acceptodds
Under review as a conference paper at ICLR 2027

KiT: A Foundation Model for Financial Time-Series Forecasting using Diffusion Transformers

26 views

Abstract

Financial candlestick forecasting is fundamental to quantitative investment, yet it remains exceptionally challenging due to extremely low signal-to-noise ratios and vast heterogeneity across markets and instruments. Existing approaches have largely attempted to introduce deep learning to capture hidden temporal features, but most adopt an auto-regressive formulation, which leads to error accumulation during inference. Meanwhile, general-purpose time-series foundation models are not tailored to the unique structure of k-line data and yield unsatisfactory performance on downstream candlestick forecasting tasks. To tackle these problems, we introduce KiT, a K-line Diffusion Transformer foundation model, and reformulate future prediction as conditional path generation via flow matching: given a historical context window, the model generates an ensemble of plausible future OHLCV trajectories. We pre-train KiT at multiple parameter scales on billions of candlestick bars spanning multiple markets and timescales. Across three markets and seven resolutions, KiT attains a mean return RankIC of and a mean volatility RankIC of , leading at every timescale and outperforming both task-specific financial forecasters and general time-series foundation models.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-Series Forecasting, Financial Forecasting, Diffusion Models, Diffusion Transformers, Generative Models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_f5cKOpenReview ↗
Rating
3
Confidence
4
Informativeness
2(−1)
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a foundation model KiT based on diffusion Transformers specifically for financial candlestick forecasting. The authors reformulate future candlestick prediction as conditional path generation via a non-autoregressive flow in continuous space, while preserving candle geometry and handling multiscale heterogeneity. KiT achieves better scores than both the general and task-specific baselines across multiple markets and frequencies.

Critical strengths
  1. This paper targets the important task of financial candlestick forecasting and identifies the research gap correctly. The proposed approach is well-motivated.

  2. KiT implements a non-autoregressive inference that avoids error accumulation encountered in many autoregressive and point-forecasting alternatives, while handling certain levels of heterogeneity across multiscale and multimarket that challenges existing diffusion-based models.

  3. Results are encouraging across multiple scenarios.

Critical weaknesses
  1. Although KiT outperforms a wide variety of baselines, the improvements compared to the best alternative Kronos-base-FT in some configurations is small. Significance tests and confidence intervals are not provided to demonstrate the significance.

  2. The paper claims that avoiding error accumulation is a major advantage, but there is no empirical validation to verify the superiority of this design over the autoregressive approaches under a fair comparison.

  3. The testing period is fixed and relatively short, making the results less convincing. Financial markets change with complex temporal regime shifts. Model forecasts may have totally different in distinct periods.

Questions for authors
  1. Does KiT handle the heterogeneity in cross-sectional dependence and regime shift?

  2. Regarding the above weakness 2, can you show the model performance of the KiT variant using autoregressive architecture while keeping the other designs the same for a fair comparison?

  3. Does KiT’s performance remain the best when evaluated on a different testing period?

Minor issue: Citation formatting reads a little confusing throughout this paper.

4 comments
AuthorsOpenReview ↗

Comment for Reviewer f5cK (part1)

Thank you for the positive assessment, and for noting that the paper "identifies the research gap correctly" and that the approach is "well-motivated." Each of your requests led to new experiments, which we will add to the revision. This part answers Weak1 and Weak2 / Q2; part 2 answers Weak3 / Q3, Q1 and the citation issue.

Weak1: Significance and confidence intervals

We will report every result with circular block-bootstrap 95% CIs over anchors, Newey–West t-statistics and paired tests on identical cases.

  • Against Kronos fine-tuned under KiT's protocol (A-share daily, 300 matched names). In the submission, Kronos-base-FT used the checkpoint with the best test score, which favors the baseline. We additionally fine-tuned Kronos-base on A-share daily bars through 2025, with checkpoint selection by validation loss and KiT's context and horizon. Volatility RankIC: 0.677 vs 0.606 (t 5.1) on 13 Apr – 30 Oct 2026, and +0.047 (t 5.0) on the 1 Jan – 10 Apr 2026 test window. 90% coverage: 0.64 vs 0.39 (t 6.5) and 0.65 vs 0.45 (t 4.2).
  • Against zero-shot Kronos with its native OHLCV input (identical names and anchors): volatility RankIC +0.187 (A-share, t 5.8) and +0.087 (U.S., t 9.7) on every anchor of 13 Apr – 30 Oct 2026; on the Jan–Apr test window (300 matched names, 13 anchors), A-share return RankIC 0.090 vs 0.010 (t 2.4).
  • Against the five general TSFMs of Tables 1–2 (Chronos-2, Chronos-Bolt, Sundial, TimesFM 2.5, TimesFM 3), re-run on full cross-sections of 1,000 A-share and 824 U.S. names: KiT's volatility RankIC is higher on 100% of anchors in all 20 comparisons (2 markets × 2 windows × 5 models; t 5.9–26.9). On A-share daily returns in the Jan–Apr test window (13 anchors), KiT leads four of the five significantly (+0.063 to +0.096, t 2.5–6.5).

Weak2 / Q2: The autoregressive variant, all else equal

  • Trained twins. We trained two KiT-S models (31M) from scratch with identical backbone, representation, data (A-share daily bars), recipe and number of updates. The only difference is the factorization: one generates the 20-bar horizon jointly, the other autoregressively, bar by bar, conditioned on its own samples. On 37 held-out anchors × 300 names (Apr 2025 – Oct 2026), cumulative-return RankIC by horizon is:
Horizon (bars)15101520
Joint (KiT).053.054.044.066.071
Autoregressive.021.031.037.035.032

The joint model ranks returns better at 19 of the 20 horizons. The AR twin's skill plateaus around 0.03–0.04, while the joint model reaches 0.071 at h = 20. Conditioning on its own samples also damps the AR twin's paths: the ratio of predicted to realized mean absolute bar return at h = 20 is 0.60, vs 0.67 for the joint model.

  • Same checkpoint, rolled out. Holding the paper's checkpoint fixed and changing only the generation mode (4 blocks of 5 bars, or 10 blocks of 2, each conditioned on the previous samples) worsens CRPS (0.738 → 0.756, t 2.2) and 90% coverage (0.651 → 0.633 / 0.626, t −2.5 / −2.9) on A-share daily bars, 13 Apr – 30 Oct 2026. It also damps late-horizon amplitude (0.864 → 0.818 / 0.820). fp32 reproduces the bf16 results, so this is not a numerical effect.
  • Cost. Rollout needs one sampler call per block: 5.3× / 11.3× / 19.7× the wall-clock time of one-shot generation for 4×5 / 10×2 / 20×1 blocks.
AuthorsOpenReview ↗

Comment for Reviewer f5cK (part2)

Weak3 / Q3: Does KiT remain the best on a different testing period?

  • Seven unseen months (13 Apr – 30 Oct 2026), paper checkpoint frozen (data that did not exist at training time; 1,000 A-share and 824 U.S. names). Volatility RankIC is 0.677 [0.646, 0.714] (A-share) and 0.861 [0.849, 0.872] (U.S.). On every anchor it is above MLE-fitted GARCH (+0.082, t 9.6 / +0.045, t 8.9), EWMA (+0.051 / +0.057), zero-shot Kronos and the five TSFMs. It also beats OHLC range estimators and HAR built on the same input on 86–100% of anchors.
  • Nineteen held-out months (Apr 2025 – Oct 2026). For a longer held-out window, we also evaluate an A-share KiT-L of the same architecture, recipe and size (training ≤ 2023-12-31, validation to 2025-03-31) on 73 anchors × ~1,000 names. On identical cases:
KiTChronos-2Chronos-BoltSundialTimesFM 2.5TimesFM 3
Return RankIC.085.048.017.073.069.068
Volatility RankIC.603.509.510.466.524.497
Q5–Q1 spread / 20 days+2.1%+0.9%+0.0%+0.7%+0.7%+1.0%

KiT's return RankIC is positive in each of the six quarters (0.055 / 0.087 / 0.106 / 0.141 / 0.069 / 0.058; overall t 5.2). Its volatility RankIC is higher on 99–100% of anchors (t 11.2–13.7), and its market-neutral spread is the only one significant at the 5% level (t 3.3; best TSFM t 1.9). Against zero-shot Kronos on 300 of these names, return RankIC is 0.082 vs 0.029 and volatility RankIC 0.611 vs 0.483 (t 10.2).

Q1: Heterogeneity, cross-sectional dependence and regime shift

  • Heterogeneity across instruments and markets. One set of parameters serves 4,830 instruments and seven resolutions through the scale-free candle state, with market, sector, instrument and resolution identity injected through AdaLN and calendar / event signals per bar. Two tests show that this transfers. (i) Three-market pretraining beats an A-share-only KiT-L of the same size on unseen A-share data: volatility RankIC 0.677 vs 0.649 (t 3.5), CRPS 0.750 vs 0.795 (t −2.2), 90% coverage 0.64 vs 0.57 (t 4.3). (ii) Zero-shot, on 161 actively traded perpetual futures outside its vocabulary (null instrument identity, 15m bars, Apr–Oct 2026), KiT's volatility RankIC is 0.709, above per-instrument MLE-fitted GARCH (0.660, t 4.8) and historical volatility (0.650, t 6.5); its return RankIC is 0.050 (t 3.6).
  • Cross-sectional dependence. KiT generates each instrument's path conditionally on its own history and identity, so cross-asset dependence is captured only implicitly, through shared market-level and calendar conditioning. We will state this as a limitation, and leave explicit cross-asset attention to future work.
  • Regime shift. Over all 2026 anchors, KiT's volatility advantage over fitted GARCH and EWMA holds on 100% of anchors in every calendar month, in every tercile of market volatility and of cross-sectional dispersion, and in rising and falling markets, in both markets (mean gain over GARCH +0.099, t 9.7, A-share; +0.045, t 10.9, U.S.). Return skill shows no decay with time since the training cutoff (slope t −0.3).

Minor: Citations We will unify the citation style (textual vs. parenthetical) and fix misplaced citations throughout.

We hope these experiments settle Weak1–Weak3. If so, we would be grateful if you would consider raising your score. Thank you again for recognizing the value of the diffusion based models!

Thanks for your response. I think this is a good opportunity to upload the revised manuscript to formally address the reviewers' concerns and strengthen the paper during the rebuttal period. I would like to review the revisions, if available.

Thank you for the suggestion. We are finalizing the revised manuscript and will upload it. For you reference, in summary, the revision will include:

  1. Significance: block-bootstrap 95% CIs, Newey–West t-statistics and paired tests for all main results, together with a Kronos baseline fine-tuned under KiT’s protocol with validation-based checkpoint selection (Weak1).

  2. Joint vs. autoregressive generation: an autoregressive twin trained from scratch with everything else identical. The joint model ranks returns better at 19 of 20 horizons and is better calibrated (Weak2, Q2).

  3. Different testing periods: evaluation on seven unseen months (Apr–Oct 2026) with the frozen checkpoint, and on 19 held-out months with a model trained through 2023 (Weak3, Q3).

  4. Heterogeneity and regime shift: the cross-market transfer and zero-shot futures results, and the per-month and per-regime analysis. Cross-sectional dependence will be stated as a limitation. (Q1)

Thank you again for recognizing the contributions of our work!

Reviewer_MkMsOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces KiT, a diffusion-transformer foundation model for candlestick forecasting. It represents OHLCV bars using five log-ratio features, uses rectified flow to generate a future path non-autoregressively, and conditions on instrument/market identity and calendar signals. Three model sizes are pretrained on approximately three billion bars across US equities, Chinese A-shares, and cryptocurrencies at seven resolutions. On the stated test window (1 January–10 April 2026), KiT reports mean return RankIC of 0.0571 and mean volatility RankIC of 0.6600. The paper also reports an A-share long-only backtest for May 2025–April 2026, with annualized net returns of 65.2% on 5-minute bars and 64.4% on 2-hour bars. Comprehensive evaluations across multiple markets and time scales at seven distinct resolutions demonstrate that KiT significantly outperforms prior methods in both return and volatility forecasting. Furthermore,theoretical backtests illustrate the strong economic relevance of its generated path forecasts.

Critical strengths
  1. The domain-specific representation is well-designed. The five-dimensional log-ratio encoding captures relative price movements, candle anatomy, and normalized volume in a common representation across instruments and time scales. Because the mapping is invertible given the previous close and the shadow coordinates are constrained to be nonnegative, it both retains the encoded OHLCV information and supports valid candlestick reconstruction.

  2. A suitable generative formulation for multi-step forecasts, with flow matching generates the future horizon jointly rather than one bar at a time. This avoids autoregressive rollout by construction and lets the model represent dependencies across future bars. Sampling multiple trajectories also provides an ensemble from which forecast distributions and path-level statistics can be derived.

  3. The model is trained on approximately three billion bars across US equities, Chinese A-shares, and cryptocurrencies, spanning seven resolutions from one minute to one day. This breadth gives the model exposure to different market structures and temporal granularities within a shared framework, directly matching the paper's goal of cross-market and cross-resolution forecasting.

  4. The paper evaluates return and volatility prediction across three markets and seven resolutions, and compares against financial-specific and general time-series models. On the reported test window, KiT achieves mean return RankIC of 0.0571 and mean volatility RankIC of 0.6600, leading the listed baselines at each resolution according to Tables 1 and 2. The scale and conditioning ablations also provide evidence about the contribution of key design choices. Overall, there is broad empirical comparison across complementary tasks.

Critical weaknesses
  1. There is temporal leakage in the backtest. The May–December 2025 backtest overlaps the validation period used to select Top-N and rebalance frequency. January–April 2026 also overlaps the stated forecasting test window. Figure 4 is therefore not an independent out-of-sample evaluation. A strictly post-December-2025 backtest or nested rolling protocol is needed.

  2. Short forecasting test window. The headline results use only 1 January–10 April 2026. The mean return RankIC gain over Kronos-base-FT is modest (0.0571 vs. 0.0521), so performance stability across non-overlapping periods and uncertainty estimates are needed.

  3. The reported annualized net returns are high, but universe construction, survivorship/delisting, turnover, market impact, capacity, and executable-price assumptions are not adequately detailed. The stated transaction costs do not resolve these issues.

  4. There are uneven baseline protocols. Baselines differ in fine-tuning, input channels, and training budgets. Specifically, GARCH uses fixed, unfitted coefficients. These differences weaken strong state-of-the-art claims and need matched comparisons or explicit qualification.

  5. Distributional forecasting is not fully evaluated.Averaging 16 generated samples may suppress uncertainty and path variation, while CFG can affect calibration. The authors should justify this estimator and clarify that CFG/model choices were not tuned on the final test data.

Questions for authors
  1. Can the authors rerun the trading backtest on a period strictly after 31 December 2025, with no outcomes from that period used to select Top-N, rebalance frequency, CFG settings, or other choices? Alternatively, can the authors provide a nested rolling protocol that separates all selection, validation, and evaluation windows?

  2. Please separate the two overlap issues explicitly: the backtest overlaps the validation window (May–December 2025) and also overlaps the stated test window (January–April 2026). Which results in Tables 1–2 are computed exclusively on the stated test window, and are any backtest choices informed by those same outcomes?

  3. Can the authors report return/volatility IC and RankIC with uncertainty across multiple non-overlapping test periods and market regimes, as well as per-resolution uncertainty, rather than relying on one short window?

  4. How are survivorship bias, delisted instruments, universe membership, turnover, market impact, capacity, and executable prices handled? Please include sensitivity to higher transaction costs.

  5. Can the authors provide a matched baseline protocol that specifies input channels, fine-tuning data, training budgets, and checkpoint selection for each method? In particular, please justify the fixed, unfitted GARCH coefficients and report a fitted GARCH baseline.

4 comments
AuthorsOpenReview ↗

Comment for Reviewer MkMs (part1)

Thank you for the detailed and constructive checklist, and for noting that the representation is "well-designed" and that flow matching "avoids autoregressive rollout by construction." We ran new experiments for each of your questions and will add all of them to the revision. This part answers Weak1–Weak2 and Q1–Q3; part 2 answers Weak3–Weak5 and Q4–Q5.

Weak1 / Q1 / Q2: Backtest strictly after 31 December 2025

Regarding Q2: all results in Tables 1–2 are computed exclusively on the stated test window (1 Jan – 10 Apr 2026), and no backtest choice (Top-N, rebalance frequency, CFG) was informed by its outcomes.

Thank you for pointing this out. We ran backtests strictly after 31 December 2025, together with further tests that rule out temporal leakage, and we will add them to the main text. All settings were fixed in advance: top-10 holdings, non-overlapping holding windows equal to the forecast horizon, K = 16 paths, CFG weights = 1.

  • Top-10 book, 1 Jan – 10 Apr 2026, fixed 100-stock A-share universe (T+1; limit-up buys skipped; limit-down sells carried over; 2.5 bp commission, 5 bp stamp tax on sells, 10 bp slippage per side). Net cumulative returns: +42.2% (5m), +42.3% (15m), +28.3% (30m), +48.9% (1h), +33.4% (2h), +15.5% (1d), vs +1.0% to +2.1% for the equal-weight universe and +3.8% to +4.2% for CSI 500. At 1m (62 rebalances), costs turn +13.0% gross into −6.3% net.
  • Market-neutral quintile spread, same window, 1,000 names: +3.45% per 20 days (t 10.2; positive on all 9 anchors whose outcomes fall in the window).
  • Two strictly sequential splits instead of a nested rolling protocol. (a) The frozen checkpoint on 13 Apr – 30 Oct 2026: the names KiT ranks lowest-risk realize lower volatility than MLE-GARCH rankings on 96% of anchors (A-share t −9.6; U.S. t −5.0). (b) An A-share KiT-L of identical architecture, recipe and size, trained through 2023-12-31 and validated to 2025-03-31: over 19 held-out months (Apr 2025 – Oct 2026), the market-neutral spread is +2.1% per 20 days (t 3.3).

Weak2 / Q3: IC/RankIC with uncertainty across periods, regimes and resolutions

We will report every result with circular block-bootstrap 95% CIs, Newey–West t-statistics and the share of anchors won.

  • Six non-overlapping quarters (model (b), 73 anchors × ~1,000 names): return RankIC 0.055 / 0.087 / 0.106 / 0.141 / 0.069 / 0.058 for 2025Q2–2026Q3; overall 0.085 [0.053, 0.114], t 5.2.
  • Regimes (paper checkpoint, all 2026 anchors): KiT's volatility RankIC exceeds fitted GARCH and EWMA on 100% of anchors in every calendar month, every tercile of market volatility and of cross-sectional dispersion, and in up and down markets, in both markets (mean gain over GARCH +0.099, t 9.7, A-share; +0.045, t 10.9, U.S.).
  • Per resolution (13 Apr – 30 Oct 2026; volatility RankIC, with the NW t of KiT minus MLE-GARCH in parentheses):
5m15m30m1h2h1d
A-share.764 (8.3).740 (6.0).676 (5.8).651 (6.6).700 (5.6).677 (9.6)
U.S..764 (9.5)––.729 (6.2).776 (7.4).861 (8.9)
Crypto–.852 (6.1).817 (5.4).764 (5.7).743 (4.8).686 (2.2)

KiT leads fitted GARCH in every market × resolution cell.

AuthorsOpenReview ↗

Comment for Reviewer MkMs (part2)

Weak3 / Q4: Universe, survivorship, turnover, capacity, execution prices and costs

We will add a dedicated appendix covering each item:

  • Universe and survivorship. The 1,000-name A-share universe is a seeded random sample of the model's instrument vocabulary, fixed before any 2026 outcome was computed. Cross-sections are point-in-time: a name that delists stays in the universe until its delisting date. The 100-stock universe of the top-10 book is fixed for the whole window. In U.S. equities, names without price-adjustment factors (mostly delisted) are excluded; because all U.S. results are cross-sectional rankings on identical cases, this affects every model equally, and we report no U.S. trading result.
  • Executable prices. The top-10 book opens positions at the anchor bar's close, the last price observed by the model, and exits at the close of the window's last bar. Limit-up buys and names in their first five listing days are skipped; limit-down sells are carried to the next window.
  • Turnover, impact and capacity. Two-way turnover is 1.5–2.0 per rebalance. Market impact beyond the 10 bp slippage is not modelled; as a bound on size, a top-10 selection from the 1,000-name universe has a median capacity of about CNY 400M at 5% of entry-day traded amount.
  • Higher costs. Costs are proportional to traded amount, so the gross–net gap gives a first-order estimate: at 2× / 3× costs the book would still earn about +32% / +22% (5m), +37% / +33% (15m), +25% / +21% (30m), +45% / +41% (1h), +31% / +30% (2h) and +15% / +14% (1d). The break-even cost multiple is about 6× at 5m and 10–21× at coarser resolutions, while the 1m book is cost-dominated. We will add exact reruns at 2× and 3× costs to the revision.

Weak4 / Q5: Matched baselines and fitted GARCH

  • GARCH. You are right that the fixed coefficients made a weak baseline. We now fit GARCH(1,1) by maximum likelihood per instrument and anchor on KiT's own context, and add EWMA and historical volatility. On 13 Apr – 30 Oct 2026, KiT leads fitted GARCH by +0.082 (A-share, t 9.6) and +0.045 (U.S., t 8.9) volatility RankIC on every anchor.
  • Input channels. We add range estimators (Parkinson, Garman–Klass, Rogers–Satchell, Yang–Zhang) and log-HAR built on the same OHLC context as KiT, with the best variant chosen on the test window itself (favoring the baselines). KiT wins in all eight cells (2 markets × 2 windows, Jan–Apr / Apr–Oct 2026), on 86–100% of anchors: over the best range estimator by +0.059 / +0.038 (A-share) and +0.018 / +0.027 (U.S.), and over the best HAR by +0.036 / +0.022 and +0.048 / +0.051 (t 3.1–7.8). With Kronos run on its native OHLCV input (zero-shot), KiT leads by +0.187 (A-share, t 5.8) and +0.087 (U.S., t 9.7).
  • Fine-tuning and checkpoint selection. In the submission, Kronos-base-FT used the checkpoint with the best test score, which favors the baseline. We additionally fine-tuned Kronos-base on A-share daily bars through 2025, with checkpoint selection by validation loss and KiT's context and horizon: volatility RankIC 0.606 vs KiT 0.677 (t 5.1); 90% coverage 0.39 vs 0.64 (t 6.5).
  • Protocol table. We will add a table listing, for every method, its input channels, fine-tuning data, training budget, number of samples and checkpoint selection.

Weak5: The 16-sample estimator and CFG

  • Averaging does not remove path variation from the volatility score. The volatility forecast is the mean over the 16 paths of each path's own realized volatility, not the volatility of the mean path. Only the terminal-return forecast uses the path mean, the natural point estimate for ranking.
  • Number of samples. Volatility RankIC rises monotonically with K (A-share 0.49 / 0.56 / 0.62 / 0.66 / 0.68 for K = 1 / 2 / 4 / 8 / 16), and return RankIC is flat from K = 4. On the same cases, K = 64 further raises volatility RankIC (A-share 0.670 → 0.694, t 6.6; U.S. 0.864 → 0.869, t 6.9) and 90% coverage (A-share 0.65 → 0.71, t 9.7; U.S. 0.72 → 0.79, t 16.9).
  • Calibration. We will add CRPS, interval coverage and PIT for the ensemble. It is far better calibrated than Kronos on identical cases (CRPS 0.60 vs 1.03, t −10.2, U.S.; 0.745 vs 0.826, t −2.6, A-share; 90% coverage 0.73 vs 0.33, U.S.; 0.64 vs 0.46, A-share).
  • CFG and model choice. CFG weights are 1 (the plain conditional model, no guidance) in every reported result, and KiT-L, the largest model, was designated as the evaluated model in advance; neither was tuned on test data.

We will also improve the presentation (data-split description, baseline protocol table, citation formatting). We hope these answers resolve your concerns. If so, we would be grateful if you would consider raising your score. if any item remains open, feel free to ask us!

Hi Reviewer MkMs!

Thank you again for your detailed review. As the discussion period ends in about one day, we would like to kindly follow up on our response. In brief, we have:

  1. run backtests strictly after 31 December 2025 and added strictly sequential splits to rule out temporal leakage (Weak1, Q1–Q2);
  2. reported results across six non-overlapping quarters, market regimes and resolutions, with bootstrap CIs and Newey–West t-statistics (Weak2, Q3);
  3. detailed the universe construction, survivorship handling, turnover, capacity, execution prices (Weak3, Q4);
  4. added a fitted GARCH, EWMA, range-based and HAR baselines, together with a matched fine-tuning protocol for Kronos (Weak4, Q5);
  5. clarified the 16-sample estimator and added calibration metrics (CRPS, coverage, PIT) (Weak5).

We would greatly appreciate it if you could let us know whether these address your concerns, and we are happy to clarify anything further before the deadline. If they do, we hope you might consider updating your score. Thank you!

Thank you for the rebuttal. I am satisfied, and I have raised the scores.

Reviewer_ZeMWOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes KiT, a diffusion-based foundation model for financial OHLCV forecasting that generates future trajectories non-autoregressively using flow matching. Its main contributions include an invertible, scale-free candlestick representation that preserves price constraints and a Diffusion Transformer with specialized conditioning mechanisms.

Pretrained on approximately 3 billion candlestick records across three financial markets and seven temporal resolutions, KiT outperforms existing financial and general-purpose forecasting models in return and volatility prediction. Ablation studies and trading backtests further support its effectiveness and practical potential.

Critical strengths
  1. KiT combines flow matching with an invertible, scale-free OHLCV representation that preserves candlestick constraints. Generating entire future trajectories non-autoregressively is a meaningful alternative to autoregressive financial forecasting, particularly for modeling uncertainty and avoiding sequential error accumulation.

  2. The model is pretrained on approximately 3 billion candlestick records across three markets and evaluated at seven temporal resolutions. Its consistent improvements over financial foundation models and diffusion-based baselines, particularly in volatility forecasting, provide strong empirical support for the proposed framework.

Critical weaknesses
  1. The forecasting test period covers only Jan to Apr 2026, limiting evidence of robustness across different market regimes. Besides, the trading backtest (May 2025 to April 2026) overlaps with the validation period used for strategy selection, raising questions about the reliability of the reported economic gains.

  2. Although generating full predictive distributions is a central motivation, the evaluation primarily focuses on point-based return and volatility metrics. The absence of quantitative assessments of distribution calibration, interval coverage, or probabilistic forecasting accuracy leaves a key claimed advantage of the diffusion framework insufficiently validated.

Questions for authors
  1. The trading backtest covers May 2025 to April 2026, overlapping with the validation period used for strategy selection. How was potential selection bias avoided? Could the authors report performance restricted to the fully held-out test period?
3 comments
AuthorsOpenReview ↗

Comment for Reviewer ZeMW

Thank you for the careful review, and for noting that KiT's improvements, "particularly in volatility forecasting, provide strong empirical support for the proposed framework." We have run new experiments for both weaknesses you raised, and we will add all of them to the revision.

Weak1 / Q1: Backtest restricted to the fully held-out period

  • Top-10 book on the held-out test window (1 Jan – 10 Apr 2026). Rank by predicted terminal return and hold the top 10 equal-weight, on a fixed 100-stock A-share universe, with T+1, limit-up buys skipped, limit-down sells carried over, 2.5 bp commission, 5 bp stamp tax on sells and 10 bp slippage per side. Net cumulative returns over the window are +42.2% (5m), +42.3% (15m), +28.3% (30m), +48.9% (1h), +33.4% (2h) and +15.5% (1d), versus +1.0% to +2.1% for the equal-weight universe and +3.8% to +4.2% for CSI 500. At 1m (62 rebalances), costs turn a +13.0% gross return into −6.3% net.
  • Market-neutral spread on the same window, 1,000 names. The top-minus-bottom quintile by KiT's forecast earns +3.45% per 20 days (t 10.2, positive on all 9 anchors whose outcomes fall in the window); on the same 9 anchors, daily return RankIC is 0.126 (t 6.8).
  • Beyond April 2026. With the checkpoint frozen on 13 Apr – 30 Oct 2026, the names KiT ranks as lowest-risk realize lower volatility than those ranked by MLE-fitted GARCH(1,1) on 96% of anchors (A-share t −9.6; U.S. t −5.0).

Weak1: Robustness across periods and regimes

To address the concern that the original test window was short, we will add a frozen forward test to the main text, and regime and long-window analyses to the appendix.

  • Frozen forward test, 13 Apr – 30 Oct 2026 (data that did not exist at training time; 16 paths, CFG = 1; 1,000 A-share and 824 U.S. names). Daily volatility RankIC is 0.677 [0.646, 0.714] (A-share) and 0.861 [0.849, 0.872] (U.S.), ahead of MLE-fitted GARCH by +0.082 (t 9.6) and +0.045 (t 8.9) on every anchor. On every anchor KiT also leads EWMA (+0.051 / +0.057), zero-shot Kronos with its native OHLCV input (+0.187 / +0.087), and each of the five general TSFMs of Tables 1–2, i.e. Chronos-2, Chronos-Bolt, Sundial, TimesFM 2.5 and TimesFM 3 (+0.11 to +0.16 / +0.06 to +0.11; A-share / U.S.).
  • Every regime. Over all 2026 anchors, KiT beats fitted GARCH and EWMA on 100% of anchors in every calendar month, in every tercile of market volatility and of cross-sectional dispersion, and in rising and falling markets, in both markets (mean gain over GARCH +0.099, t 9.7, A-share; +0.045, t 10.9, U.S.).
  • Returns over 19 held-out months. An A-share KiT-L of the same architecture, recipe and size (training ≤ 2023-12-31, validation to 2025-03-31) attains return RankIC 0.085 [0.053, 0.114] (t 5.2) over April 2025 – October 2026, positive in each of the six quarters (0.055 / 0.087 / 0.106 / 0.141 / 0.069 / 0.058), with a market-neutral spread of +2.1% per 20 days (t 3.3).

Weak2: Distributional evaluation

We agree this was the missing validation of the diffusion design, and will add a new appendix section on the unseen window:

  • Coverage and CRPS. The raw 16-path ensemble covers 64% (A-share) / 73% (U.S.) at the nominal 90% level, so it is under-dispersed, yet far better calibrated than Kronos on identical cases (CRPS 0.60 vs 1.03, t −10.2, U.S.; 0.745 vs 0.826, t −2.6, A-share; coverage 0.73 vs 0.33, U.S.; 0.64 vs 0.46, A-share). Its spread ranks realized absolute errors (Spearman 0.21 A-share, 0.40 U.S.).
  • Largely a finite-ensemble effect. With no retraining, K = 64 paths lift 90% coverage from 0.65 to 0.71 (A-share, t 9.7) and from 0.72 to 0.79 (U.S., t 16.9) on the same cases, and lower CRPS from 0.722 to 0.706 and from 0.606 to 0.582 (t −12.0 / −16.8).
  • Paths and levels. With a sampling temperature τ = 1.10 chosen on Jan–Apr, KiT's U.S. path distribution beats a fitted GARCH simulator in CRPS on realized volatility (0.039 vs 0.046, t −4.5), drawdown (0.308 vs 0.318, t −2.6) and range (0.313 vs 0.341, t −4.8). For volatility levels, KiT's U.S. QLIKE is 0.198 vs 0.260 for fitted GARCH (t −5.7), and the 90% Model Confidence Set contains KiT alone.

We hope these results resolve your concerns. If they do, we would be grateful if you would consider raising your score. If anything remains open, please let us know and we will addres it before the discussion closes.

Thank you for the rebuttal. I have adjusted the scores accordingly.

Thank you for recognizing our work and raising your score, and for the constructive experimental suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.