Pith. sign in

REVIEW 3 major objections 6 minor 21 references

A training-free sequential projection can make any probabilistic OHLC forecast free of both quantile and K-line crossings while barely moving the original numbers.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-30 20:57 UTC pith:IN6E6CGD

load-bearing objection Clean post-hoc fix for both OHLC and quantile crossings; incremental but solid, and the empirics actually back the claim. the 3 major comments →

arxiv 2607.26792 v1 pith:IN6E6CGD submitted 2026-07-29 stat.ML cs.AIcs.CEcs.LGq-fin.CP

Crossing-Free Probabilistic K-Line Forecasts Without Retraining

Runyao Yu , Yuchen Tao , Yujie Chen , Wentao Wang , Derek W. Bunn This is my paper
classification stat.ML cs.AIcs.CEcs.LGq-fin.CP
keywords probabilistic k-line forecastingOHLC consistencyquantile crossingforecast reconciliationsequential projectiontime-series foundation modelspost-hoc correction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Probabilistic forecasts of open, high, low, and close must obey two kinds of order: higher quantiles cannot fall below lower ones, and the high and low must bound the open and close at every quantile. Most models violate one or both rules, and existing fixes either retrain the model, handle only one constraint family, or rearrange values more than necessary. This paper introduces K-line–Quantile Sequential Projection (KQSP), a parameter-free post-hoc step that first projects each quantile row onto valid OHLC relations, then projects each feature column onto monotone quantiles, each time with the smallest squared change. Across a trained neural net and eight foundation models on twelve stocks, KQSP drives both crossing rates to zero, often improves accuracy metrics, and changes the raw forecasts less than cumulative-max, reordering, or hierarchical residual heads. The practical claim is that consistency for probabilistic candlesticks can be enforced after the fact, independently of how the forecasts were generated.

Core claim

KQSP applied to any model’s probabilistic OHLC matrix eliminates both quantile crossing and K-line crossing on every evaluated test set, while matching or improving average quantile loss, MAE, and RMSE and producing smaller mean and maximum corrections than cumulative maximum, reordering, and hierarchical residual constructions.

What carries the argument

K-line–Quantile Sequential Projection (KQSP): first the minimum-distance projection of each quantile’s OHLC row onto the valid candlestick set, then the minimum-distance monotone projection of each feature’s quantile column; Stage 2 is argued to preserve the OHLC inequalities already enforced in Stage 1.

Load-bearing premise

After fixing each row to a valid candlestick, the column-wise monotone projection never reintroduces a high that falls below open or close, or a low that rises above them.

What would settle it

Find a raw forecast matrix where applying the row-wise OHLC projection and then the column-wise isotonic projection yields a matrix that still has at least one K-line violation, or where KQSP’s mean correction and accuracy are worse than joint projection or reordering on the same held-out stock days.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Any existing probabilistic OHLC pipeline can be made crossing-free without changing architecture, loss, or training.
  • When the original model is trusted, KQSP is preferable to reordering or cumulative-max because it displaces forecasts less while still removing both crossing types.
  • Zero-shot foundation models that already emit paths or quantiles can be post-processed into consistent probabilistic candlesticks for trading use.
  • Joint projection is not needed in practice for these data: sequential KQSP matches its correction size and accuracy.
  • Consistency enforcement alone can reduce estimation noise enough to improve pinball loss and point metrics on many stocks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same two-stage geometry could be tried on other multi-feature quantile objects with hard bounds (e.g., bid–ask–mid bands or volume-weighted ranges) without inventing new training heads.
  • If Stage-2 order preservation is the load-bearing step, a single counter-example matrix would force either joint projection or a different sequential order as the default.
  • Finance-specific foundation models that already show low K-line crossing may need KQSP mainly as a cheap safety layer rather than a large accuracy lever.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes K-line–Quantile Sequential Projection (KQSP), a parameter-free, training-free post-hoc operator that reconciles any probabilistic OHLC (K-line) forecast matrix. KQSP first Euclidean-projects each quantile row onto the valid OHLC set CO, then projects each feature column onto the monotone quantile cone CQ. The authors argue that Stage 2 preserves the OHLC inequalities established in Stage 1, so the composition is jointly feasible. Empirically, on 12 stocks and nine backbones (a per-stock ANN plus eight zero-shot foundation models, including finance-specific Kronos), KQSP drives quantile and K-line crossing rates to zero, matches or improves AQL/MAE/RMSE under paired tests, and produces smaller mean/max L2 corrections than cumulative-max, reordering, and hierarchical residual heads, while matching joint projection on accuracy and correction size (Tables III–VI).

Significance. If the results hold, the paper cleanly separates consistency enforcement from forecast generation for probabilistic OHLC outputs. That separation is practically useful: foundation-model and production pipelines can keep existing heads and training recipes and still ship crossing-free K-line densities. Strengths include an explicit minimum-distance construction, a standard order-preservation argument for sequential feasibility, a joint-projection oracle that bounds correction error, chronological held-out tests across heterogeneous equities, and systematic comparison against both post-hoc and retrained hierarchical baselines. The contribution is incremental rather than foundational, but it is well-scoped, falsifiable, and immediately applicable.

major comments (3)
  1. [§4.2 Stage 1 / Eq. (5)] §4.2, Stage 1 (Eq. 5 and the surrounding prose): the K-line projection PO is described only as “check every possible combination in which one or more relations become equalities and select the valid result with the smallest correction.” For a four-variable convex set this is feasible, but the manuscript never gives a complete case list, closed form, or pseudocode. Because KQSP’s central claim is a deterministic, reproducible, parameter-free operator, the Euclidean projector onto CO should be fully specified (or cited to a standard QP/isotonic routine) so that independent implementations cannot diverge on boundary ties and multi-constraint violations.
  2. [§4.2 display (7)] §4.2 “Preservation of K-line consistency” and display (7): the argument that columnwise isotonic projection preserves the four OHLC inequalities is standard monotone-operator reasoning, and Tables III–V empirically show post-KQSP KCR = 0 everywhere. Still, the write-up asserts preservation without a short formal lemma (e.g., that the univariate pool-adjacent-violators map is componentwise monotone and therefore preserves coordinatewise dominance between columns). A brief lemma would make the sequential feasibility claim self-contained and would clarify why the opposite order (PQ then PO) is not used—rowwise PO can break monotonicity, so order is load-bearing.
  3. [§5.1, Table IV] §5.1 / Table IV: Hierarchical1–3 are retrained residual heads, whereas Cumulative Max, Reordering, JP, and KQSP are pure post-hoc maps of the same saved forecasts. Mean/max “correction” for the hierarchical variants is therefore measured against a different native output semantics (residuals vs. direct levels), which weakens the claim that KQSP “produces substantially smaller corrections than … hierarchical” heads. Either restrict the correction comparison to post-hoc methods only, or redefine hierarchical correction relative to an unconstrained head with matched capacity so the L2 comparison is on the same scale.
minor comments (6)
  1. [Figure 1] Figure 1’s numerical example is helpful; consider adding the squared-error arithmetic from Table II directly into the figure caption so the “smaller correction” claim is visible without flipping pages.
  2. [Eq. (4), Table IV] Eq. (4) defines E_corr with a sum of squares but is later discussed interchangeably with “mean corr.” and “max corr.” in native price units (Table IV). State explicitly whether Table IV reports √E_corr, mean absolute change, or mean E_corr per sample.
  3. [§3, §5.2] §3: the sample window ends July 23, 2026 and Kronos pretraining ends June 2024; a one-sentence leakage checklist for all eight FMs (not only Kronos) would strengthen the zero-shot claims.
  4. [Table III] Table III caption and body: “Improved or n.s. 12/12” for QCE/MAE/RMSE is slightly optimistic where individual stocks show non-significant or mixed signs (e.g., Volkswagen MAE/RMSE). Prefer “improved or non-significant under the paired test” wording already used in the text.
  5. [§4] Minor notation: Z vs. bZ vs. c★ vs. bzKQSP switches fonts frequently; a single “raw / after PO / after KQSP” notation paragraph would help.
  6. [References] References [1], [8], [14], [15], [19]–[21] are arXiv preprints; ensure venue/year are updated if accepted versions exist, and that self-citations to hierarchical heads are balanced with the classical quantile-crossing literature already cited ([5], [13]).

Circularity Check

0 steps flagged

No significant circularity: KQSP is an explicit post-hoc projection; zero crossings follow from the operator, and accuracy/correction claims are empirical on held-out tests.

full rationale

KQSP is defined as two minimum-distance projections (PO onto the OHLC set CO, then PQ onto the monotone cone CQ). Achieving QCR = KCR = 0 is the intended feasibility property of that operator (with the stated order-preservation argument for Stage 2), not a fitted or self-cited “prediction” of an external target. No parameter is estimated from the evaluation metrics and then reported as a forecast. Self-citations (e.g., hierarchical residual heads from the authors’ prior work) appear only as comparison baselines that require retraining, not as uniqueness theorems or load-bearing lemmas for KQSP. Predictive claims (AQL/MAE/RMSE, correction size vs. cumulative max, reordering, joint projection) are measured on chronological held-out splits across an ANN and multiple foundation models. The derivation chain is therefore self-contained and non-circular.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 2 invented entities

KQSP itself introduces no fitted constants. Load-bearing content is standard convex projection plus the domain definition of a valid candle. Evaluation inherits ordinary ML training choices for the ANN backbone only; those choices are not part of the KQSP claim.

free parameters (2)
  • ANN architecture and training hyperparameters (per stock) = Stock-specific validation-best among 100 sampled configs
    Hidden layers, width, learning rate, dropout, and lookback are selected on validation for the reference ANN; they affect raw forecast quality but are not parameters of KQSP.
  • Quantile grid T = nine levels 0.1..0.9
    Fixed evaluation grid {0.1,...,0.9}; choice affects measured QCR/AQL but is not learned inside KQSP.
axioms (4)
  • standard math Euclidean projection onto a closed convex set is uniquely defined and nonexpansive.
    Used to define both PO and PQ as minimum-distance maps (§4.2).
  • standard math Isotonic (monotone cone) projection preserves coordinate-wise inequalities between two sequences: if u_q ≤ v_q for all q before projection, the projected sequences still satisfy the inequality.
    Invoked to prove Stage 2 cannot reintroduce K-line crossings (display (7)).
  • domain assumption A valid OHLC vector must satisfy L ≤ min(O,C) and max(O,C) ≤ H at every quantile level.
    Defines the K-line cone CO in §2.2; without it the projection target is undefined.
  • domain assumption Chronological 70/15/15 split with scalers fit on train only yields an honest test of next-day probabilistic OHLC forecasts for the twelve equities studied.
    Underpins all reported QCR/KCR/AQL comparisons (§3).
invented entities (2)
  • K-line–Quantile Sequential Projection (KQSP) no independent evidence
    purpose: Parameter-free operator that maps any Q×4 probabilistic OHLC matrix to a jointly feasible matrix via PO then PQ.
    Central proposed method; defined by equations (3)–(6). Independent evidence is the empirical zero-crossing and correction-size results, not an external physical detection.
  • Joint Projection (JP) baseline no independent evidence
    purpose: Single quadratic program onto the intersection of quantile and K-line constraints, used as an optimality reference for correction size.
    Introduced in §5.1 as a comparison variant rather than a prior method.

pith-pipeline@v1.2.0-daily-grok45 · 18465 in / 3138 out tokens · 86855 ms · 2026-07-30T20:57:18.722782+00:00 · methodology

0 comments
read the original abstract

Probabilistic K-line forecasting describes uncertainty in four complementary prices, namely open--high--low--close (OHLC). However, it introduces two consistency problems: quantile crossing and K-line crossing. Quantile crossing occurs when a higher-quantile forecast falls below a lower-quantile forecast, while K-line crossing occurs when the forecast low exceeds the open or close, or the forecast high falls below the open or close. Existing solutions generally address only one problem through output reordering, specialized architectures, or penalized training objectives. We propose K-line--Quantile Sequential Projection (KQSP), a parameter-free and training-free reconciliation method applicable to forecasts produced by any model. Compared with other crossing solutions, KQSP preserves predictive accuracy while producing substantially smaller corrections to the original forecasts. To mitigate model bias, we evaluate KQSP using various models, including pretrained foundation models. KQSP reduces both quantile and K-line crossing rates to zero for all test data undertaken. These results show that probabilistic K-line consistency can be enforced independently of forecast generation and without retraining.

Figures

Figures reproduced from arXiv: 2607.26792 by Derek W. Bunn, Runyao Yu, Wentao Wang, Yuchen Tao, Yujie Chen.

Figure 1
Figure 1. Figure 1: Overview of the work. Left: illustration of the two crossing issues. Middle: a numerical example of KQSP, which [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Min–max-scaled close prices over the common period from January 1 to July 23, 2026. The panels show three stocks [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

21 extracted references · 2 canonical work pages

  1. [1]

    Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, and Michael Bohlke-Schneider

    Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, Danielle C. Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, and Michael B...

    Pith/arXiv arXiv 2025

  2. [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...

    Pith/arXiv arXiv 2024

  3. [3]

    Jozef Baruník, Martin Hronec, and Ondřej Tobek. 2026. Forecasting stock return distributions around the globe with quantile neural networks.International Journal of Forecasting(2026). doi:10.1016/j.ijforecast.2026.04.004

    work page doi:10.1016/j.ijforecast.2026.04.004 2026

  4. [4]

    Derek Bunn, Arne Andresen, Dipeng Chen, and Sjur Westgaard. 2016. Analysis and forecasting of electricity price risks with quantile factor models.The Energy Journal37, 1 (2016), 101–122

    2016

  5. [5]

    Victor Chernozhukov, Iván Fernández-Val, and Alfred Galichon. 2010. Quantile and probability curves without crossing.Econometrica78, 3 (2010), 1093–1125

    2010

  6. [6]

    Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou. 2024. A decoder-only foundation model for time-series forecasting. InProceedings of the 41st Interna- tional Conference on Machine Learning(Vienna, Austria)(ICML’24). JMLR.org, Article 404, 20 pages

    2024

  7. [7]

    Tilmann Gneiting and Matthias Katzfuss. 2014. Probabilistic Forecasting.Annual Review of Statistics and Its Application1 (2014), 125–151. doi:10.1146/annurev- statistics-062713-085831

    doi:10.1146/annurev- 2014

  8. [8]

    Shi Bin Hoo, Samuel Müller, David Salinas, and Frank Hutter. 2026. From Tables to Time: Extending TabPFN-v2 to Time Series Forecasting. arXiv:2501.02945 [cs.LG] https://arxiv.org/abs/2501.02945 7 Preprint, Under Review, - Yu et al. Table V: Before–after results aggregated across stocks. QCR (%) KCR (%) AQL QCE (%) MAE RMSE Model Before After𝑃Before After𝑃...

    arXiv 2026

  9. [9]

    Wenyang Huang, Tianxiao Gao, Yun Hao, and Xiuqing Wang. 2023. Transformer- Based Forecasting for Intraday Trading in the Shanghai Crude Oil Market: An- alyzing Open-High-Low-Close Prices.Energy Economics127 (2023), 107106. doi:10.1016/j.eneco.2023.107106

    arXiv 2023

  10. [10]

    Wenyang Huang, Huiwen Wang, and Shanshan Wang. 2024. A Structural VAR and VECM Modeling Method for Open-High-Low-Close Data Contained in Candlestick Chart.Financial Innovation10, 1 (2024), 97. doi:10.1186/s40854-024- 00622-6

    work page doi:10.1186/s40854-024- 2024

  11. [11]

    Junkyu Jang, Taehwan Kim, and Sung-Hyuk Park. 2024. Stock Index Forecasting Using an Explainable TAFT Model with Online Data-Driven Social Sentiment Index. InProceedings of the 5th ACM International Conference on AI in Finance (Brooklyn, NY, USA)(ICAIF ’24). Association for Computing Machinery, New York, NY, USA, 787–794. doi:10.1145/3677052.3698618

    arXiv 2024

  12. [12]

    Jinwoong Kim and Sangjin Park. 2025. IKNet: Interpretable Stock Price Prediction via Keyword-Guided Integration of News and Technical Indicators. InProceedings of the 6th ACM International Conference on AI in Finance (ICAIF ’25). Association for Computing Machinery, New York, NY, USA, 709–717. doi:10.1145/3768292. 3770343

    doi:10.1145/3768292 2025

  13. [13]

    Horowitz, and Jian Huang

    Guohao Shen, Yuling Jiao, Yuanyuan Lin, Joel L. Horowitz, and Jian Huang. 2024. Nonparametric Estimation of Non-Crossing Quantile Regression Process with Deep ReQU Neural Networks.Journal of Machine Learning Research25, 88 (2024), 1–75. https://jmlr.org/papers/v25/22-0488.html

    2024

  14. [14]

    Xiaoming Shi, Shiyu Wang, Yuqi Nie, Dianqi Li, Zhou Ye, Qingsong Wen, and Ming Jin. 2025. Time-MoE: Billion-Scale Time Series Foundation Models with Mixture of Experts. arXiv:2409.16040 [cs.LG] https://arxiv.org/abs/2409.16040

    Pith/arXiv arXiv 2025

  15. [15]

    Yu Shi, Zongliang Fu, Shuo Chen, Bohan Zhao, Wei Xu, Changshui Zhang, and Jian Li. 2026. Kronos: A foundation model for the language of financial markets. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 25366–25373

    2026

  16. [16]

    James W Taylor and Derek W Bunn. 1999. A quantile regression approach to generating prediction intervals.Management Science45, 2 (1999), 225–237

    1999

  17. [17]

    Gerald Woo, Chenghao Liu, Akshat Kumar, Caiming Xiong, Silvio Savarese, and Doyen Sahoo. 2024. Unified training of universal time series forecasting transformers. InProceedings of the 41st International Conference on Machine Learning(Vienna, Austria)(ICML’24). JMLR.org, Article 2178, 25 pages

    2024

  18. [18]

    Xuanling Yang, Zhoufan Zhu, Dong Li, and Ke Zhu. 2024. Asset pricing via the conditional quantile variational autoencoder.Journal of Business & Economic Statistics42, 2 (2024), 681–694

    2024

  19. [19]

    Runyao Yu, Chenhui Gu, Jochen Stiasny, Qingsong Wen, Wasim Sarwar Dilov, Lianlian Qi, and Jochen L. Cremer. 2026. PriceFM: Foundation Model for Proba- bilistic Electricity Price Forecasting. arXiv:2508.04875 [cs.CE]

    Pith/arXiv arXiv 2026

  20. [20]

    Bunn, Jochen Stiasny, Wentao Wang, Yujie Chen, Tara Esterl, Peter Palensky, and Jochen L

    Runyao Yu, Julia Lin, Derek W. Bunn, Jochen Stiasny, Wentao Wang, Yujie Chen, Tara Esterl, Peter Palensky, and Jochen L. Cremer. 2026. A market-rule-informed neural network for efficient imbalance electricity price forecasting.Advanced Engineering Informatics76 (2026), 105083. doi:10.1016/j.aei.2026.105083

    arXiv 2026

  21. [21]

    Bunn, Qingsong Wen, Hongye Guo, and Jochen L

    Runyao Yu, Yuchen Tao, Fabian Leimgruber, Tara Esterl, Jochen Stiasny, Derek W. Bunn, Qingsong Wen, Hongye Guo, and Jochen L. Cremer. 2026. OrderFusion: Encoding Orderbook for End-to-End Probabilistic Intraday Electricity Price Fore- casting. arXiv:2502.06830 [q-fin.CP] 8

    Pith/arXiv arXiv 2026