Pith. sign in

REVIEW 5 major objections 5 minor 2 cited by

Time Series Foundation Models for Multivariate Financial Time Series Forecasting

T0 review · 5 major / 5 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 Pretrained time-series foundation models cut financial forecast error by 25–50% on limited data, this study reports.

desk verdict Honest pilot study with a reusable paired pretrained/untrained protocol, but the headline transfer gains are not identified as transfer until the TTM pretraining leakage question is settled. read the letter →

arxiv 2507.07296 v1 pith:74Y3IAYN submitted 2025-07-09 q-fin.GN cs.LG

Ben A. Marconi This is my paper
classification q-fin.GNcs.LG
keywords timeseriesfoundationmodelstransferlearningfine-tuningzero-shotforecastingfinancialsampleefficiencyTinyMixersmultivariate
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

This paper asks whether pretrained time-series foundation models (TSFMs) can be transplanted into finance, where daily data are scarce. It reports that Tiny Time Mixers, a compact pretrained model, when fine-tuned on three financial tasks, achieves 25–50% lower mean squared error than the same architecture trained from scratch under data limits, and 15–30% lower error with longer training windows. It also reports that the pretrained model reaches comparable accuracy with 3–10 fewer years of data, and that in volatility and equity-spread tasks its zero-shot forecasts beat naive baselines. A second foundation model, Chronos, does not exhibit transfer on these tasks. The study's conclusion is that TSFMs offer genuine promise in noisy, data-constrained financial settings, but achieving competitive performance likely requires domain-specific pretraining and architectural refinements.

What carries the argument

The evaluation rests on two paired experimental protocols. The transfer-gain test compares a pretrained model with a randomly initialised copy of the same architecture in both zero-shot and fine-tuned regimes, measuring relative gains $\Delta_{\mathrm{ZS}}$ and $\Delta_{\mathrm{FT}}$; a positive gain attributes the improvement to pretraining rather than architecture. The sample-efficiency probe trains both variants on incremental amounts of historical data, so the gap between their learning curves quantifies how many years of data pretraining saves. The models themselves are the other load-bearing components: Tiny Time Mixers is a compact MLP-based time-series model pretrained on broad corpora that include Bitcoin data, and Chronos tokenises time series values into quantised bins and trains a language-model-style transformer without financial pretraining data. The paper argues that the contrast between these two models points to pretraining-corpus composition as a key driver of transfer.

What would settle it

Inspect the time stamps of TTM's pretraining datasets, especially the Bitcoin series, against the 2005–2025 evaluation windows; if any pretraining sample falls inside a test window, retrain from scratch on the truncated corpus and check whether the 25–50% gain persists. A second check is to configure TTM with a true 21-day forecast horizon instead of 30, 48, or 96 days and see whether the stated gains still hold.

Watch

Extended reading notes

Core claim

The paper's central claim is that a pretrained TSFM's weights carry reusable temporal structure that transfers to financial forecasting, while an identical architecture trained from scratch must rediscover that structure from limited data. This is established through paired experiments: fine-tuning both a pretrained Tiny Time Mixers model and a randomly initialized copy of the same architecture, then comparing their errors. Across US 10-year Treasury yield changes, EUR/USD realised volatility, and an equity spread between two country indices, the pretrained version achieved 25–50% lower MSE with limited fine-tuning data, 10–30% lower MSE with full data, and consistently needed fewer years of history to reach comparable performance. In the volatility and spread tasks, the pretrained model in zero-shot mode outperformed naive benchmarks, and in the spread task it outperformed all tested benchmark models without any fine-tuning. The paper also finds that traditional specialised models matched or exceeded TTM in two of three tasks, and that Chronos, an LLM-adapted TSFM pretrained without financial data, failed to beat naive baselines in any task.

Load-bearing premise

The clean-transfer conclusion depends on TTM's pretraining corpus not containing data from the same future period as the evaluation windows; if it does, the measured advantage is leakage rather than transfer.

Editorial extensions

If this is right

  • If the paper is right, practitioners with fewer than ten years of instrument history should prefer fine-tuning a pretrained TSFM over training an equivalent architecture from scratch.
  • Zero-shot TSFM forecasts can beat naive baselines for autocorrelated and mean-reverting financial series, meaning some forecasting value is available before any task-specific training.
  • Fine-tuning transfer gains are largest when data are scarce, suggesting the primary practical value of TSFMs lies in data-constrained settings rather than in pushing the accuracy ceiling on long histories.
  • Pretraining corpus composition matters: the model pretrained with financial data transferred well, while the model pretrained without financial data did not, supporting a finance-specific pretraining research direction.
  • Native time-series architectures may fit financial data better than LLM-adapted designs, although the paper treats this conclusion as tentative because only two models were compared.

Reading between the lines

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

  • If the transfer gains survive leakage checks, the practical consequence is that a general-purpose pretrained model can substitute for years of proprietary financial history, lowering the data barrier for newly listed or illiquid instruments.
  • The TTM-versus-Chronos gap suggests that architecture alone is not the explanation; a direct test would be to pretrain the same architecture on corpora that differ only in whether financial series are included.
  • The paper's MSE-based evaluation leaves open whether the gains translate into tradeable signals; the appendix backtest on Treasury ETFs points in that direction, but a Sharpe-ratio or directional-accuracy evaluation would settle it.
  • A cleaner comparison would align TTM's configured horizons (30, 48, and 96 days) with the stated 21-day task horizons, since a horizon mismatch could distort the measured transfer gains.
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

5 major / 5 minor

Summary. The paper evaluates two time-series foundation models, TTM and Chronos-Bolt, on three multivariate financial forecasting tasks: 21-day-ahead US 10-year Treasury yield changes, 21-day-ahead EUR/USD realized volatility, and equity spread forecasts at multiple horizons. The experimental design compares randomly initialized and pretrained versions of the same architecture under both zero-shot and fine-tuned regimes, using paired transfer-gain statistics (Equations 2 and 3) and a sample-efficiency probe that varies the training window length. The central claim is that pretrained TTM fine-tuned on limited financial data achieves 25–50% lower MSE than an identically architected untrained model, requires 3–10 fewer years of data to reach comparable performance, and in some tasks beats naive benchmarks in zero-shot mode. Classical baselines such as VAR, LSTM, ridge regression, and an error-correction model are also compared. The paper concludes that TSFMs are promising for data-constrained financial forecasting but require domain-specific pretraining to be competitive.

Significance. If the headline transfer gains survive a temporally clean evaluation, the paper would make a useful empirical contribution: it defines explicit paired transfer-gain metrics, applies a genuinely interesting pretrained-versus-untrained control for TTM, uses rolling out-of-sample evaluation, and includes a practical backtest in Appendix A.3. The three tasks span different asset classes and statistical properties, and the feature transformations are documented in detail. However, as it stands, the central quantitative claims are not identified as transfer effects because the pretraining corpus may overlap the evaluation period, and the headline numbers rest on single-seed point estimates with no dispersion or significance testing. The horizon configuration also diverges from the stated 21-day task horizon, so the reported MSE values are not measuring exactly what the abstract claims.

major comments (5)
  1. [2.6.1, Appendix A.1, 6.2] The transfer-gain identification is not established because TTM's pretraining corpus contains financial data whose temporal range is undocumented. Section 2.6.1 and Appendix A.1 list the Bitcoin data set (Zenodo record 5122101) among TTM's pretraining corpora, while the evaluation windows for the sample-efficiency probe and the rolling transfer-gain test run from roughly 2014 through 2024 (Figures 10, 15, and 20). Section 6.2 concedes that 'pretraining from scratch would eliminate potential look-ahead bias from including correlated future data in TTM's financial pretraining dataset.' Since the paper nowhere reports the date range of the Bitcoin record or its relation to the evaluation period, the positive ΔFT and ΔZS values in Tables 10–12 can be produced by temporal leakage rather than by reusable task-agnostic representations. The authors should document non-overlap of the pretraining corpus, use a checkpoint whose pretraining data are verifiably earlier than the evaluation period, or rerun the experiments with a from-scratch pretrained model.
  2. [4.1.1, Table 8; 1.1; 5] The configured forecast horizons are inconsistent with the stated task horizons, so the reported MSE values are not measuring what the abstract claims. Section 1.1 and Sections 5.1–5.2 describe 21-business-day-ahead forecasts, but Table 8 sets TTM horizons of 30, 48, and 96 for Tasks 1–3; Chronos is run at horizons of 21 and 10 (Table 9). Because MSE is averaged over all forecast steps, the TTM numbers include substantially longer horizons than the naive benchmarks and the Chronos numbers, and the headline 25–50% gains are not directly comparable across models or to the stated 21-day task. This is especially acute in Task 3, where the text reports 5- and 10-day horizons but Table 8 assigns TTM a horizon of 96. The paper should run TTM at the 21-step horizon, report horizon-specific MSE, or provide an explicit argument for why a 30/48/96-step average is equivalent to the stated task.
  3. [4.1.1, 4.3, Tables 10–12] All results rest on a single fixed seed, and the transfer gains are reported without any measure of dispersion or formal significance test. Section 4.1.1 states that a fixed random seed was used for all TTM experiments, and Tables 10–12 present ΔFT and ΔZS as point estimates. Section 5.4.4 nonetheless describes the gains as 'statistically significant' and 'robust,' but no significance test, confidence interval, or repeated-seed standard error is provided anywhere in the paper. Since the headline claims are quantitative and some of the learning-curve differences in Figures 9, 14, and 19 are modest, the authors should report means and variances over multiple seeds or bootstrap resamples, state the test used, and show error bars in the learning curves.
  4. [4.3, 5.3.1] The hyperparameter and training-window selection is anchored to a single reference date, and the paper itself documents the resulting overfitting in Task 3. The sample-efficiency probe fixes 22 January 2021 as the reference date and uses it to select the training window k, while Section 4.1.1 says context length, forecast horizon, and epochs were chosen by validation. Section 5.3.1 then notes that performance at that 2021 date shows rough parity between zero-shot and fine-tuned pretrained models and that 'hyperparameter choices and training procedures optimised for January 2021 do not generalise effectively to other time periods.' Because the same k and hyperparameters feed every rolling window in Section 4.4, the rolling transfer gains may partly reflect tuning to a single date. The authors should use a nested or rolling validation protocol, or report sensitivity of the transfer gains to the choice of reference date.
  5. [5.1.2, 5.2.2, 5.3.2] Chronos is excluded from the transfer-gain analysis after seeing its results, which undercuts the paper's comparison of the two TSFM families. Sections 5.1.2, 5.2.2, and 5.3.2 state that Chronos is excluded from subsequent transfer-gain analyses because it fails to beat the naive benchmark, and the abstract and conclusions nonetheless draw general conclusions about 'TSFMs' and about native time-series architectures versus LLM-adapted approaches from the TTM–Chronos contrast. The authors should report the Chronos ΔFT and ΔZS values even when they are negative, avoid post-hoc exclusion of a model from the primary analysis, or explicitly restrict the conclusions to TTM only.
minor comments (5)
  1. [5.3.2] The text says 'Figure 16 shows that Chronos fails...' for the Task 3 results, but the Chronos Task 3 results appear in Figure 21; the cross-reference should be corrected.
  2. [Figures 13, 24, 25] These figures are described as percentage improvement over the naive benchmark but their y-axes are labeled 'MSE'; the axis labels should be changed to 'Improvement over baseline (%).'
  3. [Figure 10] The y-axis of Figure 10 shows values as low as -400% for 'Improvement over baseline (%)'; the paper should state the exact formula used to compute this quantity and explain how negative values below -100% arise.
  4. [2.6.1, Table 4] Table 4 and Section 2.6 describe TTM as a multivariate model, but Section 2.6.1 says pretraining is univariate with a channel-independent backbone and that multivariate fine-tuning is optional; the paper should clarify how the multivariate claim applies in zero-shot mode.
  5. [Throughout] A careful proofread is needed: for example, 'V olatility' appears in several table headers, Section 3.5 contains 'are can also cause jumps,' and Section 2.4.1 begins with an unresolved capital 'An.'

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the transfer-gain and sample-efficiency results are empirical paired comparisons, not derivations from fitted parameters or load-bearing self-citations.

full rationale

The paper's central claims (25-50% MSE improvements for fine-tuned pretrained TTM over an identically-architected untrained model, 3-10 fewer years of data, and zero-shot wins over naive baselines) are established by direct experiments defined in Sections 2.8.2-2.8.3 and executed in Sections 4.3-4.4. Equations (2) and (3) define transfer gains as relative error differences between two model initialisations on the same data split; they are not identities that make the conclusion true by construction. No parameter is fitted to the evaluation targets and then renamed a prediction; no benchmark result is derived from the claim it is meant to support; and the paper contains no load-bearing self-citation, since the TTM and Chronos references ([63], [64]) are external sources describing the models under test. The acknowledged risk in Section 6.2 that TTM's pretraining corpus includes Bitcoin data (Appendix A.1) and may overlap the 2014-2024 evaluation windows is a temporal-leakage threat to the interpretation of the transfer gains, not a circularity: the pretrained checkpoint is an external artifact, and the paired comparison remains an empirical measurement. If the pretraining data do overlap the evaluation period, the headline numbers would overstate transfer, but that is a validity and correctness concern outside the circularity taxonomy. Horizon mismatches and single-seed runs are similar robustness limitations, not circular reasoning. Accordingly, no circular step can be exhibited with the required quote-and-reduction evidence, and the appropriate score is 0.

Assumptions & free parameters 9 free parameters · 6 assumptions · 0 invented entities

No new theoretical entities are introduced. The analysis rests on the choice of metrics, thresholds, normalization windows, and pretraining checkpoints, several of which are selected ad hoc or are not independently verifiable from the manuscript.

free parameters (9)
  • Transfer-gain decision threshold = 0.1
    Chosen by hand in Section 2.8.2 to declare a transfer gain significant; no statistical justification.
  • Reference date for sample-efficiency probe = 22 January 2021
    Fixed date used to choose optimal training window length k in Section 4.3; hyperparameters are optimized on this date and do not generalize in rolling evaluation (acknowledged for Task 3).
  • TTM forecast horizons = 30 (bond), 48 (FX), 96 (spread)
    Table 8; differs from stated 21-day task definitions, affecting comparability and the headline '21 business days ahead' claim.
  • TTM context lengths = 90, 512, 512
    Selected by validation performance (Section 4.1.1).
  • TTM epochs = 25 (Task 1), 50 (Task 2), 4 (Task 3)
    Tuned via validation; Task 3 overfits the reference date and deteriorates in other rolling windows.
  • Chronos forecast horizons = 21, 21, 10
    Table 9; not aligned with TTM horizons in the same tasks, making cross-model comparisons difficult.
  • Spread normalization window = 42
    Equations 4-6; chosen for rolling mean and standard deviation of log spread, not justified robustly.
  • VolTransform window = half-year (approx. 126 business days)
    Log plus rolling z-score used for volatility features in Task 2 (Section 3.7); window chosen by the author.
  • Naive baseline selection = zero for changes; previous value for volatility; previous for 5-day spread and zero for 10-day spread
    Baselines chosen per task and horizon based on series properties (Section 4.2), a post-hoc decision that affects measured gains.
assumptions (6)
  • domain assumption The official TTM checkpoint is pretrained on the claimed corpora, including Bitcoin, and is used without modification in zero-shot and fine-tuning.
    Section 2.6.1 and Appendix A.1; the pretraining corpus is not audited, and Bitcoin inclusion creates potential temporal overlap with the 2013-2024 evaluation windows.
  • domain assumption Temporal causality of the evaluation: the pretraining data do not contain information from the test periods used in this study.
    Section 6.2 concedes that pretraining from scratch would be needed to eliminate look-ahead bias, so this assumption is load-bearing and may be false.
  • domain assumption Features aligned at daily frequency from heterogeneous sources (FRED, Yahoo, Macrosynergy, ECB, Investing.com) are correctly merged and stationary after transformations.
    Sections 3.3, 3.7, 3.11; the paper asserts ADF stationarity tests pass but provides no outputs or alignment code.
  • ad hoc to paper MSE averaged over the forecast horizon is the appropriate metric for comparing forecasts at the stated 21-day horizon, even though TTM outputs 30, 48 and 96 steps.
    The paper does not describe how the multi-horizon MSE is aggregated or reconciled with the 21-day task statement.
  • standard math The ECM-to-AR(1) reduction in Appendix A.4 assumes the short-run adjustment term is zero.
    Appendix A.4; this is a stated simplification used to justify ECM's strong baseline performance in Task 3.
  • domain assumption EWA and EWC log prices are cointegrated, so the rolling standardized spread is a valid stationary target.
    Section 3.12; tested with Engle-Granger but no test statistics or critical values are shown.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Time Series Foundation Models for Multivariate Financial Time Series Forecasting." pith.science (2026). https://pith.science/paper/74Y3IAYN

@misc{pith2026250707296,
  author       = {Pith},
  title        = {Pith review of: Time Series Foundation Models for Multivariate Financial Time Series Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/74Y3IAYN}},
  note         = {Machine review of arXiv:2507.07296}
}
read the original abstract

Financial time series forecasting presents significant challenges due to complex nonlinear relationships, temporal dependencies, variable interdependencies and limited data availability, particularly for tasks involving low-frequency data, newly listed instruments, or emerging market assets. Time Series Foundation Models (TSFMs) offer a promising solution through pretraining on diverse time series corpora followed by task-specific adaptation. This study evaluates two TSFMs (Tiny Time Mixers (TTM) and Chronos) across three financial forecasting tasks: US 10-year Treasury yield changes, EUR/USD volatility, and equity spread prediction. Results demonstrate that TTM exhibits strong transferability. When fine-tuning both the pretrained version of TTM and an untrained model with the same architecture, the pretrained version achieved 25-50% better performance when fine-tuned on limited data and 15-30% improvements even when fine-tuned on lengthier datasets. Notably, TTM's zero-shot performance outperformed naive benchmarks in volatility forecasting and equity spread prediction, with the latter demonstrating that TSFMs can surpass traditional benchmark models without fine-tuning. The pretrained model consistently required 3-10 fewer years of data to achieve comparable performance levels compared to the untrained model, demonstrating significant sample-efficiency gains. However, while TTM outperformed naive baselines, traditional specialised models matched or exceeded its performance in two of three tasks, suggesting TSFMs prioritise breadth over task-specific optimisation. These findings indicate that TSFMs, though still nascent, offer substantial promise for financial forecasting-particularly in noisy, data-constrained tasks-but achieving competitive performance likely requires domain-specific pretraining and architectural refinements tailored to financial time series characteristics.

Figures

Figures reproduced from arXiv: 2507.07296 by the authors.

Figure 1
Figure 1. The Development History of Time Series Forecasting Deep Learning models, from [27] [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Types of LLM Adaptation to Time Series Methods, from [29] [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Transformer Architecture, from [19] modern hardware accelerators [19]. Third, the attention operation is highly flexible: it can be made sparse or structured (e.g. ProbSparse or LogSparse attention), combined with frequency-domain blocks (as in FEDformer [58]), or hybridised with decomposition modules (as in Autoformer [59]), enabling practitioners to tailor the trade-off between expressivity and computational cost … view at source ↗
Figures from the paper (23 more)
Figure 4
Figure 4. Figure 4: Architecture of Tiny Time Mixers (TTM), a TSFM Developed by IBM, from [63] [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Chronos Forecasting Process. From left to right: time series tokenisation via scaling and quantisa [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Illustration of Catastrophic Forgetting, where a Foundation Model ”Forgets” the Answer to a [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Types of Parameter Efficient Fine-Tuning (PEFT) Methods, from [70] [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Trade-offs When Using Different Fine-Tuning Methods [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: TTM Mean Squared Error when Fine-tuned on Different Training Dataset Sizes when Forecasting [PITH_FULL_IMAGE:figures/full_fig_p032_9.png]
Figure 10
Figure 10. Figure 10: Rolling Performance of Different Training Settings of TTM Compared to Naive Benchmark when [PITH_FULL_IMAGE:figures/full_fig_p033_10.png]
Figure 11
Figure 11. Figure 11: Chronos-Bolt (Small) Mean Squared Error when Fine-tuned on Different Training Dataset Sizes [PITH_FULL_IMAGE:figures/full_fig_p034_11.png]
Figure 12
Figure 12. Figure 12: Rolling Performance of Different Models when Forecasting 10-Year Treasury Yields 21 Business [PITH_FULL_IMAGE:figures/full_fig_p035_12.png]
Figure 13
Figure 13. Figure 13: Rolling Performance of Different Models relative to Naive Benchmark when Forecasting 10-Year [PITH_FULL_IMAGE:figures/full_fig_p035_13.png]
Figure 14
Figure 14. Figure 14: TTM Mean Squared Error when Fine-tuned on Different Training Dataset Sizes when Forecasting [PITH_FULL_IMAGE:figures/full_fig_p036_14.png]
Figure 15
Figure 15. Figure 15: Rolling Performance of Different Training Settings of TTM Compared to Naive Benchmark when [PITH_FULL_IMAGE:figures/full_fig_p037_15.png]
Figure 16
Figure 16. Figure 16: Chronos-Bolt (Small) Mean Squared Error when Fine-tuned on Different Training Dataset Sizes [PITH_FULL_IMAGE:figures/full_fig_p038_16.png]
Figure 17
Figure 17. Figure 17: Rolling Performance of Different Models when Forecasting EUR/USD Realised Volatility 21 [PITH_FULL_IMAGE:figures/full_fig_p039_17.png]
Figure 18
Figure 18. Figure 18: Rolling Performance of Different Models relative to Naive Benchmark when Forecasting [PITH_FULL_IMAGE:figures/full_fig_p039_18.png]
Figure 19
Figure 19. Figure 19: TTM Mean Squared Error when Fine-tuned on Different Training Dataset Sizes when Forecasting [PITH_FULL_IMAGE:figures/full_fig_p041_19.png]
Figure 20
Figure 20. Figure 20: Rolling Performance of Different Training Settings of TTM Compared to Naive Benchmark when [PITH_FULL_IMAGE:figures/full_fig_p042_20.png]
Figure 21
Figure 21. Figure 21: Chronos-Bolt (Small) Mean Squared Error when Fine-tuned on Different Training Dataset Sizes [PITH_FULL_IMAGE:figures/full_fig_p044_21.png]
Figure 22
Figure 22. Figure 22: Rolling Performance of Different Non-Sequential Models when Forecasting the Spread Between [PITH_FULL_IMAGE:figures/full_fig_p046_22.png]
Figure 23
Figure 23. Figure 23: Rolling Performance of Different Sequential Models when Forecasting the Spread Between MSCI [PITH_FULL_IMAGE:figures/full_fig_p047_23.png]
Figure 24
Figure 24. Figure 24: Rolling Performance of Different Non-Sequential Models Relative to Naive Benchmark when [PITH_FULL_IMAGE:figures/full_fig_p048_24.png]
Figure 25
Figure 25. Figure 25: Rolling Performance of Different Sequential Models Relative to Naive Benchmark when Forecast [PITH_FULL_IMAGE:figures/full_fig_p049_25.png]
Figure 26
Figure 26. Figure 26: Cumulative Returns of Three Signals Constructed using TTM’s forecasts Alongside a Benchmark [PITH_FULL_IMAGE:figures/full_fig_p064_26.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Forecasting Realized Volatility with Time Series Foundation Models: A Comparison with Econometric Benchmarks

    q-fin.ST 2026-07 accept novelty 6.0 of 10

    Zero-shot time series foundation models largely fail to beat econometric benchmarks for realized volatility forecasting, with only TTM achieving a narrow, calibration-driven edge.

  2. Towards Causal Market Simulators

    cs.LG 2025-11 reject novelty 4.0 of 10

    A VAE with a DAG-constrained decoder is proposed to generate counterfactual financial time series; on two synthetic AR(1) models it matches analytical counterfactual probabilities to within 0.03–0.10 L1 error.

Reference graph

Works this paper leans on

117 extracted references · 25 canonical work pages · cited by 2 Pith papers

  1. [1]

    A combination of artificial neural network and random walk models for financial time series forecasting,

    R. Adhikari and R. K. Agrawal, “A combination of artificial neural network and random walk models for financial time series forecasting,” Neural Computing and Applications, vol. 24, no. 6, pp. 1441– 1449, 2014, ISSN : 0941-0643. DOI: 10.1007/s00521-013-1386-y

    work page doi:10.1007/s00521-013-1386-y 2014

  2. [2]

    Financial time series forecasting: A comprehensive review of signal processing and optimization-driven intelligent models,

    M. Praveen, S. Dekka, D. M. Sai, D. P. Chennamsetty, and D. P. Chinta, “Financial time series forecasting: A comprehensive review of signal processing and optimization-driven intelligent models,” Computational Economics, pp. 1–27, 2025, ISSN : 0927-7099. DOI: 10.1007/s10614-025-108 99-z

    work page doi:10.1007/s10614-025-108 2025

  3. [3]

    Makridakis, “Time series prediction: Forecasting the future and understanding the past andreas s

    S. Makridakis, “Time series prediction: Forecasting the future and understanding the past andreas s. weigend and neil a. gershenfeld, eds., 1993, (addison-wesley publishing company, reading, MA, USA), 643 pp., paperback $49.50, ISBN 0-201-62,” International Journal of Forecasting, vol. 10, no. 3, pp. 463–466, 1994, ISSN : 0169-2070. DOI: 10.1016/0169-2070...

    work page doi:10.1016/0169-2070(94)90077-9 1993

  4. [4]

    Forecasting economic time series using targeted predictors,

    J. Bai and S. Ng, “Forecasting economic time series using targeted predictors,” Journal of Economet- rics, vol. 146, no. 2, pp. 304–317, 2008, ISSN : 0304-4076. DOI: 10.1016/j.jeconom.2008.0 8.010

    work page doi:10.1016/j.jeconom.2008.0 2008

  5. [5]

    Weather forecasting with ensemble methods,

    T. Gneiting and A. E. Raftery, “Weather forecasting with ensemble methods,” Science, vol. 310, no. 5746, pp. 248–249, 2005, ISSN : 0036-8075. DOI: 10.1126/science.1115255

    doi:10.1126/science.1115255 2005

  6. [6]

    Forecasting energy consumption time series using machine learning techniques based on usage patterns of residential householders,

    J.-S. Chou and D. -S. Tran, “Forecasting energy consumption time series using machine learning techniques based on usage patterns of residential householders,” Energy, vol. 165, pp. 709–726, Dec

  7. [7]

    Forecasting the future: A comprehensive review of time series prediction techniques,

    S. Arora Milind Kolambe, “Forecasting the future: A comprehensive review of time series prediction techniques,” Journal of Electrical Systems, vol. 20, no. 2s, pp. 575–586, 2024. DOI: 10.52783/je s.1478

    work page doi:10.52783/je 2024

  8. [8]

    Deep learning-based time series forecasting,

    X. Song, L. Deng, H. Wang, Y . Zhang, Y . He, and W. Cao, “Deep learning-based time series forecasting,” Artificial Intelligence Review, vol. 58, no. 1, Nov. 2024, ISSN : 1573-7462. DOI: 10.10 07/s10462-024-10989-8 . [Online]. Available: http://dx.doi.org/10.1007/s104 62-024-10989-8

    doi:10.1007/s104 2024

Show all 117 references
  1. [9]

    Traffic flow prediction with big data: A deep learning approach,

    Y . Lv, Y . Duan, W. Kang, Z. Li, and F.-Y . Wang, “Traffic flow prediction with big data: A deep learning approach,” IEEE Transactions on Intelligent Transportation Systems, pp. 1–9, 2014, ISSN : 1558-0016. DOI: 10.1109/tits.2014.2345663. [Online]. Available: http://dx.doi.o ...

    2014

  2. [10]

    A deep learning based stock trading model with 2-d cnn trend detection,

    M. U. Gudelek, S. A. Boluk, and A. M. Ozbayoglu, “A deep learning based stock trading model with 2-d cnn trend detection,” in 2017 IEEE Symposium Series on Computational Intelligence (SSCI), IEEE, Nov. 2017, pp. 1–8. DOI: 10.1109/ssci.2017.8285188. [Online]. Available: http://...

    2017

  3. [11]

    Convolutional neural networks for forex time series forecasting,

    M. Markova, “Convolutional neural networks for forex time series forecasting,” in EIGHTH IN- TERNATIONAL CONFERENCE NEW TRENDS IN THE APPLICATIONS OF DIFFERENTIAL EQUATIONS IN SCIENCES (NTADES2021), vol. 2459, AIP Publishing, 2022, p. 030 024. DOI: 10 .1063/5.0083533. [Online]...

    2022 doi

  4. [12]

    S. Bai, J. Z. Kolter, and V . Koltun,An empirical evaluation of generic convolutional and recurrent networks for sequence modeling , 2018. DOI: 10 . 48550 / ARXIV . 1803 . 01271. [Online]. Available: https://arxiv.org/abs/1803.01271

    2018 doi arXiv

  5. [13]

    X. Shi, Z. Chen, H. Wang, D.-Y . Yeung, W.-k. Wong, and W.-c. Woo, Convolutional lstm network: A machine learning approach for precipitation nowcasting, 2015. DOI: 10.48550/ARXIV.1506 .04214. [Online]. Available: https://arxiv.org/abs/1506.04214

    2015 doi arXiv

  6. [14]

    Dey and F

    R. Dey and F. M. Salem, Gate-variants of gated recurrent unit (gru) neural networks, 2017. DOI: 10.48550/ARXIV.1701.05923. [Online]. Available: https://arxiv.org/abs/1701 .05923

    2017 doi arXiv

  7. [15]

    Learning long-term dependencies with gradient descent is difficult,

    Y . Bengio, P. Simard, and P. Frasconi, “Learning long-term dependencies with gradient descent is difficult,” IEEE Transactions on Neural Networks , vol. 5, no. 2, pp. 157–166, 1994. [Online]. Available: https://ieeexplore.ieee.org/document/279181

    1994

  8. [16]

    On the difficulty of training recurrent neural networks,

    R. Pascanu, T. Mikolov, and Y . Bengio, “On the difficulty of training recurrent neural networks,” in Proceedings of the 30th International Conference on Machine Learning, 2013, pp. 1310–1318. [Online]. Available: https://proceedings.mlr.press/v28/pascanu13.pdf

    2013

  9. [17]

    Building trend fuzzy granulation-based lstm recurrent neural network for long-term time-series forecasting,

    Y . Tang, F. Yu, W. Pedrycz, X. Yang, J. Wang, and S. Liu, “Building trend fuzzy granulation-based lstm recurrent neural network for long-term time-series forecasting,”IEEE Transactions on Fuzzy Systems, vol. 30, no. 6, pp. 1599–1613, Jun. 2022, ISSN : 1941-0034. DOI: 10.1109/...

    2022

  10. [18]

    Assessment of deep recurrent neural network-based strategies for short-term building energy predictions,

    C. Fan, J. Wang, W. Gang, and S. Li, “Assessment of deep recurrent neural network-based strategies for short-term building energy predictions,” Applied Energy, vol. 236, pp. 700–710, Feb. 2019, ISSN : 0306-2619. DOI: 10.1016/j.apenergy.2018.12.004 . [Online]. Available: http:/...

    2019 doi

  11. [19]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, et al. , Attention is all you need , 2017. DOI: 10 . 48550 /ARXIV.1706.03762. [Online]. Available: https://arxiv.org/abs/1706.03762

    2017 doi arXiv

  12. [20]

    Devlin, M.-W

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova,Bert: Pre-training of deep bidirectional trans- formers for language understanding, 2018. DOI: 10.48550/ARXIV.1810.04805 . [Online]. Available: https://arxiv.org/abs/1810.04805

    2018 doi arXiv

  13. [21]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, et al., An image is worth 16x16 words: Transformers for image recognition at scale, 2020. DOI: 10.48550/ARXIV.2010.11929 . [Online]. Available: https://arxiv.org/abs/2010.11929

    2020 doi arXiv

  14. [22]

    S. Li, X. Jin, Y . Xuan,et al., Enhancing the locality and breaking the memory bottleneck of transformer on time series forecasting, 2019. DOI: 10.48550/ARXIV.1907.00235 . [Online]. Available: https://arxiv.org/abs/1907.00235

    2019 doi arXiv

  15. [23]

    Adversarial sparse transformer for time series forecasting,

    S. Wu, X. Xiao, Q. Ding, P. Zhao, Y . Wei, and J. Huang, “Adversarial sparse transformer for time series forecasting,” inAdvances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin, Eds., vol. 33, Curran Associates, Inc., 202...

    2020

  16. [24]

    Anomaly transformer: Time series anomaly detection with association discrepancy,

    J. Xu, H. Wu, J. Wang, and M. Long, “Anomaly transformer: Time series anomaly detection with association discrepancy,” inProceedings of the International Conference on Learning Representations (ICLR), 2022. [Online]. Available: https://openreview.net/forum?id=LzQQ89U1qm_

    2022

  17. [25]

    H. Zhou, S. Zhang, J. Peng, et al., Informer: Beyond efficient transformer for long sequence time- series forecasting, 2020. DOI: 10.48550/ARXIV.2012.07436. [Online]. Available: https: //arxiv.org/abs/2012.07436. 55

    2020 doi arXiv

  18. [26]

    A. Zeng, M. Chen, L. Zhang, and Q. Xu, Are transformers effective for time series forecasting? 2022. DOI: 10.48550/ARXIV.2205.13504. [Online]. Available: https://arxiv.org/abs/2 205.13504

    2022 doi arXiv

  19. [27]

    A systematic review for transformer-based long-term series forecasting,

    L. Su, X. Zuo, R. Li, X. Wang, H. Zhao, and B. Huang, “A systematic review for transformer-based long-term series forecasting,” Artificial Intelligence Review, vol. 58, no. 3, Jan. 2025, ISSN : 1573-

    2025

  20. [28]

    Z. Zhao, Y . Shi, S. Wu, F. Yang, W. Song, and N. Liu,Interpretation of time-series deep models: A survey, 2023. DOI: 10.48550/ARXIV.2305.14582. [Online]. Available: https://arxiv .org/abs/2305.14582

    2023 doi arXiv

  21. [29]

    J. Ye, W. Zhang, K. Yi,et al., A survey of time series foundation models: Generalizing time series representation with large language model, 2024. DOI: 10.48550/arxiv.2405.02358. eprint: 2405.02358

    2024 doi

  22. [30]

    Does the performance of banking sector promote economic growth? a time series analysis,

    K. Reddy, S. Mubeen, K. H. H. Raju, J. V ., and M. B. S., “Does the performance of banking sector promote economic growth? a time series analysis,” International Journal of Professional Business Review, vol. 8, no. 6, e02128, Jun. 2023, ISSN : 2525-3654. DOI: 10.26668/business...

    2023 doi

  23. [31]

    P. Liu, W. Yuan, J. Fu, Z. Jiang, H. Hayashi, and G. Neubig, Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing , 2021. DOI: 10.48550 /ARXIV.2107.13586. [Online]. Available: https://arxiv.org/abs/2107.13586

    2021 doi arXiv

  24. [32]

    Baltruˇsaitis, C

    T. Baltruˇsaitis, C. Ahuja, and L.-P. Morency, Multimodal machine learning: A survey and taxonomy,

  25. [33]

    Deep unsupervised domain adaptation with time series sensor data: A survey,

    Y . Shi, X. Ying, and J. Yang, “Deep unsupervised domain adaptation with time series sensor data: A survey,”Sensors, vol. 22, no. 15, p. 5507, Jul. 2022,ISSN : 1424-8220. DOI: 10.3390/s22155507. [Online]. Available: http://dx.doi.org/10.3390/s22155507

    2022 doi

  26. [34]

    Language models are unsuper- vised multitask learners,

    A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, and I. Sutskever, “Language models are unsuper- vised multitask learners,” OpenAI, 2019. [Online]. Available: https://cdn.openai.com/bet ter-language-models/language_models_are_unsupervised_multitask_lea rners.pdf

    2019

  27. [35]

    X. Sun, H. Cheng, J. Li, B. Liu, and J. Guan, All in one: Multi-task prompting for graph neural networks, 2023. DOI: 10.48550/ARXIV.2307.01504. [Online]. Available: https://arxi v.org/abs/2307.01504

    2023 doi arXiv

  28. [36]

    E. J. Hu, Y . Shen, P. Wallis,et al., Lora: Low-rank adaptation of large language models, arXiv preprint arXiv:2106.09685, 2021. [Online]. Available: https://arxiv.org/pdf/2106.09685

    2021 arXiv

  29. [37]

    Q. Xie, W. Han, Y . Lai, M. Peng, and J. Huang,The wall street neophyte: A zero-shot analysis of chatgpt over multimodal stock movement prediction challenges, 2023. DOI: 10.48550/ARXIV.2 304.05351. [Online]. Available: https://arxiv.org/abs/2304.05351

    2023 doi arXiv

  30. [38]

    X. Yu, Z. Chen, Y . Ling, S. Dong, Z. Liu, and Y . Lu,Temporal data meets llm – explainable financial time series forecasting , 2023. DOI: 10 . 48550 / ARXIV . 2306 . 11025. [Online]. Available: https://arxiv.org/abs/2306.11025

    2023 doi arXiv

  31. [39]

    Large language models for financial aid in financial time-series forecasting,

    M. K. Islam, A. Karmacharya, T. Sue, and J. Fox, “Large language models for financial aid in financial time-series forecasting,” in 2024 IEEE International Conference on Big Data (BigData), IEEE, Dec. 2024, pp. 4892–4895. DOI: 10.1109/bigdata62323.2024.10824953. [Online]. Avai...

    2024

  32. [40]

    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 preprint arXiv:2310.10688, 2024. [Online]. Available: https://arxiv.org/abs/2310 .10688

    2024 arXiv

  33. [41]

    X. Fu, M. Hirano, and K. Imajo, Financial fine-tuning a large time series model, 2024. DOI: 10.485 50/ARXIV.2412.09880. [Online]. Available: https://arxiv.org/abs/2412.09880

    2024 doi arXiv

  34. [42]

    G. Woo, C. Liu, A. Kumar, C. Xiong, S. Savarese, and D. Sahoo, Unified training of universal time series forecasting transformers, 2024. DOI: 10.48550/ARXIV.2402.02592 . [Online]. Available: https://arxiv.org/abs/2402.02592

    2024 doi arXiv

  35. [43]

    Anomaly detection for vietnamese financial market,

    Q. T. Huynh, T. H. Nguyen, D. T. Vu, and M. M. Ngo, “Anomaly detection for vietnamese financial market,” in 2024 18th International Conference on Advanced Computing and Analytics (ACOMPA), IEEE, Nov. 2024, pp. 58–62. DOI: 10.1109/acompa64883.2024.00016. [Online]. Available: ht...

    2024

  36. [44]

    Q. Wen, L. Sun, F. Yang,et al., Time series data augmentation for deep learning: A survey, arXiv preprint arXiv:2002.12478, 2020. [Online]. Available: https://arxiv.org/pdf/2002.124 78

    2002 arXiv

  37. [45]

    Predicting extreme financial risks on imbalanced dataset: A combined kernel fcm and kernel smote based svm classifier,

    X. Huang, C.-Z. Zhang, and J. Yuan, “Predicting extreme financial risks on imbalanced dataset: A combined kernel fcm and kernel smote based svm classifier,” Computational Economics, vol. 56, pp. 187–216, 2020. DOI: 10.1007/s10614-020-09975-3 . [Online]. Available: https://l in...

    2020 doi

  38. [46]

    Learning with imbalanced data in smart manufacturing: A comparative study,

    Y . Fathy, M. Jaber, and A. Brintrup, “Learning with imbalanced data in smart manufacturing: A comparative study,” IEEE Access, vol. 9, pp. 2734–2757, 2021. [Online]. Available: https://iee explore.ieee.org/document/9309288

    2021

  39. [47]

    A deep learning based expert framework for portfolio prediction and forecasting,

    F. Jeribi, R. J. Martin, R. Mittal, et al., “A deep learning based expert framework for portfolio prediction and forecasting,” IEEE Access, vol. 12, pp. 103 810–103 829, 2024. DOI: 10 . 1109 /ACCESS.2024.3434528. [Online]. Available: https://ieeexplore.ieee.org/docu ment/10613392

    2024

  40. [48]

    Transfer learning for class imbalance problems with inadequate data,

    S. Al-Stouhi and C. K. Reddy, “Transfer learning for class imbalance problems with inadequate data,” Knowledge and Information Systems, vol. 48, no. 1, pp. 201–228, 2016. DOI: 10.1007/s10115- 015-0870-3 . [Online]. Available: https://link.springer.com/article/10.1007 /s10115-0...

    2016 doi

  41. [49]

    A survey on transfer learning,

    S. J. Pan and Q. Yang, “A survey on transfer learning,” IEEE Transactions on Knowledge and Data Engineering, vol. 22, no. 10, pp. 1345–1359, 2010. DOI: 10.1109/TKDE.2009.191. [Online]. Available: https://ieeexplore.ieee.org/document/5288526

    2010

  42. [50]

    A brief review of domain adaptation,

    A. Farahani, S. V oghoei, K. Rasheed, and H. R. Arabnia, “A brief review of domain adaptation,” in Advances in Data Science and Information Engineering, ser. Transactions on Computational Science and Computational Intelligence, 2021. DOI: 10.1007/978-3-030-71704-9_65 . [Online...

    2021 doi

  43. [51]

    A novel deep transfer learning framework with adversarial domain adaptation: Application to financial time-series forecasting,

    D. Zhang, R. Lin, T. Wei, L. Ling, and J. Huang, “A novel deep transfer learning framework with adversarial domain adaptation: Application to financial time-series forecasting,” Neural Computing and Applications, vol. 35, pp. 24 037–24 054, 2023.DOI: 10.1007/s00521-023-09047-1

    2023 doi

  44. [52]

    Garza, C

    A. Garza, C. Challu, and M. Mergenthaler-Canseco, Timegpt-1, 2023. DOI: 10.48550/ARXIV.2 310.03589. [Online]. Available: https://arxiv.org/abs/2310.03589

    2023 doi arXiv

  45. [53]

    C.-C. M. Yeh, X. Dai, H. Chen,et al., Toward a foundation model for time series data, 2023. DOI: 10.48550/ARXIV.2310.03916. [Online]. Available: https://arxiv.org/abs/2310 .03916. 57

    2023 doi arXiv

  46. [54]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural Computation, vol. 9, no. 8, pp. 1735–1780, Nov. 1997, ISSN : 1530-888X. DOI: 10.1162/neco.1997.9.8.1735. [Online]. Available: http://dx.doi.org/10.1162/neco.1997.9.8.1735

    1997 doi

  47. [55]

    K. He, X. Zhang, S. Ren, and J. Sun, Deep residual learning for image recognition , 2015. DOI: 10.48550/ARXIV.1512.03385. [Online]. Available: https://arxiv.org/abs/1512 .03385

    2015 doi arXiv

  48. [56]

    Deep learning for time series classification: A review,

    H. Ismail Fawaz, G. Forestier, J. Weber, L. Idoumghar, and P.-A. Muller, “Deep learning for time series classification: A review,” Data Mining and Knowledge Discovery, vol. 33, no. 4, pp. 917–963, Mar. 2019, ISSN : 1573-756X. DOI: 10.1007/s10618- 019- 00619- 1 . [Online]. Avai...

    2019 doi

  49. [57]

    Temporal fusion transformers for interpretable multi- horizon time series forecasting,

    B. Lim, S. ¨O. Arık, N. Loeff, and T. Pfister, “Temporal fusion transformers for interpretable multi- horizon time series forecasting,” International Journal of Forecasting, vol. 37, no. 4, pp. 1748– 1764, Oct. 2021, ISSN : 0169-2070. DOI: 10.1016/j.ijforecast.2021.03.012. [On...

    2021 doi

  50. [58]

    T. Zhou, Z. Ma, Q. Wen, X. Wang, L. Sun, and R. Jin,Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting , 2022. DOI: 10 . 48550 / ARXIV . 2201 . 12740. [Online]. Available: https://arxiv.org/abs/2201.12740

    2022 doi arXiv

  51. [59]

    H. Wu, J. Xu, J. Wang, and M. Long, Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting, 2021. DOI: 10.48550/ARXIV.2106.13008. [Online]. Avail- able: https://arxiv.org/abs/2106.13008

    2021 doi arXiv

  52. [60]

    Y . Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam, A time series is worth 64 words: Long-term forecasting with transformers, 2022. DOI: 10.48550/ARXIV.2211.14730. [Online]. Available: https://arxiv.org/abs/2211.14730

    2022 doi arXiv

  53. [61]

    Rasul, A

    K. Rasul, A. Ashok, A. R. Williams, et al., Lag-llama: Towards foundation models for probabilistic time series forecasting , 2023. DOI: 10 . 48550 / ARXIV . 2310 . 08278. [Online]. Available: https://arxiv.org/abs/2310.08278

    2023 doi arXiv

  54. [62]

    C. Feng, L. Huang, and D. Krompass, Only the curve shape matters: Training foundation models for zero-shot multivariate time series forecasting through next curve shape prediction, 2024. DOI: 10.48550/ARXIV.2402.07570. [Online]. Available: https://arxiv.org/abs/2402 .07570

    2024 doi arXiv

  55. [63]

    Tiny time mixers (TTMs): Fast pre-trained models for enhanced zero/few-shot forecasting of multivariate time series,

    V . Ekambaram, A. Jati, P. Dayama,et al., “Tiny time mixers (TTMs): Fast pre-trained models for enhanced zero/few-shot forecasting of multivariate time series,” arXiv, 2024. DOI: 10.48550/arx iv.2401.03955. eprint: 2401.03955

    2024 doi arXiv

  56. [64]

    A. F. Ansari, L. Stella, C. Turkmen, et al., Chronos: Learning the language of time series, 2024. DOI: 10.48550/ARXIV.2403.07815. [Online]. Available: https://arxiv.org/abs/2403 .07815

    2024 doi arXiv

  57. [65]

    Moment: A family of open time-series foundation models,

    M. Goswami, K. Szafer, A. Choudhry, Y . Cai, S. Li, and A. Dubrawski, “Moment: A family of open time-series foundation models,” in International Conference on Machine Learning, 2024

    2024

  58. [66]

    A. W. Services, Chronos-t5-small, hugging face model card, https://huggingface.co/ama zon/chronos-t5-small, Accessed: 2025-05-15, 2024

    2025

  59. [67]

    K. Lv, Y . Yang, T. Liu, Q. Gao, Q. Guo, and X. Qiu,Full parameter fine-tuning for large language models with limited resources, 2023. DOI: 10.48550/ARXIV.2306.09782. [Online]. Available: https://arxiv.org/abs/2306.09782. 58

    2023 doi arXiv

  60. [68]

    S. Song, H. Xu, J. Ma, et al., How to alleviate catastrophic forgetting in llms finetuning? hierarchical layer-wise and element-wise regularization , 2025. DOI: 10 . 48550 / ARXIV . 2501 . 13669. [Online]. Available: https://arxiv.org/abs/2501.13669

    2025 doi arXiv

  61. [69]

    Lialin, V

    V . Lialin, V . Deshpande, X. Yao, and A. Rumshisky,Scaling down to scale up: A guide to parameter- efficient fine-tuning, 2023. DOI: 10.48550/ARXIV.2303.15647. [Online]. Available: https: //arxiv.org/abs/2303.15647

    2023 doi arXiv

  62. [70]

    Adik, PEFT (Parameter-Efficient Fine-Tuning), https://medium.com/@kanikaadik07 /peft-parameter-efficient-fine-tuning-55e32c60c799 , Accessed: 2025-04-12, 2023

    K. Adik, PEFT (Parameter-Efficient Fine-Tuning), https://medium.com/@kanikaadik07 /peft-parameter-efficient-fine-tuning-55e32c60c799 , Accessed: 2025-04-12, 2023

    2025

  63. [71]

    Donahue, Y

    J. Donahue, Y . Jia, O. Vinyals,et al., Decaf: A deep convolutional activation feature for generic visual recognition, 2013. DOI: 10.48550/ARXIV.1310.1531 . [Online]. Available: https: //arxiv.org/abs/1310.1531

    2013 doi arXiv

  64. [72]

    E. B. Zaken, S. Ravfogel, and Y . Goldberg,Bitfit: Simple parameter-efficient fine-tuning for transformer- based masked language-models, 2021. DOI: 10.48550/ARXIV.2106.10199. [Online]. Avail- able: https://arxiv.org/abs/2106.10199

    2021 doi

  65. [73]

    B. Zhao, H. Tu, C. Wei, J. Mei, and C. Xie, Tuning layernorm in attention: Towards efficient multi-modal llm finetuning, 2023. DOI: 10.48550/ARXIV.2312.11420. [Online]. Available: https://arxiv.org/abs/2312.11420

    2023 doi arXiv

  66. [74]

    W. J. Maddox, G. Benton, and A. G. Wilson,Rethinking parameter counting in deep models: Effective dimensionality revisited, 2020. DOI: 10.48550/ARXIV.2003.02139. [Online]. Available: htt ps://arxiv.org/abs/2003.02139

    2020 doi arXiv

  67. [75]

    C. Li, H. Farkhoor, R. Liu, and J. Yosinski,Measuring the intrinsic dimension of objective landscapes,

  68. [76]

    Aghajanyan, L

    A. Aghajanyan, L. Zettlemoyer, and S. Gupta, Intrinsic dimensionality explains the effectiveness of language model fine-tuning, 2020. DOI: 10.48550/ARXIV.2012.13255. [Online]. Available: https://arxiv.org/abs/2012.13255

    2020 doi arXiv

  69. [77]

    Edalati, M

    A. Edalati, M. Tahaei, I. Kobyzev, V . P. Nia, J. J. Clark, and M. Rezagholizadeh,Krona: Parameter efficient tuning with kronecker adapter, 2022. DOI: 10.48550/ARXIV.2212.10650. [Online]. Available: https://arxiv.org/abs/2212.10650

    2022 doi arXiv

  70. [78]

    Houlsby, A

    N. Houlsby, A. Giurgiu, S. Jastrzebski, et al., Parameter-efficient transfer learning for nlp, 2019. DOI: 10.48550/ARXIV.1902.00751. [Online]. Available: https://arxiv.org/abs/1 902.00751

    2019 doi arXiv

  71. [79]

    Pfeiffer, A

    J. Pfeiffer, A. Kamath, A. R ¨uckl´e, K. Cho, and I. Gurevych, Adapterfusion: Non-destructive task composition for transfer learning , 2020. DOI: 10 . 48550 / ARXIV . 2005 . 00247. [Online]. Available: https://arxiv.org/abs/2005.00247

    2020 doi arXiv

  72. [80]

    [Online]

    DOI: 10.48550/ARXIV.1804.08838. [Online]. Available: https://arxiv.org/a bs/1804.08838

    doi arXiv

  73. [81]

    S. He, L. Ding, D. Dong, M. Zhang, and D. Tao, Sparseadapter: An easy approach for improving the parameter-efficiency of adapters, 2022. DOI: 10.48550/ARXIV.2210.04284 . [Online]. Available: https://arxiv.org/abs/2210.04284

    2022 doi arXiv

  74. [82]

    X. Liu, Y . Zheng, Z. Du,et al., Gpt understands, too, 2021. DOI: 10.48550/ARXIV.2103.103

    2021 doi

  75. [83]

    A review on transferability estimation in deep transfer learning,

    Y . Xue, R. Yang, X. Chen, W. Liu, Z. Wang, and X. Liu, “A review on transferability estimation in deep transfer learning,” IEEE Transactions on Artificial Intelligence, vol. 5, no. 12, pp. 5894– 5914, Dec. 2024, ISSN : 2691-4581. DOI: 10.1109/tai.2024.3445892. [Online]. Avail...

    2024

  76. [84]

    A cointegration analysis of treasury bill yields,

    A. D. Hall, H. M. Anderson, and C. W. J. Granger, “A cointegration analysis of treasury bill yields,” The Review of Economics and Statistics, vol. 74, no. 1, p. 116, Feb. 1992, ISSN : 0034-6535. DOI: 10.2307/2109549. [Online]. Available: http://dx.doi.org/10.2307/2109549

    1992 doi

  77. [85]

    Y . Zhu, J. Feng, C. Zhao, M. Wang, and L. Li,Counter-interference adapter for multilingual machine translation, 2021. DOI: 10.48550/ARXIV.2104.08154. [Online]. Available: https://arx iv.org/abs/2104.08154

    2021 doi arXiv

  78. [86]

    A no-arbitrage vector autoregression of term structure dynamics with macroeconomic and latent variables,

    A. Ang and M. Piazzesi, “A no-arbitrage vector autoregression of term structure dynamics with macroeconomic and latent variables,” Journal of Monetary Economics, vol. 50, no. 4, pp. 745–787, May 2003, ISSN : 0304-3932. DOI: 10.1016/s0304-3932(03)00032-1. [Online]. Available: h...

    2003 doi

  79. [87]

    Price forecast of treasury bond market yield: Optimize method based on deep learning model,

    W. Ping, Y . Hu, and L. Luo, “Price forecast of treasury bond market yield: Optimize method based on deep learning model,” IEEE Access, vol. 12, pp. 194 521–194 539, 2024,ISSN : 2169-3536. DOI: 10.1109/access.2024.3519438. [Online]. Available: http://dx.doi.org/10.1109 /ACCESS...

    2024

  80. [88]

    Available: https://arxiv.org/abs/2103.10385

    [Online]. Available: https://arxiv.org/abs/2103.10385. 59

    arXiv

  81. [89]

    Department of the Treasury, Daily treasury par yield curve rates, 2025

    U.S. Department of the Treasury, Daily treasury par yield curve rates, 2025. [Online]. Available: https://home.treasury.gov/resource-center/data-chart-center/inter est-rates/TextView?type=daily_treasury_yield_curve

    2025

  82. [90]

    Louis, Federal reserve economic data (fred), Accessed: 2025-05-22,

    Federal Reserve Bank of St. Louis, Federal reserve economic data (fred), Accessed: 2025-05-22,

    2025

  83. [91]

    Forecasting interest rates,

    G. Duffee, “Forecasting interest rates,” in Handbook of Economic Forecasting . Elsevier, 2013, pp. 385–426. DOI: 10 . 1016 / b978 - 0 - 444 - 53683 - 9 . 00007 - 4. [Online]. Available: http://dx.doi.org/10.1016/B978-0-444-53683-9.00007-4

    2013 doi

  84. [92]

    Forecasting volatility in financial markets: A review,

    S.-H. Poon and C. W. J. Granger, “Forecasting volatility in financial markets: A review,”Journal of Economic Literature, vol. 41, no. 2, pp. 478–539, Jun. 2003, ISSN : 0022-0515. DOI: 10.1257/jel .41.2.478. [Online]. Available: http://dx.doi.org/10.1257/jel.41.2.478

    2003 doi

  85. [93]

    Chapter 49 arch models,

    T. Bollerslev, R. F. Engle, and D. B. Nelson, “Chapter 49 arch models,” inHandbook of Econometrics. Elsevier, 1994, pp. 2959–3038, ISBN : 9780444887665. DOI: 10.1016/s1573-4412(05)8001 8-2. [Online]. Available: http://dx.doi.org/10.1016/S1573-4412(05)80018-2

    1994 doi

  86. [94]

    J. S. Walia, A. Sinha, S. Srinivasan, and S. Unnikrishnan, Predicting liquidity-aware bond yields using causal gans and deep reinforcement learning with llm evaluation , 2025. DOI: 10.48550 /ARXIV.2502.17011. [Online]. Available: https://arxiv.org/abs/2502.17011

    2025 doi arXiv

  87. [95]

    The volatility of realized volatility,

    F. Corsi, S. Mittnik, C. Pigorsch, and U. Pigorsch, “The volatility of realized volatility,”Econometric Reviews, vol. 27, no. 1–3, pp. 46–78, Feb. 2008, ISSN : 1532-4168. DOI: 10.1080/07474930701 853616. [Online]. Available: http://dx.doi.org/10.1080/07474930701853616

    2008 doi

  88. [96]

    Realized volatility forecasting with neural networks,

    A. Bucci, “Realized volatility forecasting with neural networks,” Journal of Financial Econometrics, vol. 18, no. 3, pp. 502–531, 2020, ISSN : 1479-8417. DOI: 10 . 1093 / jjfinec / nbaa008. [Online]. Available: http://dx.doi.org/10.1093/jjfinec/nbaa008. 60

    2020 doi

  89. [97]

    A forecast comparison of volatility models: Does anything beat a garch(1, 1)?

    P. R. Hansen and A. Lunde, “A forecast comparison of volatility models: Does anything beat a garch(1, 1)?” SSRN Electronic Journal, 2001, ISSN : 1556-5068. DOI: 10.2139/ssrn.264571. [Online]. Available: http://dx.doi.org/10.2139/ssrn.264571

    2001 doi

  90. [98]

    Ltd., Quantamental indicators on jpmaqs, Accessed: 2025-05-22, 2025

    M. Ltd., Quantamental indicators on jpmaqs, Accessed: 2025-05-22, 2025. [Online]. Available: htt ps://macrosynergy.com/academy/quantamental-indicators-on-jpmaqs/

    2025

  91. [99]

    [Online]

    European Central Bank, Ecb data portal, 2025. [Online]. Available: https://data.ecb.euro pa.eu/data/datasets

    2025

  92. [100]

    [Online]

    Investing.com, Investing.com – financial markets worldwide, 2025. [Online]. Available: https: //www.investing.com

    2025

  93. [101]

    Modeling and forecasting realized volatility,

    T. G. Andersen, T. Bollerslev, F. X. Diebold, and P. Labys, “Modeling and forecasting realized volatility,” Econometrica, vol. 71, no. 2, pp. 579–625, Mar. 2003, ISSN : 1468-0262. DOI: 10.1111 /1468-0262.00418. [Online]. Available: http://dx.doi.org/10.1111/1468-0262 .00418

    2003

  94. [102]

    Pairs trading on international etfs,

    P. Schizas, D. D. Thomakos, and T. Wang, “Pairs trading on international etfs,” SSRN Electronic Journal, 2011, ISSN : 1556-5068. DOI: 10.2139/ssrn.1958546 . [Online]. Available: http: //dx.doi.org/10.2139/ssrn.1958546

    2011 doi

  95. [103]

    Pairs trading: Performance of a relative value arbitrage rule,

    E. G. Gatev, W. N. Goetzmann, and K. G. Rouwenhorst, “Pairs trading: Performance of a relative value arbitrage rule,” SSRN Electronic Journal, 2006, ISSN : 1556-5068. DOI: 10.2139/ssrn.141615. [Online]. Available: http://dx.doi.org/10.2139/ssrn.141615

    2006 doi

  96. [104]

    Zhu, Examining pairs trading profitability, Senior essay submitted to the Department of Economics, Yale University., 2024

    X. Zhu, Examining pairs trading profitability, Senior essay submitted to the Department of Economics, Yale University., 2024. [Online]. Available:https://economics.yale.edu/sites/defa ult/files/2024-05/Zhu_Pairs_Trading.pdf

    2024

  97. [105]

    Answering the skeptics: Yes, standard volatility models do provide accurate forecasts,

    T. G. Andersen and T. Bollerslev, “Answering the skeptics: Yes, standard volatility models do provide accurate forecasts,” International Economic Review, vol. 39, no. 4, p. 885, Nov. 1998, ISSN : 0020-

    1998

  98. [106]

    Costa and R

    M. Costa and R. Sueppel,Cross-country equity risk allocation with statistical learning, 2025. [Online]. Available: https://macrosynergy.com/research/cross-country-equity-risk -allocation-with-statistical-learning/

    2025

  99. [107]

    Smith, Cointegrated etf pairs part ii, 2017

    C. Smith, Cointegrated etf pairs part ii, 2017. [Online]. Available: https://quantoisseur.c om/2017/01/20/cointegrated-etf-pairs-part-ii/

    2017

  100. [108]

    Loshchilov and F

    I. Loshchilov and F. Hutter, Decoupled weight decay regularization, 2017. DOI: 10.48550/ARXI V.1711.05101. [Online]. Available: https://arxiv.org/abs/1711.05101

    2017 doi arXiv

  101. [109]

    On the estimation of security price volatilities from historical data,

    M. B. Garman and M. J. Klass, “On the estimation of security price volatilities from historical data,” The Journal of Business, vol. 53, no. 1, p. 67, Jan. 1980, ISSN : 1537-5374. DOI: 10.1086/296072. [Online]. Available: http://dx.doi.org/10.1086/296072

    1980 doi

  102. [113]

    Mitchell, How to use a pairs trading strategy with etfs, 2015

    C. Mitchell, How to use a pairs trading strategy with etfs, 2015. [Online]. Available: https://et fdb.com/etf-trading-strategies/how-to-use-a-pairs-trading-strateg y-with-etfs/

    2015

  103. [117]

    L. N. Smith, Cyclical learning rates for training neural networks, 2015. DOI: 10.48550/ARXIV.1 506.01186. [Online]. Available: https://arxiv.org/abs/1506.01186. A Appendix A.1 TTM Pretraining Datasets TTM models were pretrained on a wide variety of publicly available time seri...

    2015 doi arXiv

  104. [723]

    Available: http://dx.doi.org/10.1109/TFUZZ.2021.3062723

    [Online]. Available: http://dx.doi.org/10.1109/TFUZZ.2021.3062723

    2021

  105. [2017]

    [Online]

    DOI: 10.48550/ARXIV.1705.09406. [Online]. Available: https://arxiv.org/a bs/1705.09406

    doi arXiv

  106. [2018]

    DOI: 10.1016/j.energy.2018.09.144

    2018 doi

  107. [2025]

    Available: https://fred.stlouisfed.org

    [Online]. Available: https://fred.stlouisfed.org

  108. [6598]

    [Online]

    DOI: 10.2307/2527343. [Online]. Available: http://dx.doi.org/10.2307/252 7343

    doi

  109. [7462]

    [Online]

    DOI: 10.1007/s10462-024-11044-2 . [Online]. Available: http://dx.doi.org /10.1007/s10462-024-11044-2

    doi

Pith tools

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