Pith. sign in

REVIEW 5 major objections 5 minor 2 cited by

Feature-engineered Gradient Boosting beats time-series foundation models on volatile or sparse data; foundation models excel only in stable, trend-driven settings, and XAI plus causal ratings explain when and why.

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 →

On four public datasets, Gradient Boosting with hand-built features beat Chronos, Llama, and ARIMA on most accuracy metrics, while Chronos only led on financial sMAPE.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Useful exploratory benchmark with an overreaching headline: the model-family comparison is confounded by feature engineering and Chronos pretraining overlap. the 5 major comments →

arxiv 2508.20437 v1 pith:2IOI265A submitted 2025-08-28 cs.LG cs.AI

On Identifying Why and When Foundation Models Perform Well on Time-Series Forecasting Using Automated Explanations and Rating

classification cs.LG cs.AI
keywords time-series forecastingfoundation modelsexplainable AISHAPLIMEcausal ratingfeature engineeringmodel selection
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 reading

This paper asks when and why time-series forecasting models succeed or fail, and answers with a four-family comparison—statistical ARIMA, feature-engineered Gradient Boosting, and the foundation models Chronos and Llama—across finance, energy, pedestrian-traffic, and car-parts sales data. Its central claim is that domain knowledge, not model scale, decides performance: Gradient Boosting posts the best MASE in every domain, while Chronos leads only in finance, a stable, trend-driven setting. To explain why, the paper pairs standard attribution tools (SHAP, LIME) with a causal rating framework whose two diagnostics separate error differences across series (ATE) from error differences across calendar periods (WRS). If the claim is right, practitioners get a concrete rule: prefer engineered features on volatile or sparse data, and reserve foundation models for smooth, trend-driven series.

Core claim

The paper tries to establish that a forecasting model's success depends on the fit between its inductive bias and the data's statistical structure, and that the decisive ingredient is domain knowledge, not model scale. On four datasets, Gradient Boosting with hand-crafted features posts the lowest MASE in every domain, while foundation models win only in favorable settings: Chronos takes the best sMAPE in trending finance, and Llama needs fine-tuning to become competitive. The paper further claims that combining SHAP and LIME attributions with Rating-Driven Explanations (ATE, WRS) reveals why: feature-based models exploit genuinely predictive domain features, foundation-model surrogates fall

What carries the argument

The central mechanism is Rating-Driven Explanations (RDE), a causal framework that treats a protected attribute Z (series identity, month, or day-of-week) as a potential confounder between a treatment T and the model's outcome O. It yields two diagnostics: the Average Treatment Effect (ATE), estimating how much forecast errors differ across series after deconfounding, and the Weighted Rejection Score (WRS), measuring how error distributions differ across calendar groups by weighted t-tests. These ratings pair with conventional attributions—TreeSHAP for Gradient Boosting, a LightGBM surrogate carrying TsSHAP for Chronos, and a LIME variant (TS-MULE) with uniform segmentation—so that what driv

Load-bearing premise

The benchmark is treated as a fair test of model families, but Gradient Boosting receives hand-crafted domain features while foundation models receive raw values, and Chronos was pretrained on two of the four datasets—so the observed ordering is partly built into the experimental design.

What would settle it

Run the same comparison with the engineered features (lags, Fourier terms, rolling statistics) also given to Chronos and Llama, or on held-out datasets Chronos never saw in pretraining; if a foundation model then matches or beats Gradient Boosting's MASE on power or car-parts data, the paper's central claim fails. A second check: fine-tune a foundation model on the engineered features and see whether the gap to Gradient Boosting largely disappears.

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

If this is right

  • For volatile or sparse domains (power, car parts), feature-engineered Gradient Boosting is the safer default: it won every accuracy metric and showed low variance across series.
  • Foundation models such as Chronos are best reserved for stable, trend-driven settings; in finance it posted the best sMAPE and the lowest series-consistency error (ATE).
  • Average accuracy can mask instability: a model with the best MASE can still carry high WRS, meaning errors vary by month or weekday, so ratings should accompany point metrics in model selection.
  • Explanations of foundation models through surrogates show they revert to statistical baselines (expanding means, standard deviations) when patterns get complex, flagging where they will struggle before deployment.
  • Fine-tuning a general-purpose LLM (Llama) cuts errors dramatically but does not make it competitive with feature-engineered or specialized models, so the cost buys robustness, not leadership.

Where Pith is reading between the lines

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

  • The comparison is not controlled: Gradient Boosting receives engineered features while the foundation models receive raw values, and Chronos was pretrained on two of the four datasets. Feeding the same features to Chronos or Llama, or testing on data truly unseen in pretraining, would be the decisive check on the paper's rule.
  • The ATE/WRS pair could serve as a deployment screen: compute series-consistency and calendar-sensitivity ratings on validation data, then treat high WRS as a warning that point accuracy overstates reliability.
  • Because the Chronos explanations come from a LightGBM surrogate, the claim that foundation models 'default to statistical aggregates' deserves verification with faithfulness checks beyond surrogate RMSE, such as direct attribution or perturbation tests.
  • A concrete extension: if domain knowledge is the critical factor, then fine-tuning a foundation model on the same engineered features should close most of the gap to Gradient Boosting in car-parts and power data—an experiment that would separate model scale from feature engineering.
Share X Bluesky LinkedIn Reddit HN

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

5 major / 5 minor

Summary. The paper evaluates four forecasting approaches (ARIMA, gradient boosting, Chronos, and Llama with and without fine-tuning) on four public time-series datasets spanning finance, energy, pedestrian counts, and car-part sales. It combines standard accuracy metrics (MASE, sMAPE) with SHAP/LIME attributions and the authors' Rating-Driven Explanations (RDE) framework, which computes ATE and WRS scores to quantify error sensitivity across series and time groups. The central claim is that feature-engineered boosting consistently outperforms foundation models in volatile or sparse domains, while foundation models excel only in stable, trend-driven contexts such as finance; the paper further argues that domain-knowledge integration is the most critical factor in forecasting success. The evidence in Table 3 shows GBoost achieves the best MASE in all four domains, while Chronos achieves the best sMAPE only in finance; Llama variants do not lead on any metric. The paper also derives qualitative explanations from SHAP/LIME and RDE scores, concluding that RDE fills the gap between accuracy and feature attributions.

Significance. If the central claim were established, the paper would offer practitioners a simple model-selection rule: use feature-engineered boosting for irregular or sparse series and foundation models only for smooth trend data. This would be a useful and actionable finding, especially because the paper uses public datasets and standard metrics, and it explicitly credits prior work on gradient boosting (Elsayed et al.). The attempt to connect forecasting accuracy with robustness-oriented explanations (RDE) is a worthwhile direction, and the authors are candid about limitations (e.g., no comparison to other fairness-aware XAI methods, no synthetic ground-truth validation). However, the current empirical design does not isolate model family from feature engineering, and the pretraining contamination in two of the four domains further weakens the comparative claim. The paper is therefore of interest to the forecasting and XAI communities, but the headline conclusion requires substantially stronger evidence.

major comments (5)
  1. [§4.3 and Table 2] The central comparison is confounded: GBoost receives hand-crafted domain features (time-based, Fourier terms, domain statistics, tailored lags) while Chronos and Llama receive only raw values. The conclusion in §6.1 that 'domain knowledge integration remains the most critical factor' is therefore built into the experimental design, not demonstrated by it. To support the claim that model family matters, the authors should run a controlled comparison—e.g., give the same engineered feature set to the foundation models or train GBoost on raw lagged values only—and show that the ordering persists.
  2. [Table 3 footnote and §4.1] The footnote states that Chronos was exposed to the pedestrian and car parts datasets during pretraining. These are exactly the two sparse/volatile domains in which the paper claims GBoost 'consistently outperform[s]' foundation models. The contamination undermines the fairness of the comparison in those domains. The authors acknowledge the exposure but dismiss it as 'an experimental setup different than our own'; that does not remove the possibility of data leakage. The claims about pedestrian and car parts should be re-made using truly held-out datasets or otherwise justified.
  3. [Table 3 and §5.1] The finance claim is metric-dependent. Chronos has the best sMAPE (5.30) but GBoost has the best MASE (4.32). The abstract and §6.1 state that foundation models 'excel only in stable or trend-driven contexts (e.g., finance)', but this is only true for one of the two metrics used. The paper should either commit to a primary metric (e.g., MASE, which favors GBoost in all four domains), or explicitly qualify the finance conclusion as sMAPE-only.
  4. [§4.5 and Table B.1] The interpretability analysis for Chronos is performed on a LightGBM surrogate, not on Chronos itself. Table B.1 shows the surrogate can have substantially different accuracy than Chronos (e.g., pedestrian MASE 0.71 vs 1.47). The qualitative claims about 'Chronos's surrogate reveals a tendency to default to statistical aggregates' and the related 'why' explanations are therefore claims about the surrogate, not about Chronos. The paper should clearly separate these levels, and ideally validate surrogate fidelity (e.g., correlation of attributions on held-out samples) before drawing conclusions about Chronos's behavior.
  5. [§5.1 and Table 3] No significance tests or paired comparisons are provided. With only six finance series, the MASE difference between GBoost (4.32±3.00) and Chronos (5.48±5.44) may not be statistically meaningful, and the large standard deviations in car parts cast doubt on the 'consistently' claim. The authors should report paired tests (e.g., Wilcoxon signed-rank or Diebold-Mariano) across series, along with effect sizes, for each domain and metric.
minor comments (5)
  1. [Throughout] The model name is rendered inconsistently: 'LlaMA', 'LLaMA', 'Llama', and 'Llama-FT' appear in different sections. Please unify notation (e.g., 'Llama' and 'Llama-FT').
  2. [Table 3] The footnote symbol '*' is used both to mark Chronos pretraining overlap and as a footnote marker; this is confusing. Use a distinct symbol and clarify the footnote text.
  3. [§4.4] The MASE formula uses the denominator computed from the training set (m and s), but the notation is not fully defined: please state that m is the training length and s is the seasonal period, and clarify what happens when s=1 (as in car parts).
  4. [§4.2] Llama and Llama-FT use direct forecasting on Power and Pedestrian but autoregressive rollout on Finance and Car Parts. This inconsistency is not reflected in the analysis; a sentence acknowledging the potential effect on results would help.
  5. [Appendix C] Figures C.6 and C.7 captions label perturbation methods, but the text sometimes uses 'zero replacement' interchangeably with 'zero perturbation'. Define the perturbation strategies once and use consistent terminology.

Circularity Check

0 steps flagged

No circular derivation: the empirical comparison is self-contained; self-citations to the RDE framework are acknowledged and not load-bearing.

full rationale

The paper's central claims are empirical rankings from Table 3, computed on external public datasets (Monash, Yahoo Finance, Tantiv4) with standard metrics (sMAPE, MASE). Gradient Boosting receives hand-engineered features while Chronos and Llama receive raw values, and Table 3's footnote discloses that Chronos was exposed to pedestrian and car parts during pretraining. These are confounding/validity concerns that weaken the causal conclusion 'domain knowledge integration remains the most critical factor,' but they are not circular: the outcome is not defined into existence by the inputs, and no fitted parameter is renamed as a prediction. No equation in the paper defines X in terms of Y. The RDE framework is cited from the authors' prior work (Lakkaraju, Valluru, and Srivastava 2025), but the paper re-specifies ATE and WRS in Section 4.6 and Appendix A, and explicitly states in Section 6.1 that RDE was not compared against other fairness-oriented XAI approaches or evaluated on synthetic benchmarks with ground-truth fairness properties. Thus the self-citations identify provenance rather than supply the argument. The surrogate LightGBM used for Chronos SHAP analysis is a standard surrogate-explanation technique, not a circular prediction. Accordingly, there is no exhibited reduction of a claimed result to its own inputs, and the appropriate score is low, reflecting only minor, non-load-bearing self-citation.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The central claim rests on an experimental setup, not on a derivation. The ledger lists the main choices that could change the outcome: per-domain feature engineering, window sizes, hyperparameters, seasonal periods used in MASE, the causal graph underlying RDE, and the surrogate used for Chronos explanations. None of these are fitted constants in a mathematical derivation, but they are load-bearing settings.

free parameters (7)
  • ARIMA orders (p,d,q) and seasonal orders = selected per series by grid search
    ARIMA performance depends on these choices; the grid is constrained by domain-specific maxima and seasonality s inferred from autocorrelation or defaults, so comparisons are tuned.
  • GBoost engineered feature set = not enumerated
    The feature set (lags, moving averages, Fourier terms, domain statistics) is the central reason GBoost wins; it is hand-selected per domain and not fully specified.
  • Context length C and horizon H per domain = Finance C=20,H=5; Power C=1440,H=360; Pedestrian C=72,H=18; Car C=8,H=2
    Window sizes are chosen by the authors and affect which models can perform well; no sensitivity analysis is given.
  • LightGBM hyperparameters = 500 estimators, learning rate 0.05, L1 loss
    Chosen by hand; no tuning or ablation.
  • Llama fine-tuning hyperparameters = r=16, alpha=32, dropout=0.05, NF4, paged AdamW
    Chosen for LoRA fine-tuning; not varied.
  • MASE seasonal periods s = power 1440, pedestrian 24, finance 5, car 1
    The seasonal period determines MASE denominators and changes the reported values, especially for car parts where some denominators can make MASE infinite.
  • RDE confidence interval weights = 1, 0.8, 0.6 for 95%, 75%, 60% CIs
    Used to compute WRS; arbitrary weighting scheme from prior RDE work.
axioms (5)
  • domain assumption The causal graph in Figure 1 correctly identifies treatment T, outcome O, and protected attribute Z with no unmeasured confounding.
    The ATE/WRS ratings are interpreted causally; if the graph is wrong, the ratings do not measure what the paper claims. The graph and G-computation identification are adopted from prior RDE work without validation here.
  • domain assumption The LightGBM surrogate is a faithful approximation of Chronos for SHAP analysis.
    Section 4.5 uses a surrogate because Chronos is not directly explainable; fidelity is said to be checked with RMSE, but the RMSE values are not reported, so the reader cannot verify that SHAP values describe Chronos rather than the surrogate.
  • domain assumption MASE and sMAPE aggregated across heterogeneous series are valid ordering criteria.
    The central conclusions compare models by mean scores, but no significance tests are run, and the car parts domain produces infinite MASE for several models, which is not handled statistically.
  • domain assumption The 80-20 split and windowed rollout protocol are fair for all models.
    Iterative single-step forecasting for ARIMA/GBoost differs from windowed direct forecasting for Llama and Chronos, so the task is not identical across models.
  • domain assumption Student t-test on residual distributions is valid for WRS.
    Residuals are likely autocorrelated and non-normal, and multiple tests across group pairs are weighted but not adjusted for multiple comparisons or autocorrelation.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of On Identifying Why and When Foundation Models Perform Well on Time-Series Forecasting Using Automated Explanations and Rating." pith.science (2026). https://pith.science/paper/2IOI265A

@misc{pith2026250820437,
  author       = {Pith},
  title        = {Pith review of: On Identifying Why and When Foundation Models Perform Well on Time-Series Forecasting Using Automated Explanations and Rating},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2IOI265A}},
  note         = {Machine review of arXiv:2508.20437}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Time-series forecasting models (TSFM) have evolved from classical statistical methods to sophisticated foundation models, yet understanding why and when these models succeed or fail remains challenging. Despite this known limitation, time series forecasting models are increasingly used to generate information that informs real-world actions with equally real consequences. Understanding the complexity, performance variability, and opaque nature of these models then becomes a valuable endeavor to combat serious concerns about how users should interact with and rely on these models' outputs. This work addresses these concerns by combining traditional explainable AI (XAI) methods with Rating Driven Explanations (RDE) to assess TSFM performance and interpretability across diverse domains and use cases. We evaluate four distinct model architectures: ARIMA, Gradient Boosting, Chronos (time-series specific foundation model), Llama (general-purpose; both fine-tuned and base models) on four heterogeneous datasets spanning finance, energy, transportation, and automotive sales domains. In doing so, we demonstrate that feature-engineered models (e.g., Gradient Boosting) consistently outperform foundation models (e.g., Chronos) in volatile or sparse domains (e.g., power, car parts) while providing more interpretable explanations, whereas foundation models excel only in stable or trend-driven contexts (e.g., finance).

Figures

Figures reproduced from arXiv: 2508.20437 by Biplav Srivastava, John Aydin, Kausik Lakkaraju, Michael Widener.

Figure 1
Figure 1. Figure 1: Proposed Generalized Causal Graph. The RDE framework (Lakkaraju, Valluru, and Srivastava 2025) integrates causal reasoning with metrics to explain model behavior under different conditions. The generalized causal graph used for our experiments is shown in [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: LIME plots for ARIMA model in finance using 20 [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: SHAP values for the Gradient Boosting model (fi [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: LIME plots for Chronos model in finance using [PITH_FULL_IMAGE:figures/full_fig_p007_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.

Forward citations

Cited by 2 Pith papers

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

  1. The Spectrum Is Not Enough: When Context Helps Time-Series Forecasting

    cs.LG 2026-07 conditional novelty 7.0

    Spectral indices cannot predict when context helps time-series forecasting; a coverage-deficit diagnostic measuring analog-over-linear gain can.

  2. The Spectrum Is Not Enough: When Context Helps Time-Series Forecasting

    cs.LG 2026-07 conditional novelty 6.0

    Power-spectrum indices cannot predict whether added context helps forecasting; a coverage-deficit diagnostic based on analog-vs-linear gain separates beyond-spectrum context value.

Reference graph

Works this paper leans on

59 extracted references · 41 canonical work pages · cited by 1 Pith paper · 3 internal anchors

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Abbasimehr, H.; and Paki, R. 2022. Improving time series forecasting using LSTM and attention models. Journal of Ambient Intelligence and Humanized Computing, 13(1): 673--691

  4. [4]

    Agapitos, A.; Brabazon, A.; and O’Neill, M. 2017. Regularised gradient boosting for financial time-series modelling. Computational Management Science, 14(3): 367--391

  5. [5]

    L.; Zippo, A

    Alfeo, A. L.; Zippo, A. G.; Catrambone, V.; Cimino, M. G.; Toschi, N.; and Valenza, G. 2023. From local counterfactuals to global feature importance: efficient, robust, and model-agnostic explanations for brain connectivity networks. Computer Methods and Programs in Biomedicine, 236: 107550

  6. [6]

    Alikhademi, K.; Richardson, B.; Drobina, E.; and Gilbert, J. E. 2021. Can explainable AI explain unfairness? A framework for evaluating explainable AI. arXiv preprint arXiv:2106.07483

  7. [7]

    F.; Stella, L.; Turkmen, C.; Zhang, X.; Mercado, P.; Shen, H.; Shchur, O.; Rangapuram, S

    Ansari, A. F.; Stella, L.; Turkmen, C.; Zhang, X.; Mercado, P.; Shen, H.; Shchur, O.; Rangapuram, S. S.; Arango, S. P.; Kapoor, S.; et al. 2024. Chronos: Learning the language of time series. arXiv preprint arXiv:2403.07815

  8. [8]

    B.; D \' az-Rodr \' guez, N.; Del Ser, J.; Bennetot, A.; Tabik, S.; Barbado, A.; Garc \' a, S.; Gil-L \'o pez, S.; Molina, D.; Benjamins, R.; et al

    Arrieta, A. B.; D \' az-Rodr \' guez, N.; Del Ser, J.; Bennetot, A.; Tabik, S.; Barbado, A.; Garc \' a, S.; Gil-L \'o pez, S.; Molina, D.; Benjamins, R.; et al. 2020. Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI. Information fusion, 58: 82--115

  9. [9]

    Arsenault, P.-D.; Wang, S.; and Patenaude, J.-M. 2025. A survey of explainable artificial intelligence (XAI) in financial time series forecasting. ACM Computing Surveys, 57(10): 1--37

  10. [10]

    Bernagozzi, M.; Srivastava, B.; Rossi, F.; and Usmani, S. 2021 a . Gender Bias in Online Language Translators: Visualization, Human Perception, and Bias/Accuracy Tradeoffs. IEEE Internet Computing, 25(5): 53--63

  11. [11]

    Bernagozzi, M.; Srivastava, B.; Rossi, F.; and Usmani, S. 2021 b . Vega: a virtual environment for exploring gender bias vs. accuracy trade-offs in ai translation services. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, 15994--15996

  12. [12]

    O.; Pfister, T.; Zheng, Y.; Ye, W.; and Liu, Y

    Cao, D.; Jia, F.; Arik, S. O.; Pfister, T.; Zheng, Y.; Ye, W.; and Liu, Y. 2023. Tempo: Prompt-based generative pre-trained transformer for time series forecasting. arXiv preprint arXiv:2310.04948

  13. [13]

    B.; \.I can, \"O .; and Bulut, E

    C elik, T. B.; \.I can, \"O .; and Bulut, E. 2023. Extending machine learning prediction capabilities by explainable AI in financial time series prediction. Applied Soft Computing, 132: 109876

  14. [14]

    City of Melbourne . 2024. Pedestrian Counting System. https://www.melbourne.vic.gov.au/pedestrian-counting-system. Accessed: 2024-08-10

  15. [15]

    Das, A.; Kong, W.; Sen, R.; and Zhou, Y. 2023. A decoder-only foundation model for time-series forecasting. arXiv preprint arXiv:2310.10688

  16. [16]

    H.; Gifford, W

    Ekambaram, V.; Jati, A.; Dayama, P.; Mukherjee, S.; Nguyen, N. H.; Gifford, W. M.; Reddy, C.; and Kalagnanam, J. 2024. Tiny Time Mixers (TTMs): Fast Pre-trained Models for Enhanced Zero/Few-Shot Forecasting of Multivariate Time Series. CoRR

  17. [17]

    Elsayed, S.; Thyssens, D.; Rashed, A.; Schmidt - Thieme, L.; and Jomaa, H. S. 2021. Do We Really Need Deep Learning Models for Time Series Forecasting? CoRR, abs/2101.02118

  18. [18]

    Garza, A.; and Mergenthaler-Canseco, M. 2023. TimeGPT-1. arXiv preprint arXiv:2310.03589

  19. [19]

    I.; Hyndman, R

    Godahewa, R.; Bergmeir, C.; Webb, G. I.; Hyndman, R. J.; and Montero-Manso, P. 2021. Monash Time Series Forecasting Archive. arXiv:2105.06643

  20. [20]

    Goswami, M.; Szafer, K.; Choudhry, A.; Cai, Y.; Li, S.; and Dubrawski, A. 2024. Moment: A family of open time-series foundation models. arXiv preprint arXiv:2402.03885

  21. [21]

    Greenwell, B. M. 2017. pdp: An R Package for Constructing Partial Dependence Plots . The R Journal , 9(1): 421--436

  22. [22]

    Gruver, N.; Finzi, M.; Qiu, S.; and Wilson, A. G. 2023. Large language models are zero-shot time series forecasters. Advances in Neural Information Processing Systems, 36

  23. [23]

    R.; Mueller, S

    Hoffman, R. R.; Mueller, S. T.; Klein, G.; Jalaeian, M.; and Tate, C. 2023. Explainable ai: roles and stakeholders, desirements and challenges. Frontiers in Computer Science, 5: 1117848

  24. [24]

    Hyndman, R.; Koehler, A.; Ord, K.; and Snyder, R. 2008. Forecasting with exponential smoothing: the state space approach. Springer

  25. [25]

    Y.; Shi, X.; Chen, P.-Y.; Liang, Y.; Li, Y.-F.; Pan, S.; et al

    Jin, M.; Wang, S.; Ma, L.; Chu, Z.; Zhang, J. Y.; Shi, X.; Chen, P.-Y.; Liang, Y.; Li, Y.-F.; Pan, S.; et al. 2023. Time-llm: Time series forecasting by reprogramming large language models. arXiv preprint arXiv:2310.01728

  26. [26]

    Lakkaraju, K.; Gupta, A.; Srivastava, B.; Valtorta, M.; and Wu, D. 2023. The Effect of Human v/s Synthetic Test Data and Round-Tripping on Assessment of Sentiment Analysis Systems for Bias. In 2023 5th IEEE International Conference on Trust, Privacy and Security in Intelligent Systems and Applications (TPS-ISA), 380--389. Los Alamitos, CA, USA: IEEE Compu...

  27. [27]

    L.; Zeng, Z.; Srivastava, B.; and Valtorta, M

    Lakkaraju, K.; Kaur, R.; Zehtabi, P.; Patra, S.; Valluru, S. L.; Zeng, Z.; Srivastava, B.; and Valtorta, M. 2025. On Creating a Causally Grounded Usable Rating Method for Assessing the Robustness of Foundation Models Supporting Time Series. arXiv preprint arXiv:2502.12226

  28. [28]

    Lakkaraju, K.; Kaur, R.; Zeng, Z.; Zehtabi, P.; Patra, S.; Srivastava, B.; and Valtorta, M. 2024. Rating Multi-Modal Time-Series Forecasting Models (MM-TSFM) for Robustness Through a Causal Lens. arXiv preprint arXiv:2406.12908

  29. [29]

    Lakkaraju, K.; Srivastava, B.; and Valtorta, M. 2024. Rating Sentiment Analysis Systems for Bias Through a Causal Lens. IEEE Transactions on Technology and Society, 1--1

  30. [30]

    L.; and Srivastava, B

    Lakkaraju, K.; Valluru, S. L.; and Srivastava, B. 2025. Holistic Explainable AI (H-XAI): Extending Transparency Beyond Developers in AI-Driven Decision Making. arXiv:2508.05792

  31. [31]

    Li, Y.; Zhu, Z.; Kong, D.; Han, H.; and Zhao, Y. 2019. EA-LSTM: Evolutionary attention-based LSTM for time series prediction. Knowledge-Based Systems, 181: 104785

  32. [32]

    M.; and Lee, S.-I

    Lundberg, S. M.; and Lee, S.-I. 2017. A unified approach to interpreting model predictions. Advances in neural information processing systems, 30

  33. [33]

    K.; and Kalita, J

    Mersha, M.; Lam, K.; Wood, J.; AlShami, A. K.; and Kalita, J. 2024. Explainable artificial intelligence: A survey of needs, techniques, applications, and future direction. Neurocomputing, 599: 128111

  34. [34]

    P.; Adji, T

    Minarso, V. P.; Adji, T. B.; and Setiawan, N. A. 2022. Hybrid SVD-ARIMA Method for Sales Forecasting with Sparse Data on E-Commerce Products. In 2022 IEEE International Conference on Cybernetics and Computational Intelligence (CyberneticsCom), 387--392

  35. [35]

    E.; Higdon, B

    Mokhtari, K. E.; Higdon, B. P.; and Ba s ar, A. 2019. Interpreting financial time series with SHAP values. In Proceedings of the 29th annual international conference on computer science and software engineering, 166--172

  36. [36]

    J.; Appajigowda, C.; Srivastava, B.; and Johri, L

    Muppasani, B.; Anand, C. J.; Appajigowda, C.; Srivastava, B.; and Johri, L. 2023. A dataset and baseline approach for identifying usage states from non-intrusive power sensing with MiDAS IoT-based sensors. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, 15545--15550

  37. [37]

    Pan, Z.; Jiang, Y.; Garg, S.; Schneider, A.; Nevmyvaka, Y.; and Song, D. 2024. IP-LLM: Semantic space informed prompt learning with LLM for time series forecasting. In Forty-first International Conference on Machine Learning

  38. [38]

    R.; Khorasani, A.; Adamopoulos, G.; Bhagwatkar, R.; Bilo s , M.; Ghonia, H.; Hassen, N

    Rasul, K.; Ashok, A.; Williams, A. R.; Khorasani, A.; Adamopoulos, G.; Bhagwatkar, R.; Bilo s , M.; Ghonia, H.; Hassen, N. V.; Schneider, A.; et al. 2023. Lag-llama: Towards foundation models for time series forecasting. arXiv preprint arXiv:2310.08278

  39. [39]

    C.; Jati, A.; Mukherjee, S.; Aggarwal, N.; Sarpatwar, K.; Ganapavarapu, G.; and Vaculin, R

    Raykar, V. C.; Jati, A.; Mukherjee, S.; Aggarwal, N.; Sarpatwar, K.; Ganapavarapu, G.; and Vaculin, R. 2023. TsSHAP: Robust model agnostic feature-based explainability for time series forecasting. arXiv preprint arXiv:2303.12316

  40. [40]

    Why Should I Trust You?

    Ribeiro, M. T.; Singh, S.; and Guestrin, C. 2016. "Why Should I Trust You?": Explaining the Predictions of Any Classifier. arXiv:1602.04938

  41. [41]

    Robins, J. 1986. A new approach to causal inference in mortality studies with a sustained exposure period—application to control of the healthy worker survivor effect. Mathematical modelling, 7(9-12): 1393--1512

  42. [42]

    Rubin, D. B. 1974. Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of educational Psychology, 66(5): 688

  43. [43]

    Schlegel, U.; Arnout, H.; El-Assady, M.; Oelke, D.; and Keim, D. A. 2019. Towards a rigorous evaluation of XAI methods on time series. In 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), 4197--4201. IEEE

  44. [44]

    TS-MULE: Local Interpretable Model-Agnostic Explanations for Time Series Forecast Models

    Schlegel, U.; Lam, D. V.; Keim, D. A.; and Seebacher, D. 2021. TS-MULE: Local Interpretable Model-Agnostic Explanations for Time Series Forecast Models. arXiv:2109.08438

  45. [45]

    Shih, S.-Y.; Sun, F.-K.; and Lee, H.-y. 2019. Temporal pattern attention for multivariate time series forecasting. Machine Learning, 108: 1421--1441

  46. [46]

    H.; and Stoffer, D

    Shumway, R. H.; and Stoffer, D. S. 2017. ARIMA models. In Time series analysis and its applications: with R examples, 75--163. Springer

  47. [47]

    Sivill, T.; and Flach, P. 2022. LIMESegment: Meaningful, Realistic Time Series Explanations. In Camps-Valls, G.; Ruiz, F. J. R.; and Valera, I., eds., Proceedings of The 25th International Conference on Artificial Intelligence and Statistics, volume 151 of Proceedings of Machine Learning Research, 3418--3433. PMLR

  48. [48]

    M.; and Speed, T

    Splawa-Neyman, J.; Dabrowska, D. M.; and Speed, T. P. 1990. On the application of probability theory to agricultural experiments. Essay on principles. Section 9. Statistical Science, 465--472

  49. [49]

    Srivastava, B.; and Rossi, F. 2018. Towards Composable Bias Rating of AI Systems. In 2018 AI Ethics and Society Conference (AIES 2018), New Orleans, Louisiana, USA, Feb 2-3

  50. [50]

    Srivastava, B.; and Rossi, F. 2020. Rating AI Systems for Bias to Promote Trustable Applications. In IBM Journal of Research and Development

  51. [51]

    Student. 1908. The probable error of a mean. Biometrika, 1--25

  52. [52]

    Talukder, S.; Yue, Y.; and Gkioxari, G. 2024. TOTEM: TOkenized Time Series EMbeddings for General Time Series Analysis. arXiv preprint arXiv:2402.16412

  53. [53]

    U.; Islam, S.; and Khan, R

    Tasin, I.; Nabil, T. U.; Islam, S.; and Khan, R. 2023. Diabetes prediction using machine learning and explainable AI techniques. Healthcare technology letters, 10(1-2): 1--10

  54. [54]

    O.; Ildiz, M

    Tire, K.; Taga, E. O.; Ildiz, M. E.; and Oymak, S. 2024. Retrieval augmented time series forecasting. arXiv preprint arXiv:2411.08249

  55. [55]

    Woo, G.; Liu, C.; Kumar, A.; Xiong, C.; Savarese, S.; and Sahoo, D. 2024. Unified training of universal time series forecasting transformers. arXiv preprint arXiv:2402.02592

  56. [56]

    Xiao, M.; Jiang, Z.; Qian, L.; Chen, Z.; He, Y.; Xu, Y.; Jiang, Y.; Li, D.; Weng, R.-L.; Peng, M.; et al. 2025. Retrieval-augmented large language models for financial time series forecasting. arXiv preprint arXiv:2502.05878

  57. [57]

    Yu, X.; Chen, Z.; Ling, Y.; Dong, S.; Liu, Z.; and Lu, Y. 2023. Temporal Data Meets LLM--Explainable Financial Time Series Forecasting. arXiv preprint arXiv:2306.11025

  58. [58]

    Zhang, Y.; Petrosian, O.; Liu, J.; Ma, R.; and Krinkin, K. 2022. FI-SHAP: explanation of time series forecasting and improvement of feature engineering based on boosting algorithm. In Proceedings of SAI intelligent systems conference, 745--758. Springer

  59. [59]

    Zhou, T.; Niu, P.; Sun, L.; Jin, R.; et al. 2023. One fits all: Power general time series analysis by pretrained lm. Advances in neural information processing systems, 36: 43322--43355

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.