acceptodds
Under review as a conference paper at ICLR 2027

FiT: from market language to market decisions

35 views

Abstract

Financial time-series foundation models have recently enabled language-like modeling of financial markets through discrete representations of candle stick dynamics. However, existing models remain limited in faithfully representing market structure, anticipating future dynamics, and adapting these representations to downstream decision-oriented tasks. We propose **FiT**, **Fi**nancial **T**ransformer, which builds a unified pipeline from market representation to future prediction and downstream adaptation. On the data side, we introduce the FiT Dataset, expanding market coverage to more than 100 exchanges and temporal resolution down to 1-second K-lines, thereby enabling broader generalization across sampling frequencies. With more than 200 billion K-line records—17 times the size of the Kronos corpus—making it the largest financial K-line corpus to date, by the best of our knowledge. For pretraining, we propose Rotated Leech Spherical Quantization (RLSQ) for K-line tokenization designed to better capture the geometry of candlestick dynamics. Compared with the Kronos tokenizer, it substantially improves reconstruction performance, reducing reconstruction error by 62% while improving directional accuracy and K-line structural validity by 21% and 22%, respectively. For the predictor, we augment next-token prediction (NTP) with Future Summary Prediction (FSP), yielding 52% and 95% improvements in zero-shot IC and RankIC, respectively, over the best-performing baselines. On the post-training period, we perform BiRank GRPO to adapt FiT to downstream tasks, producing task-specific experts that are subsequently consolidated into the base model through multi-teacher on-policy distillation (MOPD). This procedure improves downstream performance while largely preserving the capabilities and generalization of the pre-trained model. We further develop a hybrid optimization strategy tailored to large-scale financial time-series pretraining, combining Muon, Lion, and AdamW across parameter groups, which reduces validation loss by 5.6% relative to an AdamW-only baseline. Together, these advances trace a complete path from market language to market dynamics to market decisions: RLSQ learns a high-fidelity discrete representation of K-line behavior, FSP learns to anticipate its future evolution, and BiRank GRPO with MOPD translates these capabilities into downstream tasks. FiT thus provides a unified foundation for financial K-line modeling that moves beyond representing the market toward anticipating its dynamics and adapting them to downstream decisions.

Primary area:
learning on time series and dynamical systems
Keywords:
financial time series, time series foundation models, K-line (candlestick) data, vector quantization, Leech lattice, future summary prediction, Muon optimizer, on-policy distillation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_RABeOpenReview ↗
Rating
1
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes FiT, a foundation model for financial K-line data. The pipeline has three main components: 1) RLSQ, a 24D Leech lattice codebook for tokenization using a Householder gradient surrogate; 2) FSP, an auxiliary future-summary prediction task during pretraining; and 3) BiRank GRPO combined with multi-teacher distillation (MOPD) for trading adaptation. The authors also introduce a large K-line dataset across 103 exchanges.

Critical strengths
  1. Using the first shell of the Leech lattice (RLSQ) instead of standard BSQ is well-motivated geometrically, as the guaranteed 60° separation helps prevent small price movements from collapsing into identical tokens.
  2. The theoretical analysis of the gradient surrogate is solid. Theorems 1 and 3 clearly explain how STE relates to RLSQ and acts as an implicit commitment term.
  3. This paper studies a meaningful direction since large financial market models are appreciated by both the AI technique community and financial products.
Critical weaknesses
  1. Unsubstantiated claims and missing comparisons.

    • The paper claims a 62% reconstruction gain and 21% directional gain over Kronos (lines 19–21, 125–127). However, I could not find any actual comparison with the Kronos tokenizer. Table 5 only compares RLSQ against STE on the same architecture.
    • The reported gains from FSP (52% zero-shot IC, 95% RankIC) are confounded. Table 3 only benchmarks against external models like Chronos-Bolt and Kronos. There is no ablation without FSP (λFSP=0\lambda_{\text{FSP}} = 0) under the same model and data budget, so it is unclear whether the improvement comes from FSP or just dataset size and architecture.
    • For MOPD, the text claims distillation preserves base capabilities, but there is no ablation comparing individual task experts to the distilled student model.
    • Section 4.3 (cross-market and cross-frequency generalization) appears completely unfinished. The text is full of blank placeholders like "Table – reports reconstruction error" and "hold out – exchanges entirely".
  2. Inconsistencies in main results

    • The claim of achieving SOTA across all tasks (line 128) is contradicted by Table 3: FiT trails TimeVAE on synthetic generation (TSTR IC: 0.0684 vs. 0.0690) and Passive Limit on queue-fill (2.69 bp vs. 2.75 bp).
    • Scaling inversion: In the CSI2000 trading task (Table 3), scaling from 258M to 680M leads to a sharp performance drop (annualized return falls from 44.9% to 25.8%, Sharpe drops from 1.21 to 0.72). Why does the 680M model degrade so significantly? This points to potential training instability in the RL stage that needs explanation.
  3. Missing controls promised in text

    • Appendix A.4 explicitly promises shuffled-reward and sharpening baselines to ensure RL gains are not just artifacts of distribution sharpening (Choshen et al., 2020). However, these controls are missing from the submission.
  4. Draft quality and inconsistencies

    • Unresolved citations and references: There are over 14 missing citations ("?") and 15 broken cross-references ("??"). The references for lattice bounds (lines 894, 902) are also missing, making Theorem 2 hard to verify.
    • Model name discrepancy: In line 1168, the text suddenly refers to the model as "Kairos" instead of FiT, suggesting unvetted text reuse.
    • Loss formulation mismatch: Appendix B.7 introduces a dynamic primal-dual curriculum, but Section 4.2 states static weights were used. It is unclear which setup was actually trained.
    • Placeholders remain in the reproducibility section (e.g., line 522 cuts off mid-sentence, line 525 has "[N] random seeds").
Questions for authors

Please refer to Weaknesses.

Reviewer_x6r2OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper propose FiT, which is a financial time-series foundation model that combines several techniques for representation learning, future prediction, and downstream adaptation. Specifically, FiT uses RLSQ for tokenization, FSP for pretraining, and BiRank GRPO with MOPD for downstream adaptation. The paper also introduce a FiT Dataset, which contains over 200 billion K-line records from more than 100 exchanges, with the temporal resolution down to 1 second K-line.

Critical strengths

None

Critical weaknesses

W1: Poor presentation and reproducibility. The paper is very difficult to follow due to unclear writing, insufficient explanations of technical details, and numerous formatting issues. The paper contains many missing or incorrect references, and even contains unfinished placeholders in several sections which includes missing experimental numbers and settings (e.g., “[N] random seeds” / “we release .” in reproducibility statement). More importantly, the paper lacks sufficient and clear enough details on dataset construction, training procedures, and hyperparameter settings, which makes the reported results difficult to reproduce. Although the paper does provide a link to an anonymous code repository to support reproducibility, and state their training details are provided in the repo, the repo does not contain any implementation code but only a single line readme. Overall, these issues make it very difficult to understand and verify the proposed method.

W2: Insufficient evidence and missing experiments. The main table only reports selected baseline results (the best / second best for each task), without providing the full list of evaluated baselines or explaining how they were selected. This makes it difficult to understand whether the comparisons are fair and comprehensive. Moreover, the paper does not provide the full performance results of the baselines or report standard deviations across runs (and the number of random seeds is also left as a placeholder). This makes it difficult to understand the reliability of the reported improvements and whether they are consistent across different runs. Moreover, some improvement numbers reported in the abstract and main text do not match the corresponding results (e.g., yielding 52% and 95% improvements over in zero-shot IC and RankIC is actually only 41.6% and 43.6%), and several claimed results cannot be found in the paper (e.g, Section 4.3 discuss the generalization experiment, but the corresponding tables and figures are missing).

W3: Unclear novelty and contribution. The proposed method seems to be a combination of lots of existing techniques/components from different fields, and introduces these components as its contributions. However, I found it difficult to identify the key research problem the authors are trying to solve, or why these specific techniques need to be combined together. Specifically, the paper lacks ablation studies to show the effectiveness of each component (although some comparisons are provided for RLSQ), making it hard to understand which components actually contribute to the final improvements and whether the combination is necessary if these ablations are missing. Overall, this work seems to be more focused on maximizing performance by combining different techniques, rather than addressing a specific research problem.

Questions for authors

Q1: Missing and Unclear Experiment Results.

  • Can the author provide the full generalization experiment results and its settings (e.g., unseen sampling frequencies, held-out exchanges, etc.) that are missing from section 4.3 ? These experiments are important for supporting the generalization claims of the proposed model.
  • Can the author explain how the improvements in zero-shot IC and RankIC mentioned in both abstract and main result are calculated? Based on Table 3, the largest FiT model achieves 0.0831 IC and 0.0916 RankIC, compared to 0.0587 and 0.0638 for the best baselines. The corresponding improvements are only around 41.6% and 43.6% rather than the reported 52% and 95%, respectively.

Q2: Comprehensiveness and significance of main results.

  • Can the author provide the full list of baselines of each task tested in Table 3 and how are these baselines selected? It is important to know whether the main comparisons are comprehensive enough to cover strong methods under each task as the proposed method is claimed as a foundation model.
  • Can the author provide the performance standard deviations across different seeds of each task in Table 3? Can the author provide a significant test against the best performing baseline on each task? Without both information, it is hard to determine whether the reported advantages are stable across different training seeds.

Q3: Contributions of each proposed component. Can the author provide ablation studies showing how much each proposed component actually contributes to the final performance? The listed experiments below are important as these components are all presented as part of the major contributions, and some of them are mentioned at the opening of Section 4 but actually missing.

  • How does the NTP+FSP compare with using NTP alone with the same FiT backbone? This is an important check to make sure the performance gain is from the added FSP objective rather than a having different model structure or using different pretraining dataset compared with Kronos.
  • How much does BiRank GRPO improve over native AR-style forecasting which is used by Kronos? This is an important sanity check to understand whether the performance comes from better pretraining or just a better downstream objective.
  • How well does MOPD preserve the performance of task-specific experts and does adding MOPD on top of the pretrained checkpoint actually improve performance? This is important to understand whether the distillation is just an average of different teachers and if this post-training step ruins the pretraining weight.
Reviewer_BmwnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FiT, a financial time-series foundation model integrating market representation, forecasting, and decision-making.

Main contributions: (1) A dataset of 208.5B K-line recrds. (2) RLSQ, a Leech-lattice-based tokenizer; (3) Future Summary Prediction (FSP) for improved forecasting. (4) BiRank GRPO and multi-teacher distillation for downstream adaptation. (5) hybrid optimization.

Validation: Experiments report improved reconstruction, forecasting, and trading performance across eight task families. However, several results are incomplete, and some claimed improvements do not match the reported numbers.

Critical strengths
  1. Well-motivated tokenizer design with strong theoretical and empirical support. The proposed RLSQ combine Leech-lattice quantization with rotation-based gradients. The theoretical analysis and controlled experiments against STE provide convincing evidence of improved reconstruction quality, making this the paper's strongest technical contribution.

  2. The dataset covers 208.5 billion K-line records across 103 exchanges, and the model is evaluated on eight downstream task families. Thus, large-scale dataset and broad downstream evaluation are involved. This scale and breadth demonstrate the potential practical significance of the proposed framework.

Critical weaknesses
  1. Insufficient ablation studies for key contributions. The paper lacks controlled experiments isolating the effects of FSP, BiRank GRPO, and MOPD. Without comparisons against NTP-only pretraining, models without RL, or individual task experts, it is difficult to establish whether these components deliver the claimed benefits.

  2. Several generalization experiments in Section 4.3 reference missing tables and figures, leaving important claims unverified. Additionally, the claimed 52% IC and 95% RankIC improvements do not match the results in Table 3 (approximately 42% and 44%, respectively).

Questions for authors

Could the authors provide a controlled comparison between NTP-only and NTP+FSP pretraining under identical settings? This is essential to determine how much of the reported forecasting improvement is attributable to FSP.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.