Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

FinCast: A Foundation Model for Financial Time-Series Forecasting

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read FinCast, a 1-billion-parameter financial time-series foundation model, reports average 20% lower zero-shot forecasting error than existing time-series foundation models, without task-specific fine-tuning.

desk verdict Credible 1B financial time-series foundation model with a real training run, but the headline 20% zero-shot gain rests on an undocumented data-disjointness assertion. read the letter →

arxiv 2508.19609 v1 pith:JEZEU6VK submitted 2025-08-27 cs.LG cs.AIq-fin.CP

Zhuohang Zhu , Haodong Chen , Qiang Qu , Vera Chung This is my paper
classification cs.LGcs.AIq-fin.CP
keywords financialtime-seriesforecastingfoundationmodelzero-shotgeneralizationmixtureofexpertsdecoder-onlytransformernon-stationaritypoint-quantilelossfrequencyembedding
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

FinCast is a 1-billion-parameter decoder-only transformer pretrained on more than 20 billion time points spanning stocks, cryptocurrencies, forex, futures, macroeconomic indicators, and other series. The paper's central claim is that this model generalizes across financial domains and sampling frequencies without task-specific fine-tuning: on a 3,632-series zero-shot benchmark it reduces mean-squared forecast error by about 20% relative to general-purpose time-series foundation models, and its zero-shot variant also beats supervised baselines on stock-price benchmarks. The design intended to make this work combines a point-quantile loss, a token-level sparse mixture-of-experts layer, and learnable frequency embeddings that encode temporal resolution. If the claim holds, financial forecasting would no longer require per-domain training, and a single pretrained model could be adapted with only minimal fine-tuning.

What carries the argument

The central object is the FinCast decoder-only transformer with three interacting mechanisms. A point-quantile loss (PQ-loss) jointly optimizes a Huber point forecast, quantile forecasts, first-difference trend consistency, and expert-balance regularization, which the paper says prevents forecast collapse and captures tail risk. A token-level sparse mixture-of-experts (MoE) routes each patch token to its top-2 of 4 expert MLPs, letting experts specialize across financial domains and resolutions. Learnable frequency embeddings encode each series' sampling frequency into a vector added to all input tokens, giving the model an explicit cue to resolution-dependent dynamics.

What would settle it

Run the zero-shot benchmark a second time with the pretraining corpus and benchmark series matched at subsequence level; if any benchmark series shares a time window or near-duplicate instrument with pretraining data, restrict to the fully disjoint subset and compare. If the 20% MSE gap over the best general-purpose model disappears, the zero-shot generalization claim is falsified. A second check: train FinCast with PQ-loss replaced by plain MSE on the same pretraining data; the ablation predicts at least a 7.6% degradation, so observing no degradation would refute the loss's causal role.

Watch

Extended reading notes

Core claim

The paper's central claim is that FinCast is the first foundation model built specifically for financial time-series forecasting. Trained on 2.4 million series and over 20 billion time points at resolutions from seconds to months, it reports the best results on both zero-shot and supervised benchmarks, with average MSE reductions of 20% and 23% respectively. The load-bearing mechanism is a joint training objective: a point-quantile loss that couples Huber point estimation with quantile-based uncertainty, a sparse token-level MoE that lets experts specialize by domain, and learnable frequency embeddings that condition representations on temporal resolution. A one-epoch fine-tune restricted to

Load-bearing premise

The 20% zero-shot improvement rests on the benchmark series being truly unseen during pretraining; the paper asserts this exclusion but does not document how the series were selected or deduplicated.

Editorial extensions

If this is right

  • Zero-shot forecasting becomes a practical default for financial series: one pretrained model covers stocks, crypto, forex, and futures without per-domain training.
  • The zero-shot variant itself beats supervised baselines on the studied stock benchmarks, implying that pretraining on diverse financial data transfers better than training directly on the target dataset.
  • Minimal fine-tuning suffices: updating only the output block and the last 10% of MoE layers for one epoch yields the best supervised results, so adaptation cost is small.
  • Inference on an 8GB consumer GPU is reported feasible at up to 5x faster than compared models along the accuracy-latency frontier, making deployment in constrained settings plausible.
  • The ablation attributes 9.32% MSE degradation to removing sparse routing, 7.62% to removing PQ-loss, and 4.38% to removing frequency embeddings, so each component contributes to the zero-shot result.

Reading between the lines

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

  • The paper does not test whether the pretrained weights also transfer to downstream financial tasks beyond forecasting, such as regime detection or risk measurement, but the quantile head already produces distributional forecasts that could support those uses.
  • A direct extension would replace the learnable frequency embedding with a fixed sinusoidal frequency encoding; the ablation removes the component entirely, so it does not isolate the value of learning the embedding.
  • If the claimed benchmark exclusion is confirmed, the scaling-law motivation cited in the paper predicts that further enlarging the pretraining corpus should yield continued zero-shot gains.
  • The benchmark's composition is not specified beyond domain and resolution, so an external check would construct a held-out set with explicit provider and date disjointness and rerun the comparison.
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. FinCast is a 1B-parameter decoder-only Transformer with a token-level sparse MoE, learnable frequency embeddings, and a point-quantile loss, pretrained on roughly 20B time points from financial and non-financial series. The paper claims that it outperforms general-purpose time-series foundation models in zero-shot settings and supervised baselines on stock-data benchmarks, with average MSE reductions of 20% and 23%. The architecture is described in reasonable detail and the tables are arithmetically internally consistent. The central concerns are that the zero-shot benchmark's exclusion from pretraining is asserted but never documented, and that the headline percentage reductions are computed against the average of all baselines rather than the strongest baseline, with no uncertainty estimates.

Significance. If the zero-shot results held with a genuinely disjoint benchmark, FinCast would be a useful contribution: it is among the first large financial-specific forecasting models, and the design choices (PQ-loss, frequency embeddings, sparse MoE) are sensible and partially ablated. The stated intention to release code and weights is a further strength. However, as presented, the central generalization claim is not yet established: the evaluation protocol cannot rule out overlap between pretraining and benchmark, and the reported margins depend on averaging over weak baselines. The paper is therefore more a promising system description than a fully supported foundation-model evaluation.

major comments (4)
  1. [Section 4.1] The load-bearing premise for the 20% zero-shot reduction is that the 3,632 benchmark series are disjoint from the pretraining corpus. The paper only asserts this. It does not report the benchmark's constituent symbols/providers, date ranges, selection rules, or any deduplication procedure against the 2.4M-series corpus in Table 1. Since both corpora draw from the same public-API ecosystems over the same domain/resolution grid, an overlap would make the zero-shot numbers in-distribution. The statement that benchmark scale makes overfitting unlikely does not address leakage. Please provide a machine-checkable split and an overlap audit (e.g., ticker + timestamp matching) between the benchmark and pretraining sets.
  2. [Abstract and Section 1; Tables 2-3] The advertised '20%', '23%', and '26%' reductions are computed against the arithmetic mean of all baselines, not against the strongest baseline. Against the best zero-shot baseline (Chronos-Small), Table 2's average MSE reduction is 11.6%; against PCIE, the best supervised baseline, FinCast zero-shot reduces average MSE by 5.2% and FinCast-finetune by 8.9%. Moreover, Table 3 shows FinCast zero-shot is worse than PCIE on US_14L at horizon 10 (0.1509 vs 0.1458), which contradicts 'consistently outperforms ... without task-specific fine-tuning.' Report per-baseline and best-baseline margins.
  3. [Tables 2 and 3] All results are single-run point estimates on highly noisy financial data, with no seeds, error bars, confidence intervals, or significance tests. Several decisive margins are small (e.g., US_71 h=10: FinCast zero-shot 0.0675 vs PCIE 0.0690; Table 3). Without repeated runs or at least a paired test, the claimed uniform superiority is not statistically supported.
  4. [Section 4, Tables 2-3] No persistence/random-walk baseline is included. A last-value or drift forecast is the standard lower bound for financial series; the 'state-of-the-art' comparison is incomplete without it. Also, the supervised benchmark is adopted from the authors' own arXiv paper [46] without specifying whether baseline numbers are copied or re-run; please clarify.
minor comments (6)
  1. [Section 3.6, Eq. (20)] The sentence 'where each λ controls' is inconsistent with the equation, which writes Ltotal = Lpoint + Ltrend + Lquantile + LMOE without weights; Eq. (22) also omits λ. Please fix notation.
  2. [Section 3.3] The mapping from temporal resolution to frequency index f is never specified. Add a table (e.g., which integer corresponds to 1-minute, 1-hour, daily, weekly) for reproducibility.
  3. [Section 4.1] The phrase 'The benchmark's scale and diversity make overfitting unlikely' is logically invalid as a leakage defense; replace it with the quantitative overlap control requested above.
  4. [Section 4.5] 'Some collapse to flat-line outputs due to only using MSE for optimization' overgeneralizes; for example, Chronos is trained with tokenized cross-entropy, not MSE. Please qualify this claim to the models for which it is accurate.
  5. [Figure 6] The inference-speed panel lacks axis labels/units and numeric timing details; please report actual latency and hardware configuration in the caption or text.
  6. [References] [39] is cited as a source of standard practice for input length 128, but [39] is a stock-movement prediction paper and does not appear to make that recommendation; please re-check the citation.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation chain; minor self-citation in supervised benchmark is not load-bearing.

full rationale

We walked the claimed derivation chain and found no step in which a prediction or first-principles result reduces by construction to its inputs. The zero-shot claim (Abstract; §4.1) is an empirical comparison on a benchmark that the paper asserts is excluded from pretraining ('The benchmark dataset is excluded from the pretraining datasets to ensure a strict zero-shot setting'). This exclusion is not demonstrated with series lists, providers, or deduplication evidence, and the benchmark is drawn from the same public data ecosystems and same domain/resolution grid as the pretraining corpus. That is a data-provenance/validity risk, not a circularity: the paper's equations do not define the benchmark as a function of the fitted model, and the reported MSE/MAE are computed from model outputs, not from fitted parameters renamed as predictions. The supervised benchmark (§4.2) adopts the standardized benchmark and baseline numbers from the authors' own prior paper [46], which is a minor self-citation. However, this reuse does not make the central claim circular—the comparison is logically distinct from the model's construction, and the zero-shot contribution (the paper's strongest claim) does not depend on [46]. Architectural components (PQ-loss, MoE, frequency embeddings) are validated through ablations (§4.3), not defined in terms of the evaluation metrics. No other circularity patterns (self-definition, uniqueness imported from authors, ansatz smuggled via citation, renaming) are present. Therefore, the appropriate finding is a minor self-citation that is not load-bearing, warranting a score of 2 rather than 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper introduces no ontological entities (no new particles, forces, dimensions, or conserved quantities); the MoE experts and frequency embeddings are architectural components, not invented entities. The empirical claims rest on the six unreported free parameters listed above and on four domain assumptions, the most load-bearing being the asserted disjointness between the pretraining corpus and the zero-shot benchmark.

free parameters (6)
  • Loss weights lambda_trend, lambda_quantile, lambda_MOE
    Eq. 20 sums four loss terms 'where each lambda controls the relative contribution,' but no values are reported anywhere; Lpoint in Eq. 22 has no lambda at all. These weights shape training and every reported result.
  • Frequency index mapping f
    Section 3.3: 'Each input sequence is assigned a discrete frequency index f,' but the mapping from sampling rates (seconds to months) to integers is never specified; the learned frequency embeddings depend entirely on this hand-assigned index.
  • Patch size P
    Section 3.3 defines N = floor(L/P) non-overlapping patches; P is never disclosed, despite the masking ratio (15%) and context lengths (1024/256) depending on it.
  • Quantile set Q
    Eq. 21 says 'Q is a set of quantiles (e.g. deciles),' but the actual set is not given, and the output head producing y-hat_q is not described: Section 3.5 (Eq. 18) maps each token to a single point forecast in R^H.
  • Huber delta (delta)
    Eq. 22 uses delta in the Huber point loss without a stated value.
  • Architecture dimensions (D_model, layers, heads)
    The model is described only as '1 billion parameters, 4 experts, top-k=2'; D_model, layer count, and head count are never reported.
assumptions (4)
  • domain assumption The zero-shot benchmark is disjoint from the pretraining corpus such that FinCast has never seen those series or time periods.
    Section 4.1 states 'The benchmark dataset is excluded from the pretraining datasets to ensure a strict zero-shot setting' without describing how the benchmark was built or any deduplication mechanism. If this fails, the 20% zero-shot claim collapses to in-distribution performance.
  • domain assumption Channel independence (Eq. 4): ignoring cross-series correlation does not hurt the forecasting targets.
    Section 3.1 applies the same scalar mapping to each coordinate series, following patch-based prior work [28]; financial co-movement across instruments is thereby discarded, a substantive modeling choice for stock/crypto/forex data.
  • domain assumption Instance normalization per patch preserves the information needed to forecast H steps ahead.
    Section 3.3 normalizes each patch by its own mu and sigma and inverts at the output (Eq. 19); this assumes scale within the forecast window stays consistent with the input patch.
  • domain assumption Public-API price data (crypto, forex, futures, stocks at 1min-1wk frequencies) is a sufficient and unbiased proxy for financial time-series forecasting as a general capability.
    Section 3.7.1 builds the entire 20B-point corpus from public interfaces; all generality claims about financial forecasting inherit this sampling.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FinCast: A Foundation Model for Financial Time-Series Forecasting." pith.science (2026). https://pith.science/paper/JEZEU6VK

@misc{pith2026250819609,
  author       = {Pith},
  title        = {Pith review of: FinCast: A Foundation Model for Financial Time-Series Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JEZEU6VK}},
  note         = {Machine review of arXiv:2508.19609}
}
read the original abstract

Financial time-series forecasting is critical for maintaining economic stability, guiding informed policymaking, and promoting sustainable investment practices. However, it remains challenging due to various underlying pattern shifts. These shifts arise primarily from three sources: temporal non-stationarity (distribution changes over time), multi-domain diversity (distinct patterns across financial domains such as stocks, commodities, and futures), and varying temporal resolutions (patterns differing across per-second, hourly, daily, or weekly indicators). While recent deep learning methods attempt to address these complexities, they frequently suffer from overfitting and typically require extensive domain-specific fine-tuning. To overcome these limitations, we introduce FinCast, the first foundation model specifically designed for financial time-series forecasting, trained on large-scale financial datasets. Remarkably, FinCast exhibits robust zero-shot performance, effectively capturing diverse patterns without domain-specific fine-tuning. Comprehensive empirical and qualitative evaluations demonstrate that FinCast surpasses existing state-of-the-art methods, highlighting its strong generalization capabilities.

Figures

Figures reproduced from arXiv: 2508.19609 by the authors.

Figure 1
Figure 1. FinCast Model Architecture. ○A: Input preprocessing, tokenization, and applies Learnable Frequency Embedding. ○B : Causal Attention, masking future attention scores. ○C : Sparse MOE, activation based on each token. ○D : Process decoder outputs and reverse norm. ○E : PQ-Loss jointly optimizes output head. to adaptively scale each feature dimension within the attention computation [8]. To enforce autoregressive behavi… view at source ↗
Figure 2
Figure 2. Expert activation patterns across datasets. Each [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Point and Quantile Outputs During Training [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Zero shot forecasting examples from Zero Shot Forecast Benchmark, [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Supervised forecasting examples from Supervised Forecast Benchmark, [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Inference Speed vs Performance In contrast, FinCast demonstrates strong pattern sensitivity and trend awareness, accurately adapting to complex pattern shifts and diverse domains with different temporal resolutions [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  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.

Reference graph

Works this paper leans on

48 extracted references · 18 canonical work pages · cited by 1 Pith paper

  1. [46]

    Zhuohang Zhu, Haodong Chen, Qiang Qu, Xiaoming Chen, and Vera Chung

    work page

  2. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 1 (2023), 35

    arXiv 2023

  3. [2]

    Maddix, Hao Wang, Michael W

    Abdul Fatir Ansari, Lorenzo Stella, Caner Turkmen, Xiyuan Zhang, Pedro Mercado, Huibin Shen, Oleksandr Shchur, Syama Sundar Rangapuram, Sebas- tian Pineda Arango, Shubham Kapoor, Jasper Zschiegner, Danielle C. Maddix, Hao Wang, Michael W. Mahoney, Kari Torkkola, Andrew Gordon Wilson, Michael Bohlke-Schneider, and Yuyang Wang. 2024. Chronos: Learning the L...

    arXiv 2024

  4. [3]

    Ariyo, Adewumi O

    Adebiyi A. Ariyo, Adewumi O. Adewumi, and Charles K. Ayo. 2014. Stock Price Prediction Using the ARIMA Model. In 2014 UKSim-AMSS 16th International Conference on Computer Modelling and Simulation . IEEE, IEEE, UK, 106–112

    work page 2014

  5. [5]

    Christopher Chatfield. 2013. The analysis of time series: theory and practice . Springer, USA

    work page 2013

  6. [7]

    John H Cochrane. 1997. Time series for macroeconomics and finance

    work page 1997

  7. [8]

    Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou. 2024. A decoder- only foundation model for time-series forecasting. arXiv:2310.10688 [cs.CL] https://arxiv.org/abs/2310.10688

    arXiv 2024

  8. [9]

    Kaize Ding, Jianling Wang, Jundong Li, Kai Shu, Chenghao Liu, and Huan Liu

Show all 48 references
  1. [10]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. An Image is Worth 16x16 Words: Transformers for Image Recognit...

    2021 arXiv

  2. [11]

    Kelvin Du, Rui Mao, Frank Xing, and Erik Cambria. 2024. Explainable Stock Price Movement Prediction using Contrastive Learning. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management (Boise, ID, USA) (CIKM ’24). Association for Computing...

    2024

  3. [12]

    Yupeng Fang, Ruirui Liu, Huichou Huang, Peilin Zhao, and Qingyao Wu. 2024. A Spatio-Temporal Diffusion Model for Missing and Real-Time Financial Data Infer- ence. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management (Boise, ID, USA) (...

    2024

  4. [13]

    1998.Time series models for business and economic forecasting

    Philip Hans Franses. 1998.Time series models for business and economic forecasting. Cambridge university press, UK

    1998

  5. [14]

    Philip Hans Franses and Dick Van Dijk. 1996. Forecasting stock market volatility using (non-linear) Garch models. Journal of forecasting 15, 3 (1996), 229–235

    1996

  6. [15]

    Felix A Gers, Jürgen Schmidhuber, and Fred Cummins. 2000. Learning to forget: Continual prediction with LSTM. Neural computation 12, 10 (2000), 2451–2471

    2000

  7. [16]

    Webb, Rob J

    Rakshitha Godahewa, Christoph Bergmeir, Geoffrey I. Webb, Rob J. Hyndman, and Pablo Montero-Manso. 2021. Monash Time Series Forecasting Archive. arXiv:2105.06643 [cs.LG] https://arxiv.org/abs/2105.06643

    2021 arXiv

  8. [17]

    Kaan Gokcesu and Hakan Gokcesu. 2021. Generalized Huber Loss for Robust Learning and its Efficient Minimization for a Robust Statistics. arXiv:2108.12627 [stat.ML] https://arxiv.org/abs/2108.12627

    2021 arXiv

  9. [18]

    Rae, Oriol Vinyals, and Laurent Sifre

    Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Jo- hannes Welbl, Aidan Clark, Tom Hennigan, Eric Noland, Katie Millican, George van den Driessche, Bogdan Damoc, Aurelia Guy, Simon Os...

    2022 arXiv

  10. [19]

    Min Hou, Chang Xu, Yang Liu, Weiqing Liu, Jiang Bian, Le Wu, Zhi Li, Enhong Chen, and Tie-Yan Liu. 2021. Stock Trend Prediction with Multi-granularity Data: A Contrastive Learning Approach with Adaptive Fusion. In Proceedings of the 30th ACM International Conference on Informa...

    2021

  11. [20]

    Yulong Jia, Guanxing Li, Ganlong Zhao, Xiangru Lin, and Guanbin Li. 2024. Graph- VAE: Unveiling Dynamic Stock Relationships with Variational Autoencoder- based Factor Modeling. In Proceedings of the 33rd ACM International Con- ference on Information and Knowledge Management (B...

    2024

  12. [21]

    Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. 2020. Scaling Laws for Neural Language Models. arXiv:2001.08361 [cs.LG] https: //arxiv.org/abs/2001.08361

    2020 arXiv

  13. [22]

    Koa, Yunshan Ma, Ritchie Ng, and Tat-Seng Chua

    Kelvin J.L. Koa, Yunshan Ma, Ritchie Ng, and Tat-Seng Chua. 2023. Diffusion Variational Autoencoder for Tackling Stochasticity in Multi-Step Regression Stock Price Prediction. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management (CIKM...

    2023

  14. [23]

    Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. 2020. GShard: Scaling Giant Models with Conditional Computation and Automatic Sharding. arXiv:2006.16668 [cs.CL] https://arxiv.org/abs/2006.16668

    2020 arXiv

  15. [24]

    Yuchen Liu, Shimin Di, Lei Chen, Xiaofang Zhou, and Fei Lin. 2024. A Universal and Interpretable Method for Enhancing Stock Price Prediction. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management (Boise, ID, USA) (CIKM ’24). Associatio...

    2024

  16. [25]

    Gael M Martin, David T Frazier, Worapree Maneesoonthorn, Rubén Loaiza- Maya, Florian Huber, Gary Koop, John Maheu, Didier Nibbering, and Anastasios Panagiotelis. 2024. Bayesian forecasting in economics and finance: A modern review. International Journal of Forecasting 40, 2 (2...

    2024

  17. [26]

    Larry R Medsker, Lakhmi Jain, et al. 2001. Recurrent neural networks. Design and Applications 5, 64-67 (2001), 2

    2001

  18. [27]

    Tung Nguyen, Jason Jewik, Hritik Bansal, Prakhar Sharma, and Aditya Grover

  19. [28]

    Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2022. A time series is worth 64 words: Long-term forecasting with transformers. arXiv preprint arXiv:2211.14730 1 (2022), 10

    2022 arXiv

  20. [29]

    Hui Niu, Siyuan Li, and Jian Li. 2022. MetaTrader: An Reinforcement Learning Approach Integrating Diverse Policies for Portfolio Optimization. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management (CIKM ’22). ACM, Atlanta, USA, 1573–1583...

    2022

  21. [30]

    Omer Berat Sezer, Mehmet Ugur Gudelek, and Ahmet Murat Ozbayoglu. 2020. Financial time series forecasting with deep learning: A systematic literature review: 2005–2019. Applied soft computing 90 (2020), 106181

    2020

  22. [31]

    Xiaoming Shi, Shiyu Wang, Yuqi Nie, Dianqi Li, Zhou Ye, Qingsong Wen, and Ming Jin. 2024. Time-MoE: Billion-Scale Time Series Foundation Models with Mixture of Experts. arXiv:2409.16040 https://arxiv.org/abs/2409.16040

    2024 arXiv

  23. [32]

    Stephen J Taylor. 2011. Asset price dynamics, volatility, and prediction . Princeton university press, USA

    2011

  24. [33]

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. 2023. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 1 (2023), 5

    2023 arXiv

  25. [34]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 1 (2023), 1

    2023 arXiv

  26. [35]

    Ruey S Tsay. 2005. Analysis of financial time series . John wiley & sons, USA

    2005

  27. [36]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems (Long Beach, California, ...

    2017

  28. [37]

    Jingjing Wang, Yanhao Wang, Wenjun Jiang, Yuchen Li, and Kian-Lee Tan. 2020. Efficient Sampling Algorithms for Approximate Temporal Motif Counting (Ex- tended Version). arXiv:2007.14028 [cs.SI] https://arxiv.org/abs/2007.14028

    2020 arXiv

  29. [38]

    Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. 2022. Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Fore- casting. arXiv:2106.13008 [cs.LG] https://arxiv.org/abs/2106.13008 FinCast: A Foundation Model for Financial Time-Series Forecas...

    2022 arXiv

  30. [39]

    Yumo Xu and Shay B. Cohen. 2018. Stock Movement Prediction from Tweets and Historical Prices. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , Iryna Gurevych and Yusuke Miyao (Eds.). Association for Computatio...

    2018 doi

  31. [40]

    Xiaoyu You, Mi Zhang, Daizong Ding, Fuli Feng, and Yuanmin Huang. 2021. Learning to Learn the Future: Modeling Concept Drifts in Time Series Pre- diction. In Proceedings of the 30th ACM International Conference on Informa- tion & Knowledge Management (Virtual Event, Queensland...

    2021

  32. [41]

    Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. 2023. Are transformers effective for time series forecasting?. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and...

    2023 doi

  33. [42]

    Biao Zhang and Rico Sennrich. 2019. Root Mean Square Layer Normalization. arXiv:1910.07467 [cs.LG] https://arxiv.org/abs/1910.07467

    2019 arXiv

  34. [43]

    Liheng Zhang, Charu Aggarwal, and Guo-Jun Qi. 2017. Stock Price Prediction via Discovering Multi-Frequency Trading Patterns. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (Halifax, NS, Canada) (KDD ’17). Association for C...

    2017

  35. [44]

    Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. 2021. Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting. arXiv:2012.07436 [cs.LG] https://arxiv.org/ abs/2012.07436

    2021 arXiv

  36. [45]

    Peng Zhu, Yuante Li, Yifan Hu, Qinyuan Liu, Dawei Cheng, and Yuqi Liang. 2024. LSR-IGRU: Stock Trend Prediction Based on Long Short-Term Relationships and Improved GRU. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management (Boise, ID, ...

    2024 doi

  37. [47]

    Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, and William Fedus. 2022. ST-MoE: Designing Stable and Trans- ferable Sparse Expert Models. arXiv:2202.08906 [cs.CL] https://arxiv.org/abs/ 2202.08906

    2022 arXiv

  38. [2020]

    In Proceedings of the 29th ACM International Conference on Information & Knowledge Management (Virtual Event, Ireland) (CIKM ’20)

    Graph Prototypical Networks for Few-shot Learning on Attributed Net- works. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management (Virtual Event, Ireland) (CIKM ’20). Association for Com- puting Machinery, New York, NY, USA, 295–304. doi...

  39. [2023]

    arXiv:2307.01909 [cs.LG] https://arxiv.org/abs/2307.01909

    ClimateLearn: Benchmarking Machine Learning for Weather and Climate Modeling. arXiv:2307.01909 [cs.LG] https://arxiv.org/abs/2307.01909

    arXiv

  40. [2025]

    arXiv:2504.17313 [cs.CE] https://arxiv.org/abs/2504.17313

    Tokenizing Stock Prices for Enhanced Multi-Step Forecast and Prediction. arXiv:2504.17313 [cs.CE] https://arxiv.org/abs/2504.17313

    arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.