilon or for the missing-data filling rule.

pith-pipeline@v1.3.0-alltime-deepseek · 24631 in / 15470 out tokens · 146115 ms · 2026-08-01T03:24:50.424428+00:00 · methodology

0 comments
read the original abstract

We propose ABF-T-GLCP, a model-agnostic framework for forecasting and uncertainty quantification in nonstationary multivariate time series. The central idea is to learn an adaptive predictive state representation for point forecasting and reuse it for conformal calibration. The forecasting module combines horizon-specific temporal experts through a learned gate and refines predictions using sparse predictive transfer across related series. The uncertainty module, Gate-Localized Conformal Prediction (GLCP), uses the learned gate state, together with temporal recency, to select locally relevant calibration residuals, thereby coupling uncertainty calibration to the predictive regimes used by the forecasting model. This shared representation allows point forecasts and prediction intervals to adapt consistently under evolving temporal dynamics while retaining the model-agnostic nature of conformal prediction and yielding approximate local coverage under mild stability conditions. Experiments on a large-scale high-frequency commodity forecasting benchmark show consistent gains in point forecasting accuracy and substantially narrower prediction intervals with empirical coverage close to the nominal level. Additional results indicate that the framework extends beyond the motivating financial application.

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

62 extracted references · 1 canonical work pages

  1. [1]

    Energy and AI , volume=

    Conformal prediction for electricity price forecasting in the day-ahead and real-time balancing market , author=. Energy and AI , volume=. 2025 , publisher=

    2025

  2. [2]

    Applied Energy , volume=

    A review of solar and wind energy forecasting: From single-site to multi-site paradigm , author=. Applied Energy , volume=. 2025 , publisher=

    2025

  3. [3]

    IEEE Transactions on Neural Systems and Rehabilitation Engineering , volume=

    EEG-based seizure prediction via model uncertainty learning , author=. IEEE Transactions on Neural Systems and Rehabilitation Engineering , volume=. 2022 , publisher=

    2022

  4. [4]

    International Journal of Forecasting , volume=

    Electricity price forecasting: A review of the state-of-the-art with a look into the future , author=. International Journal of Forecasting , volume=. 2014 , publisher=

    2014

  5. [5]

    Tibshirani and Larry Wasserman , title =

    Jing Lei and Max G’Sell and Alessandro Rinaldo and Ryan J. Tibshirani and Larry Wasserman , title =. Journal of the American Statistical Association , volume =. 2018 , publisher =

    2018

  6. [6]

    2005 , publisher=

    Algorithmic learning in a random world , author=. 2005 , publisher=

    2005

  7. [7]

    Information and Inference: A Journal of the IMA , volume=

    The limits of distribution-free conditional predictive inference , author=. Information and Inference: A Journal of the IMA , volume=. 2021 , publisher=

    2021

  8. [8]

    Advances in Neural Information Processing Systems , volume=

    Conformal prediction under covariate shift , author=. Advances in Neural Information Processing Systems , volume=

  9. [9]

    arXiv preprint arXiv:2604.13748 , year=

    Forecasting Multivariate Time Series under Predictive Heterogeneity: A Validation-Driven Clustering Framework , author=. arXiv preprint arXiv:2604.13748 , year=

    Pith/arXiv arXiv

  10. [10]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

    Doubly robust calibration of prediction sets under covariate shift , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2024 , publisher=

    2024

  11. [11]

    International Conference on Machine Learning , pages=

    Sequential predictive conformal inference for time series , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  12. [12]

    arXiv preprint arXiv:2410.13115 , year=

    Online conformal inference for multi-step time series forecasting , author=. arXiv preprint arXiv:2410.13115 , year=

    arXiv

  13. [13]

    The Twelfth International Conference on Learning Representations , year=

    Copula conformal prediction for multi-step time series prediction , author=. The Twelfth International Conference on Learning Representations , year=

  14. [14]

    Forty-second International Conference on Machine Learning , year=

    Relational Conformal Prediction for Correlated Time Series , author=. Forty-second International Conference on Machine Learning , year=

  15. [15]

    arXiv preprint arXiv:2509.02844 , year=

    Conformal Prediction for Time-series Forecasting with Change Points , author=. arXiv preprint arXiv:2509.02844 , year=

    arXiv

  16. [16]

    and Jordan, Michael I

    Jacobs, Robert A. and Jordan, Michael I. and Nowlan, Steven J. and Hinton, Geoffrey E. , title =. Neural Computation , volume =. 1991 , month =. doi:10.1162/neco.1991.3.1.79 , url =

    doi:10.1162/neco.1991.3.1.79 1991

  17. [17]

    International Journal of Forecasting , volume=

    Forecast combinations: An over 50-year review , author=. International Journal of Forecasting , volume=. 2023 , publisher=

    2023

  18. [18]

    International conference on artificial intelligence and statistics , pages=

    Mixture-of-linear-experts for long-term time series forecasting , author=. International conference on artificial intelligence and statistics , pages=. 2024 , organization=

    2024

  19. [19]

    Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V

    Duet: Dual clustering enhanced multivariate time series forecasting , author=. Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1 , pages=

  20. [20]

    Fuzzy Sets and Systems , volume=

    Forecasting time series collections via fuzzy clustering , author=. Fuzzy Sets and Systems , volume=. 2026 , publisher=

    2026

  21. [21]

    Knowledge-based systems , volume=

    Time series clustering based on prediction accuracy of global forecasting models , author=. Knowledge-based systems , volume=. 2025 , publisher=

    2025

  22. [22]

    1986 , issn =

    Generalized autoregressive conditional heteroskedasticity , journal =. 1986 , issn =. doi:https://doi.org/10.1016/0304-4076(86)90063-1 , url =

    doi:10.1016/0304-4076(86)90063-1 1986

  23. [23]

    Andersen and Tim Bollerslev and Francis X

    Torben G. Andersen and Tim Bollerslev and Francis X. Diebold and Paul Labys , journal =. Modeling and Forecasting Realized Volatility , urldate =

  24. [24]

    Journal of Financial Econometrics , volume =

    Corsi, Fulvio , title =. Journal of Financial Econometrics , volume =. 2009 , month =. doi:10.1093/jjfinec/nbp001 , url =

    doi:10.1093/jjfinec/nbp001 2009

  25. [25]

    and Bollerslev, Tim and Diebold, Francis X

    Andersen, Torben G. and Bollerslev, Tim and Diebold, Francis X. and Labys, Paul , title =. Econometrica , volume =. doi:https://doi.org/10.1111/1468-0262.00418 , url =. https://onlinelibrary.wiley.com/doi/pdf/10.1111/1468-0262.00418 , year =

    arXiv

  26. [26]

    Journal of Business & Economic Statistics , volume=

    High-frequency lead-lag effects and cross-asset linkages: a multi-asset lagged adjustment model , author=. Journal of Business & Economic Statistics , volume=. 2021 , publisher=

    2021

  27. [27]

    Journal of Financial Econometrics , volume=

    Multifactor Timing with Deep Learning , author=. Journal of Financial Econometrics , volume=. 2026 , publisher=

    2026

  28. [28]

    Quantitative Finance , volume=

    Cross-impact of order flow imbalance in equity markets , author=. Quantitative Finance , volume=. 2023 , publisher=

    2023

  29. [29]

    Pattern Recognition , volume=

    Financial time series forecasting with multi-modality graph neural network , author=. Pattern Recognition , volume=. 2022 , publisher=

    2022

  30. [30]

    International Journal of Forecasting , volume=

    Forecasting realized volatility with spillover effects: Perspectives from graph neural networks , author=. International Journal of Forecasting , volume=. 2025 , publisher=

    2025

  31. [31]

    Proceedings of the 31st ACM international conference on information & knowledge management , pages=

    Temporal and heterogeneous graph neural network for financial time series prediction , author=. Proceedings of the 31st ACM international conference on information & knowledge management , pages=

  32. [32]

    Journal of Business & Economic Statistics , pages=

    The efficient tail hypothesis: An extreme value perspective on market efficiency , author=. Journal of Business & Economic Statistics , pages=. 2025 , publisher=

    2025

  33. [33]

    arXiv preprint arXiv:2505.08008 , year=

    Separation-based causal discovery for extremes , author=. arXiv preprint arXiv:2505.08008 , year=

    Pith/arXiv arXiv

  34. [34]

    Proceedings of the 6th ACM International Conference on AI in Finance , pages=

    DeltaLag: Learning Dynamic Lead-Lag Patterns in Financial Markets , author=. Proceedings of the 6th ACM International Conference on AI in Finance , pages=

  35. [35]

    Journal of the American Statistical Association , volume=

    Distribution-free predictive inference for regression , author=. Journal of the American Statistical Association , volume=. 2018 , publisher=

    2018

  36. [36]

    Advances in Neural Information Processing Systems , volume=

    Conformalized quantile regression , author=. Advances in Neural Information Processing Systems , volume=

  37. [37]

    Advances in Neural Information Processing Systems , volume=

    Adaptive conformal inference under distribution shift , author=. Advances in Neural Information Processing Systems , volume=

  38. [38]

    Journal of Machine Learning Research , volume=

    Conformal inference for online prediction with arbitrary distribution shifts , author=. Journal of Machine Learning Research , volume=

  39. [39]

    International Conference on Machine Learning , pages=

    Conformal prediction interval for dynamic time-series , author=. International Conference on Machine Learning , pages=. 2021 , organization=

    2021

  40. [40]

    International Conference on Machine Learning , pages=

    Improved online conformal prediction via strongly adaptive online learning , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  41. [41]

    International Conference on Machine Learning , pages=

    Adaptive conformal predictions for time series , author=. International Conference on Machine Learning , pages=. 2022 , organization=

    2022

  42. [42]

    Biometrika , volume=

    Localized conformal prediction: A generalized inference framework for conformal prediction , author=. Biometrika , volume=. 2023 , publisher=

    2023

  43. [43]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

    Conformal prediction with local weights: randomization enables robust guarantees , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2025 , publisher=

    2025

  44. [44]

    The Fourteenth International Conference on Learning Representations , year=

    Adaptive Conformal Prediction via Mixture-of-Experts Gating Similarity , author=. The Fourteenth International Conference on Learning Representations , year=

  45. [45]

    International Journal of Forecasting , volume=

    25 years of time series forecasting , author=. International Journal of Forecasting , volume=. 2006 , publisher=

    2006

  46. [46]

    International Journal of Machine Learning and Cybernetics , volume=

    Deep learning for time series forecasting: a survey , author=. International Journal of Machine Learning and Cybernetics , volume=. 2025 , publisher=

    2025

  47. [47]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Wpmixer: Efficient multi-resolution mixing for long-term time series forecasting , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  48. [48]

    Advances in Neural Information Processing Systems , volume=

    Conformal prediction for time-series forecasting with change points , author=. Advances in Neural Information Processing Systems , volume=

  49. [49]

    International Conference on Learning Representations , volume=

    Timemixer: Decomposable multiscale mixing for time series forecasting , author=. International Conference on Learning Representations , volume=

  50. [50]

    The Twelfth International Conference on Learning Representations , year=

    Pathformer: Multi-scale Transformers with Adaptive Pathways for Time Series Forecasting , author=. The Twelfth International Conference on Learning Representations , year=

  51. [51]

    Journal of Business & Economic Statistics , volume=

    The efficient tail hypothesis: An extreme value perspective on market efficiency , author=. Journal of Business & Economic Statistics , volume=. 2026 , publisher=

    2026

  52. [52]

    Econometrica , volume=

    Modeling and forecasting realized volatility , author=. Econometrica , volume=. 2003 , publisher=

    2003

  53. [53]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

    Econometric analysis of realized volatility and its use in estimating stochastic volatility models , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2002 , publisher=

    2002

  54. [54]

    Journal of Econometrics , volume=

    Does anything beat 5-minute RV? A comparison of realized measures across multiple asset classes , author=. Journal of Econometrics , volume=. 2015 , publisher=

    2015

  55. [55]

    Journal of financial econometrics , volume=

    A simple approximate long-memory model of realized volatility , author=. Journal of financial econometrics , volume=. 2009 , publisher=

    2009

  56. [56]

    International Journal of Forecasting , volume=

    Forecasting realized volatility of agricultural commodity futures with infinite Hidden Markov HAR models , author=. International Journal of Forecasting , volume=. 2022 , publisher=

    2022

  57. [57]

    The Eleventh International Conference on Learning Representations , year=

    A Time Series is Worth 64 Words: Long-term Forecasting with Transformers , author=. The Eleventh International Conference on Learning Representations , year=

  58. [58]

    The Thirteenth International Conference on Learning Representations , year=

    Time-MoE: Billion-Scale Time Series Foundation Models with Mixture of Experts , author=. The Thirteenth International Conference on Learning Representations , year=

  59. [59]

    and Hyndman, Rob J

    Godahewa, Rakshitha and Bergmeir, Christoph and Webb, Geoffrey I. and Hyndman, Rob J. and Montero-Manso, Pablo , title =. 2020 , publisher =. doi:10.5281/zenodo.4656144 , url =

    work page doi:10.5281/zenodo.4656144 2020

  60. [60]

    Advances in neural information processing systems , volume=

    Conformalized quantile regression , author=. Advances in neural information processing systems , volume=

  61. [61]

    International conference on machine learning , pages=

    Adaptive conformal predictions for time series , author=. International conference on machine learning , pages=. 2022 , organization=

    2022

  62. [62]

    Adaptive Conformal Prediction via Mixture-of-Experts Gating Similarity , author=