Pith. sign in

REVIEW 4 major objections 6 minor 79 references

Evaluation of a Foundational Model and Stochastic Models for Forecasting Sporadic or Spiky Production Outages of High-Performance Machine Learning Services

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A pretrained time-series foundational model beats classical stochastic models at forecasting rare, spiky production outages.

desk verdict Genuine first application of TimesFM to sparse outage counts, but the headline FM gain is not out-of-sample as presented and the root-cause label window needs reconciling. read the letter →

arxiv 2507.01067 v1 pith:PYE5XJKG submitted 2025-06-30 cs.LG cs.AIcs.DCcs.SYeess.SY

Keun Soo Yim This is my paper
classification cs.LGcs.AIcs.DCcs.SYeess.SY
keywords timeseriesforecastingfoundationalmodelTimesFMproductionoutagessoftwarereliabilityrootcauseanalysisextremeeventsspiky
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

On seven years of monthly production-outage counts from planet-scale machine-learning services, the paper claims that a pre-trained time-series foundational model (TimesFM) forecasts total outages more accurately than previous-value, moving-average, and autoregressive baselines: with a seven-month lookback it reports 1.3% lower MAE, 12.4% lower MSE, and 6.4% lower RMSE than the second-best model. Broken down by eight root-cause types, the foundational model has the lowest errors on most types, but no single model or lookback length wins for every type; moving averages win when outages are stationary or periodic peaks occur, while the foundational model best tracks smoother, longer spikes. Using iterated multi-step forecasts up to twelve months ahead with optimized parameters, the best model estimates the year-end outage count of one root-cause type with errors below 6%, which the paper uses to validate the effect of mitigation efforts. The paper frames the result as evidence that one-parameter software-reliability growth models are insufficient for agile services and that pre-trained foundational models can be applied to sparse, spiky reliability data.

What carries the argument

The load-bearing object is the pre-trained TimesFM foundational model: a decoder-only transformer that turns time series into patches (analogous to tokens in language models), is trained on large-scale and synthetic time series, and adapts to varying context lengths through its patching and data-granularity inputs. The paper optimizes it with a log1p transformation of inputs, exponentiation of outputs, flooring negative predictions at zero, and optional fine-tuning, and compares it against previous-value, moving-average, and autoregressive baselines. The argument also rests on a characterization step: fitting 56 standard distributions to the outage CDFs shows that one-parameter models fail while two-parameter beta and wrapped Cauchy distributions fit, which motivates replacing software-reliability growth models with time-series forecasting. Normalized mean absolute error, mean squared error, and root mean squared error over monthly outage counts carry the quantitative comparison.

What would settle it

Re-run the root-cause evaluation using only the 82% of labeled outages with a single root cause, or use an independently re-labeled sample of the postmortems; if the foundational model's per-type error margins shrink or reverse under cleaner labels, the claimed per-type superiority is an artifact of label noise rather than of the model.

Watch

Extended reading notes

Core claim

In the paper's own terms, the discovery is that the pre-trained decoder-only transformer TimesFM, when optimized with a logarithmic transformation and a floor at zero, is a more accurate forecaster for monthly production outages of large machine-learning services than classical stochastic models over the full six-year test period, with 1.3-12.4% accuracy gains over the best classical model depending on the error metric. For root-cause-specific series the paper finds no universally best model: the optimal model and lookback depend on the shape of the events, and it characterizes which patterns each model tracks well. It also shows that one-parameter growth models fit the observed outage distributions poorly, that two-parameter distributions (beta and wrapped Cauchy) fit the monthly CDFs, and that a tuned model can estimate a year-long outage statistic for a selected root cause with less than 6% value error, supporting the practical use of such forecasts for assessing mitigation impact.

Load-bearing premise

The comparisons across root-cause types assume the manually assigned root-cause labels are accurate and consistent over the study period, and that counting an outage under every applicable cause does not distort per-type comparisons.

Editorial extensions

If this is right

  • Forecasting monthly outage totals with the tuned foundational model gives 1.3-12.4% lower error than the best classical stochastic model on the full test period.
  • No single model or lookback dominates all eight root-cause series, so practical deployment should choose the model and lag according to the event pattern of each series.
  • Small per-type fine-tuning sets (about 60 samples) do not improve accuracy, suggesting pre-training plus simple transforms, not domain fine-tuning, is what drives the gains.
  • Year-end, iterated multi-step estimates up to twelve months ahead can land within 6% value error for a selected root-cause type, making it possible to track the effect of mitigation work during the year.
  • The sub-6% year-end result is demonstrated for the experiment root-cause type; other types need their own optimized model and lag selection before the bound can be assumed to hold.

Reading between the lines

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

  • If the result holds across services, reliability forecasting can move away from bespoke single-parameter growth models toward zero-shot foundational models, with the immediate engineering question becoming automatic selection of model type and lookback given the target series' pattern.
  • Because eighteen percent of labeled outages have more than one root cause, the per-type rankings are sensitive to how multi-cause incidents are counted; re-running the comparison on the single-cause subset or on independently re-labeled postmortems would test whether the foundational model's per-type advantage is robust.
  • A mixture-of-experts forecaster that picks among moving-average, autoregressive, and foundational-model predictions based on recent spike shape could plausibly beat every single model; the paper's pattern analysis (smooth long spikes vs. double spikes vs. periodic peaks) gives a concrete feature set for such a selector.
  • The sub-6% year-end estimate is demonstrated on one root-cause type; testing iterated multi-step forecasting on the other seven types would reveal whether the accuracy bound comes from the model or from that particular series being relatively smooth.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. This paper evaluates TimesFM, a pre-trained transformer foundation model for time-series forecasting, against simple stochastic baselines (PV, MA, AR) on seven years of monthly production outage counts from a large-scale machine-learning service. The authors first use distribution fitting to argue that one-parameter SRGMs are insufficient for this agile-software data, then compare the forecasting models for total monthly outages and for per-root-cause outage counts. They report that TimesFM achieves 1.3-12.4% lower errors than the best classical model for the full test period, that the winning model and lag vary by root-cause type and test window, and that an optimized model estimates the year-end experiment-caused outage total with less than 6% error. The paper also studies fine-tuning, log transformation, floor, and covariance inputs as optimization techniques.

Significance. If the central claims hold, this would be a useful empirical demonstration that a pre-trained foundation model can be competitive with, and often beat, classical stochastic models on sparse and spiky reliability data. The main strengths are the real seven-year production dataset, the eight root-cause categories, the full lag sweeps over several model families, and the explicit discussion of when each model class tracks spikes or fails to do so. The contribution is entirely empirical and the dataset is proprietary, so independent reproducibility is limited; no code or data is provided. The headline ranking is plausible but not yet established, because the best lag values appear to be selected on the test interval, the per-root-cause data periods are internally inconsistent, and no uncertainty quantification is given. These issues are fixable in revision.

major comments (4)
  1. [Section VI.A / Table IV] The headline result that FM(7) beats MA(6) by 1.3% in MAE and 12.4% in MSE is not an out-of-sample comparison. Section V.B defines train/validation/test splits, but Section VI.A reports the best lag for each model on the same six-year test interval used to compute the errors, and no validation-based lag selection is shown. Because only a few lags are compared per family, the reported margin can reflect selection noise. Table V makes the window dependence explicit: on the last two years, MA(9) has MAE 1.326 versus 1.534 for FM(7). Please report lag selection on the validation split only, or use nested cross-validation, and provide confidence intervals or a paired significance test for the headline differences.
  2. [Section III.A footnote 1 vs. Section IV / Tables VI-VIII] The manuscript is internally inconsistent about label coverage. Footnote 1 in Section III.A states that root-cause labels exist only for the last three years, yet Section IV says the per-root-cause curve fitting uses the last five years and Table VI is explicitly headed 'LAST 5 YEARS' for all eight root-cause types. If labels are not available for the first two of those five years, the five-year per-root-cause series in Table VI (and the per-type rankings derived from it) do not have a stable denominator. The Table IX year-end experiment-type estimate also depends on this per-type series. Please clarify exactly how many years of labeled root-cause data exist, state which months are labeled, and report the label coverage per root-cause type.
  3. [Section III (Accuracy Metrics) / Table IX] The Error% formula is inconsistent with Table IX. Section III defines Error% = sum_i |o_i - o_hat_i|/(o_i * I), which divides each month's error by that month's actual value before summing. Table IX instead computes Error% as (sum of predicted values - sum of actual values)/sum of actual values: for example, AR(3) at A=-12 predicts a sum of 16.87 versus an actual sum of 17.88, giving -6%. The text also claims that the best model 'always has <6% estimation errors', but Table IX(d) shows FM(11) at A=-12 with -9%. The signed values in Table IX and the sign convention are unexplained. Please reconcile the formula, the table, and the abstract claim.
  4. [Section VI (all tables)] No confidence intervals or significance tests are reported for any MAE or MSE comparison. With 72 test months (and only 24 months in Tables V, VII, and VIII) of sparse, high-variance outage counts, differences such as the 1.3% MAE gain for total outages, or the per-root-cause winners in Tables VI-VIII, may be sampling noise. Please add bootstrap confidence intervals or a paired test (e.g., Diebold-Mariano on the monthly errors) for at least the headline total-outage comparison and the per-root-cause winner claims.
minor comments (6)
  1. [Section III vs. Section VII] Reference [17] is used in Section III as a citation for extreme events, but reference [17] is the Outage-Watch paper; the extreme-event citations appear to be [18]-[20]. Please renumber the references.
  2. [Figure 4 caption] The caption of Figure 4 says 'x-axis: week index', but the experiments in Section VI are described as monthly forecasting. Please clarify the correct time unit on the axis.
  3. [Table VI title] Table VI is titled 'ABSOLUTE OUTAGE COUNTS', while Section III defines normalized outage counts and normalized error metrics. Please clarify whether the per-root-cause values are normalized by the total outage count, by per-type totals, or left in raw counts.
  4. [Section IV / Table I] For the daily granularity, the best K-S p-values are 0.0352 (beta) and 0.0074 (wrapped Cauchy), which do not indicate a good fit. Observation #1 should explicitly state that the daily fit is poor and that the two-parameter conclusion rests mainly on the weekly and monthly granularities.
  5. [Section V.B / Section VI.B] The fine-tuning experiments report gains of less than 0.02% in average errors but no confidence intervals; the conclusion that per-type fine-tuning 'does not result in any accuracy gains' should be softened given the small validation set and the absence of significance testing. Hyperparameters for fine-tuning (epochs, learning rate, batch size) are also not reported.
  6. [General] The manuscript still carries the 'REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER' header and has numerous grammatical slips (e.g., 'foundational model was not yet used', 'a type to extreme events'). Please copy-edit the text before submission.

Circularity Check

1 steps flagged · score 5.0 of 10

Year-end '<6%' claim uses test-window-selected lags; core FM-vs-stochastic benchmark remains non-circular.

  1. fitted input called prediction [Section VI.C (Estimating Year-End Outage Counts, Table IX); model selection in Section VI.B (Table VII); Introduction/Abstract]
    "Finally, we use the models with optimized parameters to estimate the monthly outages of the seventh year of a specific root cause type (i.e., long-term, iterated multi-step forecasting). The best model always has <6% estimation errors although it is done up to 12 months ahead. ... When using the last two years of data as the test dataset, the result is different. As summarized in Table VII ... AR shows the highest accuracy for the experiment root cause type. ..."

    The 'optimized parameters' for the experiment-type year-end estimate are AR(3), MA(8), and FM(11), which are exactly the best, second-best, and third-best configurations selected on the last-two-year test window in Table VII. The target of Table IX is the seventh year, which lies inside that same two-year test window. Section V.B defines train/validation/test splits, but no validation-based lag selection is shown; the paper instead reports lags that minimize error on the test interval itself and then presents errors on a subset of that interval as a forecast. The <6% figure is therefore an in-sample, test-selected result rather than an independent out-of-sample prediction.

full rationale

The paper is largely a self-contained empirical evaluation. The forecasting comparison is not derived from any equation that already contains the outcome; TimesFM, MA, AR, and PV are standard external baselines. The only author self-citation ([14]) is contextual (outage post-mortem practice and operational causes) and does not carry any of the accuracy claims. The curve-fitting section fits standard distributions and reports K-S p-values; this is legitimate fitting, and the observation that two-parameter distributions fit better is not circular. The circular element I can exhibit concretely is in Section VI.C: the 'optimized parameters' for the experiment-type year-end estimate are AR(3), MA(8), and FM(11), exactly the best/second/third configurations selected on the last-two-year test window in Table VII, while year 7 is inside that window. Thus the <6% error is a test-selected, in-sample number rather than an out-of-sample prediction. This does not invalidate the independent FM-vs-MA comparison, which stands or falls on test-window and uncertainty analysis (a soundness concern, not circularity). The score reflects one partial circular step, not a definitional collapse.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the proprietary seven-year outage dataset, on manually assigned root-cause labels, and on the choice of lag parameters selected using the test period. No new physical or conceptual entities are introduced.

free parameters (7)
  • FM lag (lookback) for total outages = 7 months
    Selected as the lag with lowest test error in Table IV; this is a post-hoc choice on the test set.
  • MA lag for total outages = 6 months
    Selected as the second-best lag in Table IV based on test-period errors.
  • AR lag for experiment-caused year-end estimation = 3 months
    Chosen because AR(3) is used in Table IX; lags are evaluated on test data in Tables VI-VII.
  • MA lag for experiment-caused year-end estimation = 8 months
    Used in Table IX for MA(8); selected from test-period comparisons.
  • FM lag for experiment-caused year-end estimation = 11 months
    Used in Table IX for FM(11); FM(11) is reported as the fine-tuned freq=1 model in Table VIII.
  • Fine-tuning frequency flag = 0 or 1
    The frequency parameter is tuned and shown to affect accuracy for some root cause types (Section VI-B).
  • Distribution parameters (beta, wrapped Cauchy, etc.) = e.g., beta a=0.83818 b=0.749637; wrapcauchy rho=0.1
    Curve fitting in Section IV fits 56 distributions to the same outage data; these parameters are descriptive, not used for forecasting, but they motivate the two-parameter claim.
assumptions (5)
  • domain assumption The monthly aggregation of outage counts is stationary and forecastable (ADF test p<0.05 for monthly and weekly).
    Section III-A uses Augmented Dickey-Fuller tests to justify using monthly counts; the test is standard but the choice of aggregation level is a modeling decision.
  • domain assumption Manual post-mortem root-cause labels for outages are correct and complete enough for per-type analysis.
    Section III states every outage has a post-mortem with root cause types identified; the paper does not measure label accuracy or inter-rater agreement.
  • domain assumption TimesFM's pre-trained weights are a valid starting point for fine-tuning on outage counts.
    The paper relies on the external TimesFM model [8] and assumes its inductive biases transfer to sparse count data; no independent check is offered.
  • domain assumption The first-year data split for training/validation and six-year test split is appropriate for evaluating the models.
    Section V-B describes the split; no sensitivity analysis is given.
  • standard math Standard statistical tests (K-S, ADF) have their usual validity for this sample size.
    Used in Section IV and III-A; with small samples these tests may have limited power, but the methods are standard.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Evaluation of a Foundational Model and Stochastic Models for Forecasting Sporadic or Spiky Production Outages of High-Performance Machine Learning Services." pith.science (2026). https://pith.science/paper/PYE5XJKG

@misc{pith2026250701067,
  author       = {Pith},
  title        = {Pith review of: Evaluation of a Foundational Model and Stochastic Models for Forecasting Sporadic or Spiky Production Outages of High-Performance Machine Learning Services},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PYE5XJKG}},
  note         = {Machine review of arXiv:2507.01067}
}
read the original abstract

Time series forecasting models have diverse real world applications (e.g., from electricity metrics to software workload). Latest foundational models trained for time series forecasting show strengths (e.g., for long sequences and in zero-shot settings). However, foundational model was not yet used for forecasting rare, spiky events, i.e., a challenging target because those are a corner case of extreme events. In this paper, we optimize a state-of-the-art foundational model to forecast sporadic or spiky production outages of high-performance machine learning services powering billions of client devices. We evaluate the forecasting errors of the foundational model compared with classical stochastic forecasting models (e.g., moving average and autoregressive). The analysis helps us understand how each of the evaluated models performs for the sporadic or spiky events. For example, it identifies the key patterns in the target data that are well tracked by the foundational model vs. each of the stochastic models. We use the models with optimal parameters to estimate a year-long outage statistics of a particular root cause with less than 6% value errors.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

79 extracted references · 67 canonical work pages

  1. [1]

    Outage detection via real -time social stream analysis: leveraging the power of online complaints ,

    E. Augustine, C. Cushing, A. Dekhtyar, K. McEntee, K. Paterson, and M. Tognetti, “Outage detection via real -time social stream analysis: leveraging the power of online complaints ,” in Proceedings of the 21st ACM International Conference on World Wide Web (WWW '12 Companion), pp. 13–22, 2012

    work page 2012

  2. [2]

    #facebookdown: Time to panic or detox? Understanding users’ reactions to social media outage ,

    M. Liao and S. S . Sundar, “#facebookdown: Time to panic or detox? Understanding users’ reactions to social media outage ,” In Extended Abstracts of the 2022 ACM CHI Conference on Human Factors in Computing Systems (CHI EA '22). Article 357, pp. 1–8, 2022

    work page 2022

  3. [3]

    Software Reliability Modeling Survey,

    W. Farr, “Software Reliability Modeling Survey,” Ch. 3, Ed. M. R. Lyu, Handbook of Software reliability Engineering , IEEE Computer Society Press, McGraw-Hill, 1996

    work page 1996

  4. [4]

    Software Reliability Growth Model,

    A. Wood, “Software Reliability Growth Model,” Tandem Technical Report, 1996

    work page 1996

  5. [5]

    D. D. Hanagal and N. N. Bhalerao, Software Reliability Growth Model, Springer, 2021

    work page 2021

  6. [6]

    Theory and Practice

    Birolini, A.: Reliability Engineering. Theory and Practice. 3rd ed. Springer, Berlin, Heidelberg 1999

    work page 1999

  7. [7]

    https://agilemanifesto.org/

    Manifesto for Agile Software Development. https://agilemanifesto.org/

    work page

  8. [8]

    A decoder -only foundation model for time-series forecasting,

    A. Das, W. Kong, R. Sen, and Y . Zhou, “A decoder -only foundation model for time-series forecasting,” arXiv:2310.10688, 2024

    arXiv 2024

Show all 79 references
  1. [9]

    Time -Dependent Error -Detection Rate Model for Software Reliability and Other Performance Measur es,

    A. L. Goel and K. Okumoto, “Time -Dependent Error -Detection Rate Model for Software Reliability and Other Performance Measur es,” IEEE Transactions on Reliability, vol. R-28, no. 3, pp. 206-211, 1979

    1979

  2. [10]

    Software Reliability Growth Modeling: Models and Applications,

    S. Yamada and S. Osaki, “Software Reliability Growth Modeling: Models and Applications,” IEEE Transactions on Software Engineering, vol. SE- 11, no. 12, pp. 1431-1437, 1985

    1985

  3. [11]

    In: Software Reliability Handbook

    Littlewood, B.: Modelling Growth in Software Reliability. In: Software Reliability Handbook. Elsevier, London, New York 1990 (Edt: Paul Rook), pp. 137-153, 401-412

    1990

  4. [12]

    The use of reliability growth models in project management,

    E. Miranda, “The use of reliability growth models in project management,” in Proceedings of the IEEE ISSRE, pp. 291-298, 1998

    1998

  5. [13]

    What bugs cause production cloud incidents? ,

    H. Liu, S. Lu, M. Musuvathi, and S. Nath, “What bugs cause production cloud incidents? ,” in Proceedings of the Workshop on Hot Topics in Operating Systems (HotOS), pp. 155–162, 2019. 13 > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE -CLICK HERE TO EDIT) <

    2019

  6. [14]

    Norming to Performing: Failure Analysis and Deployment Automation of Big Data Software Developed by Highly Iterative Models,

    K. S. Yim, “Norming to Performing: Failure Analysis and Deployment Automation of Big Data Software Developed by Highly Iterative Models,” in Proceedings of the IEEE 25th International Symposium on Software Reliability Engineering, pp. 144–155, 2014

    2014

  7. [15]

    How to fight production incidents? An empirical study on a large -scale cloud service ,

    S. Ghosh, M. Shetty, C. Bansal, and S . Nath, “How to fight production incidents? An empirical study on a large -scale cloud service ,” in Proceedings of the 13th Sympos ium on Cloud Computing (SoCC ), pp. 126–141, 2022

    2022

  8. [16]

    Debugging Incidents in Google’s Distributed Systems: How experts debug production issues in complex distributed systems,

    C. Chan and B . Cooper, “Debugging Incidents in Google’s Distributed Systems: How experts debug production issues in complex distributed systems,” Queue, 18(2):40, 20 pages, 2020

    2020

  9. [17]

    Outage-Watch: Early Prediction of Out ages using Extreme Event Regularizer,

    S. Agarwal, S. Chakraborty, S. Garg, S . Bisht, C. Jain, A. Gonuguntla, and S. Saini, “Outage-Watch: Early Prediction of Out ages using Extreme Event Regularizer,” in Proceedings of the ACM Joint European Software Engineering Conference and Symposium on the Foundations of Soft...

    2023

  10. [18]

    Albeverio, V

    S. Albeverio, V. Jentsch, and H . Kantz, Extreme Events i n Nature and Society, Springer Science & Business Media, 2006

    2006

  11. [19]

    Extreme events: dynamics, statistics and prediction,

    M. Ghil, P. Yiou, S. Hallegatte, B. D. Malamud, P. Naveau, A. Soloviev, P. Friederichs, V. Keilis-Borok, D . Kondrashov, V. Kossobokov, et al., “Extreme events: dynamics, statistics and prediction,” Nonlinear Processes in Geophysics, 18(3):295–350, 2011

    2011

  12. [20]

    Haan and A

    L. Haan and A . Ferreira, Extreme value theory: an introduction , Vol. 3, Springer, 2006

    2006

  13. [21]

    Time -series extreme event forecasting with neural networks at Uber,

    N. Laptev, J. Yosinski, L. E. Li, and S. Smyl, “Time -series extreme event forecasting with neural networks at Uber,” in International Conference on Machine Learning (ICML) – Time Series Workshop , Vol. 34 , pp. 1–5, 2017

    2017

  14. [22]

    Forecast verification for extreme value distributions with an application to probabilistic peak wind prediction,

    P. Friederichs and T. L. Thorarinsdottir, “Forecast verification for extreme value distributions with an application to probabilistic peak wind prediction,” Environmetrics, 23(7):579–594, 2012

    2012

  15. [23]

    Modeling extre me events in time series prediction ,

    D. Ding, M . Zhang, X . Pan, M . Yang, and X . He, “Modeling extre me events in time series prediction ,” in Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp. 1114–1122, 2019

    2019

  16. [24]

    P. J. Brockwell and R. A. Davis, Time Series: Theory and Methods , Springer, 1991

    1991

  17. [25]

    M. H. Hayes, Statistical Digital Signal Processing and Modeling , Wiley, 1996

    1996

  18. [26]

    G. E. P. Box, G. M. Jenkins, and G. C. Reinsel, Time Series Analysis: Forecasting and Control, 3rd ed., Prentice-Hall, 1994

    1994

  19. [27]

    Are Transformers Effective for Time Series Forecasting? ,

    A. Zeng, M. Chen, L. Zhang, and Q. Xu, “Are Transformers Effective for Time Series Forecasting? ,” in Proceedings of the AAAI conference on artificial intelligence, 37:11121–11128, 2023

    2023

  20. [28]

    One Fits All:Power General Time Series Analysis by Pretrained LM,

    T. Zhou, P.S. Niu, X. Wang, L. Sun, R. Jin, “One Fits All:Power General Time Series Analysis by Pretrained LM,” arXiv:2302.11939, 2023

    2023 arXiv

  21. [29]

    Bidirectional recurrent neural networks,

    M. Schuster and K. K. Paliwal, “Bidirectional recurrent neural networks,” IEEE Transactions on Signal Processing, 45(11):2673–26, 1997

    1997

  22. [30]

    Outage prediction and diagnosis for cloud service systems,

    Y. Chen, X. Yang, Q. Lin, H. Zhang, F. Gao, Z. Xu, Y. Dang, D. Zhang, H. Dong, Y. Xu, et al., “Outage prediction and diagnosis for cloud service systems,” in Proceedings of the World Wide Web Conference , pp. 2659– 2665, 2019

    2019

  23. [31]

    Real-time incident prediction for online service systems,

    N. Zhao, J. Chen, Z. Wang, X. Peng, G. Wang, Y. Wu, F. Zhou, Z. Feng, X. Nie, W. Zhang, K. Sui, and D. Pei, “Real-time incident prediction for online service systems,” in Proceedings of the 28th ACM Joint Meeting on European Software Engineering Conference and Symposium on the...

    2020

  24. [32]

    Fighting the Fog of War: Automated Incident Detection for Cloud Systems ,

    L. Li, X. Zhang, X. Zhao, H. Zhang, Y. Kang, P. Zhao, B. Qiao, S. He, P. Lee, J . Sun, et al. , “ Fighting the Fog of War: Automated Incident Detection for Cloud Systems ,” in Proceedings of the USENIX Annual Technical Conference (ATC), pp. 131–146, 2021

    2021

  25. [33]

    Adaptive performance anomaly detection for online service systems via pattern sketching,

    Z. Chen, J . Liu, Y . Su, H . Zhang, X . Ling, Y . Yang, and M . R. Lyu, “Adaptive performance anomaly detection for online service systems via pattern sketching,” in Proceedings of the 44th Inter -national Conference on Software Engineering, pp. 61–72, 2022

    2022

  26. [34]

    Fault detection and diagnosis using combined autoencoder and long short -term memory network ,

    P. Park, P. Di Marco, H. Shin, and J. Bang, “Fault detection and diagnosis using combined autoencoder and long short -term memory network ,” Sensors, 19(21)4612, 2019

    2019

  27. [35]

    Deep convolutional neural network model based chemical process fault diagnosis,

    H. Wu and J. Zhao, “Deep convolutional neural network model based chemical process fault diagnosis,” Comput. Chem. Eng. , 115:185 –197, 2018

    2018

  28. [36]

    Cross-dataset time series anomaly detection for cloud systems ,

    X. Zhang, J. Kim, Q. Lin, K. Lim, S. O. Kanaujia, Y. Xu, K. Jamieson, A. Albarghouthi, S. Qin, M. J. Freedman, et al., “Cross-dataset time series anomaly detection for cloud systems ,” in Proceedings of the USENIX Annual Technical Conference (ATC), pp. 1063–1076, 2019

    2019

  29. [37]

    Large scale predictive analytics for hard disk remaining useful life estimation ,

    P. Anantharaman, M . Qiao, and D . Jadav, “Large scale predictive analytics for hard disk remaining useful life estimation ,” in Proceedings of the 2018 IEEE International Congress on Big Data (BigData Congress), pp. 251–254, 2018

    2018

  30. [38]

    Making disk failure predictions smarter! ,

    S. Lu, B. Luo, T. Patel, Y. Yao, D . Tiwari, and W . Shi, “Making disk failure predictions smarter! ,” in Proceedings of the USENIX FAST , pp. 151–167, 2020

    2020

  31. [39]

    Predicting Failures in Hard Drives with LSTM Networks,

    F. D. d. S. Lima, G. M. R. Amaral, L. G. d. M. Leite, J. P. P. Gomes , and J. d. C. Machado, “Predicting Failures in Hard Drives with LSTM Networks,” in Proceedings of the Brazilian Conference on Intelligent Systems (BRACIS), pp. 222-227, 2017

    2017

  32. [40]

    A multimodal anomaly detector for robot-assisted feeding using an LSTM-based variational autoencoder ,

    D. Park, Y. Hoshi, and C. C. Kemp, “A multimodal anomaly detector for robot-assisted feeding using an LSTM-based variational autoencoder ,” IEEE Robotics and Automation Letters, 3(3):1544–1551, 2018

    2018

  33. [41]

    Robust anomaly detection for multivariate time series through stochastic recurrent neural network,

    Y. Su, Y. Zhao, C. Niu, R. Liu, W. Sun, and D . Pei, “Robust anomaly detection for multivariate time series through stochastic recurrent neural network,” in Proceedings of the 25th ACM SIGKDD intern ational conference on knowledge discovery & data mining, pp. 2828–2837, 2019

    2019

  34. [42]

    TranAD: deep transformer networks for anomaly detection in multivariate time series data ,

    S. Tuli, G . Casale, and N . R. Jennings, “TranAD: deep transformer networks for anomaly detection in multivariate time series data ,” in Proceedings of the VLDB Endowment, 15(6):1201–1214, 2022

    2022

  35. [43]

    ARIMA models to predict next -day electricity prices,

    J. Contreras, R. Espinola, F. J. Nogales and A. J. Conejo, “ARIMA models to predict next -day electricity prices,” IEEE Transactions on Power Systems, 18(3):1014-1020, 2003

    2003

  36. [44]

    ARIMA -Based Time Series Model of Stochastic Wind Power Generation,

    P. Chen, T. Pedersen, B. Bak -Jensen and Z. Chen, "ARIMA -Based Time Series Model of Stochastic Wind Power Generation," IEEE Transactions on Power Systems, 25(2):667-676, 2010

    2010

  37. [45]

    Stock Price Prediction Using the ARIMA Model,

    A. A. Ariyo, A. O. Adewumi and C. K. Ayo, “Stock Price Prediction Using the ARIMA Model,” in Proceedings of the UKSim -AMSS 16th International Confe rence on Computer Modelling and Simulation , pp. 106-112, 2014

    2014

  38. [46]

    Workload Prediction Using ARIMA Model and Its Impact on Cloud Applications’ QoS,

    R. N. Calheiros, E. Masoumi, R. Ranjan and R. Buyya, “Workload Prediction Using ARIMA Model and Its Impact on Cloud Applications’ QoS,” IEEE Transactions on Cloud Computing, 3(4):449-458, 2015

    2015

  39. [47]

    Forecasting seasonals and trends by exponentially weighted moving averages,

    C. Holt, “Forecasting seasonals and trends by exponentially weighted moving averages,” International Journal of Forecasting, 20:5–10, 2004

    2004

  40. [48]

    Greedy function approximation: a gradient boosting machine,

    J. H. Friedman, “Greedy function approximation: a gradient boosting machine,” Annals of statistics, pp. 1189–1232, 2001

    2001

  41. [49]

    Recurrent Networks and NARMA Modeling,

    J. Connor, L. E. Atlas, and D. R. Martin , “Recurrent Networks and NARMA Modeling,” NIPS, pp. 301–308, 1991

    1991

  42. [50]

    On the properties of neural machine translation: Encoder -decoder approache s,

    K. Cho, B. V. Merrienboer, D. Bahdanau, and Y. Bengio, “On the properties of neural machine translation: Encoder -decoder approache s,” ArXiv:1409.1259, 2014

    2014 arXiv

  43. [51]

    Long short -term memory,

    S. Hochreiter and J. Schmidhuber, “Long short -term memory,” Neural Computation, 9:1735–1780, 1997

    1997

  44. [52]

    Modeling Long- and Short- term Temporal Patterns with Deep Neural Networks,

    G. Lai, W.-C. Chang, Y. Yang, and H. Liu, “Modeling Long- and Short- term Temporal Patterns with Deep Neural Networks,” in Proceedings of the International ACM SIGIR Conference on Research and Development in Information Retrieval, 2017

    2017

  45. [53]

    An empirical evaluation of generic convolutional and recurrent networks for sequence modeling ,

    S. Bai, J. Z. Kolter, and V . Koltun, “An empirical evaluation of generic convolutional and recurrent networks for sequence modeling ,” arXiv:1803.01271, 2018

    2018 arXiv

  46. [54]

    Time series is a special sequence: Forecasting with sample convolution and interaction ,

    M. Liu, A . Zeng, Z . Xu, Q . Lai, and Q . Xu, “Time series is a special sequence: Forecasting with sample convolution and interaction ,” arXiv:2106.09305, 2021

    2021 arXiv

  47. [55]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in Neural Information Processing Systems, 30, 2017

    2017

  48. [56]

    LLM4TS: Aligning Pre-Trained LLMs as Data -Efficient Time-Series Forecasters,

    C. Chang, W. -Y. Wang, W. -C. Peng, and T. -F. Chen, “LLM4TS: Aligning Pre-Trained LLMs as Data -Efficient Time-Series Forecasters,” arXiv:2308.08469, 2024

    2024 arXiv

  49. [57]

    Large Language Models Are Zero-Shot Time Series Forecasters,

    N. Gruver, M. Finzi, S. Qiu, and A. G. Wilson, “Large Language Models Are Zero-Shot Time Series Forecasters,” arXiv:2310.07820, 2024

    2024 arXiv

  50. [58]

    A time series is worth 64 words: Long -term forecasting with transformers,

    Y. Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam, “A time series is worth 64 words: Long -term forecasting with transformers,” in Proceedings of the ICLR, OpenReview.net, 2023

    2023

  51. [59]

    Enhancing the locality and breaking the memory bottleneck of transformer on time series forecasting,

    S. Li, X. Jin, Y. Xuan, X. Zhou, W. Chen, Y. -X. Wang, and X. Yan, “Enhancing the locality and breaking the memory bottleneck of transformer on time series forecasting,” Advances in Neural Information Processing Systems, 32, 2019

    2019

  52. [60]

    Informer: Beyond efficient transformer for long sequence time -series 14 > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE -CLICK HERE TO EDIT) < forecasting,

    H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W. Zhang, “Informer: Beyond efficient transformer for long sequence time -series 14 > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE -CLICK HERE TO EDIT) < forecasting,” in Proceedings of the Thirty -Fifth AA...

    2021

  53. [61]

    Pyraformer: Low -complexity pyramidal attention for long -range time series modeling and forecasting,

    S. Liu, H. Yu, C. Liao, J. Li, W. Lin, A. X. Liu, and S. Dustdar, “Pyraformer: Low -complexity pyramidal attention for long -range time series modeling and forecasting,” in Proceedings of the Inter national Conference on Learning Representations (ICLR), 2021

    2021

  54. [62]

    Triformer: Triangular, variable -specific attentions for long sequence multivariate time series forecasting –full version,

    R.-G. Cirstea, C. Guo, B. Yang, T. Kieu, X. Dong, and S. Pan, “Triformer: Triangular, variable -specific attentions for long sequence multivariate time series forecasting –full version,” arXiv:2204.13767, 2022

    2022 arXiv

  55. [63]

    Fedformer: Frequency enhanced decomposed transformer for long -term series forecasting,

    T. Zhou, Z. Ma, Q. Wen, X. Wang, L. Sun, and R. Jin, “Fedformer: Frequency enhanced decomposed transformer for long -term series forecasting,” in Proceedings of the International Conference on Machine Learning (ICML), 2022

    2022

  56. [64]

    Chronos: Learning the Language of Time Series,

    A. F. Ansari, L. S tella, C. Turkmen, X. Zhang, P. Mercado, H. Shen, O. Shchur, S. S. Rangapuram, S. P. Arango, S. Kapoor, J. Zschiegner, D. C. Maddix, H. Wang, M. W. Mahoney, K. Torkkola, A. G. Wilson, M. Bohlke-Schneider, and Y. Wang, “Chronos: Learning the Language of Time ...

    2024 arXiv

  57. [65]

    Autoformer: Decomposition transformers with auto -correlation for long -term series forecasting ,

    J. Xu, J . Wang, M . Long, et al. , “ Autoformer: Decomposition transformers with auto -correlation for long -term series forecasting ,” Advances in Neural Information Processing Systems, 34, 2021

    2021

  58. [66]

    Transformers in Time Series: A Survey,

    Q. Wen, T. Zhou, C. Zhang, W. Chen, Z. Ma, J. Yan, and L. Sun, “Transformers in Time Series: A Survey,” arXiv:2202.07125, 2023

    2023 arXiv

  59. [67]

    TSMixer: An All-MLP Architecture for Time Series Forecasting,

    S.-A. Chen, C.-L. Li, N. Yoder, S. O. Arik, and T. Pfister, “TSMixer: An All-MLP Architecture for Time Series Forecasting,” arXiv:2303.06053, 2023

    2023 arXiv

  60. [68]

    SAMformer: Unlocking the Potential of Transformers in Time Series Forecasting with Sharpness -Aware Minimization and Channel-Wise Attention,

    R. Ilbert, A. Odonnat, V. Feofanov, A. Virmaux, G. Paolo, T. Palpanas, and I. Redko, “SAMformer: Unlocking the Potential of Transformers in Time Series Forecasting with Sharpness -Aware Minimization and Channel-Wise Attention,” arXiv:2402.10198, 2024

    2024 arXiv

  61. [69]

    ETDataset, https://github.com/zhouhaoyi/ETDataset

  62. [70]

    https://archive.ics.uci.edu/ml/datasets/ElectricityLoadDiagrams20112014

  63. [71]

    https://ncei.noaa.gov/data/local-climatological-data

  64. [72]

    Fast Outage Analysis of Large -scale Production Clouds with Service Correlation Mining,

    Y. Wang, G. Li, Z. Wang, Y. Kang, Y. Zhou, H. Zhang, F. Gao, J. Sun, L. Yang, P. Lee, Z. Xu, P. Zhao, B. Qiao, L. Li, X. Zhang, and Q. Lin, “Fast Outage Analysis of Large -scale Production Clouds with Service Correlation Mining,” in Proceedings of the 43rd International Confer...

    2021

  65. [73]

    Root cause detection in a service - oriented architecture ,

    M. Kim, R . Sumbaly, and S . Shah, “Root cause detection in a service - oriented architecture ,” in Proceedings of the ACM SIGMETRICS/international conference on Measurement and modeling of computer systems (SIGMETRICS), pp. 93–104, 2013

    2013

  66. [74]

    ESRO: Experience Assisted Service Reliability against Outages ,

    S. Chakraborty, S . Agarwal, S . Garg, A . Sethia, U . N. Pandey, V . Aggarwal, and S . Saini, “ESRO: Experience Assisted Service Reliability against Outages ,” in Proceedings of the 38th IEEE/ACM International Conference on Automated Software Engineering (ASE) , pp. 255–267, 2024

    2024

  67. [75]

    Assess and Summarize: Improve Outage Understan ding with Large Language Models,

    P. Jin, S. Zhang, M. Ma, H. Li, Y. Kang, L. Li, Y. Liu, B. Qiao, C. Zhang, P. Zhao, S. He, F. Sarro, Y. Dang, S. Rajmohan, Q. Lin, and D . Zhang, “Assess and Summarize: Improve Outage Understan ding with Large Language Models,” in Proceedings of the ACM Joint European Software...

    2023

  68. [76]

    Recurrent neural network for rain estimation using commercial microwave links,

    H. V. Habi and H. Messer, "Recurrent neural network for rain estimation using commercial microwave links," IEEE Trans. Geosci. Remote Sens. , vol. 59, no. 5, pp. 3672-3681, May 2021

    2021

  69. [77]

    L Shooman, Probabilistic Reliability: an Engineering Approach , Robert E

    M. L Shooman, Probabilistic Reliability: an Engineering Approach , Robert E. Krieger Publishing Company, Malabar, Florida, 1990

    1990

  70. [78]

    "Is my internet down?

    E. Cem Kirci, M . Vahlensieck, and L . Vanbever, “ "Is my internet down?": sifting through user -affecting outages with Google trends ,” in Proceedings of the 22nd ACM Internet Measurement Conference (IMC) , pp. 290–297, 2022

    2022

  71. [79]

    Crosslayer network outage classification using machine learning ,

    J. M. Evang, A. H. Ahmed, A . Elmokashfi, and H . Bryhni, “Crosslayer network outage classification using machine learning ,” in Proceedings of the 2022 Applied Networking Research Workshop (ANRW '22) , Article 2, pp. 1–7, 2022

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.