pith. sign in

arxiv: 2502.06830 · v5 · submitted 2025-02-05 · 💱 q-fin.CP · cs.AI· cs.LG

OrderFusion: Encoding Orderbook for End-to-End Probabilistic Intraday Electricity Price Forecasting

Runyao Yu , Yuchen Tao , Fabian Leimgruber , Tara Esterl , Jochen Stiasny , Derek W. Bunn , Qingsong Wen , Hongye Guo
show 1 more author
Jochen L. Cremer
This is my paper

Pith reviewed 2026-05-23 03:57 UTC · model grok-4.3

classification 💱 q-fin.CP cs.AIcs.LG
keywords orderbookprobabilistic forecastingintraday electricity pricescontinuous intraday marketsquantile estimationbuy-sell dynamics
0
0 comments X

The pith

Order fusion encodes the orderbook to learn buy-sell interactions for probabilistic intraday electricity price forecasts.

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

The paper presents OrderFusion, an end-to-end probabilistic forecasting model for continuous intraday electricity prices that directly processes the orderbook. It builds an interaction-aware representation of buy and sell dynamics instead of relying on aggregated features or treating the data as a simple multivariate time series. The approach also applies hierarchical estimation to keep predicted quantiles from crossing. Experiments across high- and low-liquidity European CID price indices show consistent gains over conventional baselines, with ablation results pointing to the fusion component as a main contributor.

Core claim

Order fusion supplies an end-to-end parameter-efficient probabilistic forecasting model that learns an interaction-aware representation of the buy-sell dynamics and hierarchically estimates quantiles under non-crossing constraints, yielding improved forecasts on CID price indices.

What carries the argument

The order fusion methodology, which encodes the full orderbook to produce an interaction-aware representation of buy-sell dynamics for the forecasting task.

If this is right

  • The model delivers consistent improvements on CID price indices in both high- and low-liquidity European markets.
  • Hierarchical non-crossing quantile estimation removes a common artifact in probabilistic forecasts.
  • Ablation studies isolate the contribution of the fusion component over baseline representations.

Where Pith is reading between the lines

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

  • The same orderbook encoding could be tested in other continuous double-auction settings such as certain commodity or FX markets.
  • Pairing the learned interaction representation with separate renewable or demand forecasts might tighten imbalance risk estimates.
  • Parameter efficiency opens the possibility of retraining the model on rolling windows at higher intraday frequencies.

Load-bearing premise

The full buy-sell interaction structure of the orderbook is the key missing element in aggregated or time-series treatments, and learning this structure through fusion will produce measurable forecast gains.

What would settle it

If a model that uses only conventional aggregated features or multivariate time-series inputs matches or exceeds OrderFusion performance on the same CID price index datasets, the claim that interaction structure supplies the decisive advantage would be falsified.

Figures

Figures reproduced from arXiv: 2502.06830 by Derek W. Bunn, Fabian Leimgruber, Hongye Guo, Jochen L. Cremer, Jochen Stiasny, Qingsong Wen, Runyao Yu, Tara Esterl, Yuchen Tao.

Figure 1
Figure 1. Figure 1: Analysis of orderbook and price indices. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Structure of OrderFusion. The model takes the 2 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Result analysis on the testing data for three price indices. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
read the original abstract

Probabilistic intraday electricity price forecasting is becoming increasingly important for short-term power-system operation. With increasing renewable generation, demand-side flexibility, and storage assets, market participants need to adjust their positions under uncertainty closer to delivery. Continuous intraday (CID) markets support this process by providing updated price signals, helping participants manage imbalance exposure and operational risk. Unlike auction markets, CID trading in many jurisdictions is characterized by the continuous posting of buy and sell orders. This dynamic orderbook microstructure of price formation presents special challenges for price forecasting. Conventional methods represent the orderbook via domain features aggregated from buy and sell trades, or by treating it as a multivariate time series, but such representations neglect the full buy-sell interaction structure of the orderbook. This research therefore develops a new order fusion methodology, which is an end-to-end and parameter-efficient probabilistic forecasting model that learns a interaction-aware representation of the buy-sell dynamics. Furthermore, as quantile crossing is often a problem in probabilistic forecasting, this approach hierarchically estimates the quantiles with non-crossing constraints. Extensive experiments on CID price indices across high- and low-liquidity European markets demonstrate consistent improvements over conventional baselines, and ablation studies highlight the contributions of the main components.The methodology is available at: https://runyao-yu.github.io/OrderFusion/.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 3 minor

Summary. The paper proposes OrderFusion, an end-to-end probabilistic forecasting architecture for intraday electricity prices that fuses buy- and sell-side orderbook data to learn interaction-aware representations of market microstructure. It augments this with a hierarchical non-crossing quantile layer and reports consistent outperformance versus aggregated-feature and multivariate time-series baselines on CID price indices from high- and low-liquidity European markets, supported by ablation experiments.

Significance. If the reported gains hold under rigorous scrutiny, the work supplies a parameter-efficient mechanism for incorporating full buy-sell interaction structure that conventional representations omit, with direct relevance to short-term risk management in renewable-heavy power systems. The public release of the methodology code is a clear strength for reproducibility.

minor comments (3)
  1. [Abstract] Abstract: the claim of 'consistent improvements' would be strengthened by reporting the magnitude of gains (e.g., average CRPS reduction) and the number of markets or test periods evaluated.
  2. [Section 4] Section 4 (Experiments): confirm that the temporal train/validation/test split respects causality and that no future information leaks into the orderbook encoding at each forecast origin.
  3. [Figure 2] Figure 2 and ablation tables: add standard-error bands or statistical significance markers to the performance differences so readers can judge whether the fusion component's contribution is distinguishable from noise.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the constructive summary, positive assessment of significance, and recommendation of minor revision. No major comments were listed in the report.

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained

full rationale

The paper introduces an end-to-end neural architecture (OrderFusion) trained directly on market orderbook data and evaluated via standard train/test splits on CID price indices. No derivation step reduces a claimed prediction or uniqueness result to a fitted parameter by construction, nor does any load-bearing premise rest on self-citation chains. The hierarchical non-crossing quantile layer is presented as an auxiliary constraint rather than the source of the reported gains. Ablation studies and market-specific comparisons are described as external validation, keeping the central claim independent of its own inputs.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the effectiveness of a learned neural representation of orderbook interactions plus the non-crossing quantile constraint; the model itself contains many fitted parameters whose contribution is asserted via empirical gains.

free parameters (1)
  • neural network weights and fusion parameters
    All parameters of the end-to-end OrderFusion model are fitted to training data; the claimed performance gains depend on these learned values.
axioms (1)
  • domain assumption Orderbook microstructure contains learnable buy-sell interaction signals that improve probabilistic price forecasts beyond aggregated or separate time-series representations
    Invoked in the motivation for developing the fusion methodology rather than using conventional orderbook encodings.

pith-pipeline@v0.9.0 · 5807 in / 1337 out tokens · 42410 ms · 2026-05-23T03:57:58.277215+00:00 · methodology

discussion (0)

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

Forward citations

Cited by 3 Pith papers

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

  1. Scenario generation of intraday electricity price paths for optimal trading in continuous markets

    stat.AP 2026-05 unverdicted novelty 6.0

    A kernel-based regression model plus scenario generation from forecast errors and a new Support Vector Sorting step produces ensemble price trajectories that improve both statistical accuracy and trading profits over ...

  2. A Market-Rule-Informed Neural Network for Efficient Imbalance Electricity Price Forecasting

    q-fin.CP 2026-05 unverdicted novelty 5.0

    A market-rule-informed neural network for imbalance electricity price forecasting matches generic deep learning accuracy while using substantially fewer parameters and less training time.

  3. Deep Learning for Electricity Price Forecasting: A Review of Day-Ahead, Intraday, and Balancing Electricity Markets

    q-fin.CP 2026-02 unverdicted novelty 3.0

    A structured review organizes deep learning models for electricity price forecasting via a backbone-head-loss taxonomy and identifies gaps in intraday and balancing market applications.

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages · cited by 3 Pith papers · 3 internal anchors

  1. [1]

    Padding values are set to a constantc= 10,000to ensure they do not occur in the data

    Dual Masking Layer:As the number of matched trades varies between samples, we apply pre-padding to align all input sequences to a maximum lengthT max. Padding values are set to a constantc= 10,000to ensure they do not occur in the data. Thus, the input dimension is standardized toR Tmax×3. To identify valid timesteps, we define a binary padding maskB (s) ...

    work page

  2. [2]

    2: C(s) i,k = ( X(s) i ifk= 0, C(s) i,k−1 |C (¯s) i,k−1 ifk≥1, (9) wherekdenotes the degree of interactions, and¯sis the opposite side ofs

    Iterative Fusion Layer:As buyers and sellers iteratively adjust their bids and offers based on observed quotes from the opposite side, reflecting strategic interactions [13], we design a series of iterative fusion layers to enable representation learning of such buy-sell interactions, as illustrated in Fig. 2: C(s) i,k = ( X(s) i ifk= 0, C(s) i,k−1 |C (¯s...

    work page

  3. [3]

    We apply global average pooling to obtain the attention- weighted average representationU i ∈R F : Ui = 1 Tmax TmaxX j=1 Ci[j].(15) C

    Aggregation Layer:All of the fused representations at different degrees are combined via residual connection [31] to produce the higher-level representationC i ∈R Tmax×F : Ci = KX k=1 C(+) i,k +C (−) i,k ,(14) whereKdenotes the maximum degree of interactions, and the summation is element-wise addition. We apply global average pooling to obtain the attenti...

    work page 2022

  4. [4]

    Short-term electricity trading for system balancing: An empirical analysis of the role of intraday trading in balancing germany’s electricity system,

    C. Koch and L. Hirth, “Short-term electricity trading for system balancing: An empirical analysis of the role of intraday trading in balancing germany’s electricity system,”Renewable and Sustainable Energy Reviews, vol. 113, p. 109 275, 2019

    work page 2019

  5. [6]

    Variance sta- bilizing transformations for electricity spot price fore- casting,

    B. Uniejewski, R. Weron, and F. Ziel, “Variance sta- bilizing transformations for electricity spot price fore- casting,”IEEE Transactions on Power Systems, vol. 33, no. 2, pp. 2219–2229, 2018.DOI: 10.1109/TPWRS. 2017.2734563

    work page doi:10.1109/tpwrs 2018

  6. [7]

    PriceFM: Foundation Model for Probabilistic Electricity Price Forecasting

    R. Yu et al.,Pricefm: Foundation model for proba- bilistic electricity price forecasting, 2025. arXiv: 2508. 04875[cs.CE]. [Online]. Available: https://arxiv.org/ abs/2508.04875

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  7. [8]

    Forecasting the intra- day spread densities of electricity prices,

    E. Abramova and D. Bunn, “Forecasting the intra- day spread densities of electricity prices,”Energies, vol. 13, no. 3, 2020,ISSN: 1996-1073.DOI: 10.3390/ en13030687 [Online]. Available: https : / / www. mdpi . com/1996-1073/13/3/687

    work page 2020

  8. [9]

    Multivariate simulation-based forecasting for intraday power markets: Modeling cross- product price effects,

    S. Hirsch and F. Ziel, “Multivariate simulation-based forecasting for intraday power markets: Modeling cross- product price effects,”Applied Stochastic Models in Business and Industry, vol. 40, no. 6, pp. 1571–1595, 2024

    work page 2024

  9. [10]

    Bayesian hierarchical probabilistic forecasting of intraday electricity prices,

    D. Nickelsen and G. M ¨uller, “Bayesian hierarchical probabilistic forecasting of intraday electricity prices,” Applied Energy, vol. 380, p. 124 975, 2025

    work page 2025

  10. [11]

    Ensemble forecasting for intraday electricity prices: Simulating trajectories,

    M. Narajewski and F. Ziel, “Ensemble forecasting for intraday electricity prices: Simulating trajectories,”Ap- plied Energy, vol. 279, p. 115 801, 2020

    work page 2020

  11. [12]

    Price jumps on the intraday energy market - design and implementation of an alarm system with machine learning methods,

    R. Lackes, J. Sengewald, and M. Wilz, “Price jumps on the intraday energy market - design and implementation of an alarm system with machine learning methods,” in 29th Pacific Asia Conference on Information Systems, PACIS 2025, Kuala Lumpur, Malaysia, July 6-9, 2025, M. D. Myers, R. A. Alias, W. F. Boh, R. M. Davisin, B. Tan, and N. Z. A. Rahim, Eds., 202...

    work page 2025

  12. [13]

    Beating the na ¨ıve—combining lasso with na ¨ıve intraday elec- tricity price forecasts,

    G. Marcjasz, B. Uniejewski, and R. Weron, “Beating the na ¨ıve—combining lasso with na ¨ıve intraday elec- tricity price forecasts,”Energies, vol. 13, no. 7, 2020, ISSN: 1996-1073.DOI: 10.3390/en13071667 [Online]. Available: https://www.mdpi.com/1996-1073/13/7/1667

    work page doi:10.3390/en13071667 2020

  13. [14]

    Econometric modelling and forecasting of intraday electricity prices,

    M. Narajewski and F. Ziel, “Econometric modelling and forecasting of intraday electricity prices,”Journal of Commodity Markets, vol. 19, p. 100 107, 2020

    work page 2020

  14. [15]

    Trading on short-term path forecasts of intraday electricity prices,

    T. Serafin, G. Marcjasz, and R. Weron, “Trading on short-term path forecasts of intraday electricity prices,” Energy Economics, vol. 112, p. 106 125, 2022

    work page 2022

  15. [16]

    Unraveling in guessing games: An experi- mental study,

    R. Nagel, “Unraveling in guessing games: An experi- mental study,”The American economic review, vol. 85, no. 5, pp. 1313–1326, 1995

    work page 1995

  16. [17]

    Efficient capital markets,

    E. F. Fama, “Efficient capital markets,”Journal of finance, vol. 25, no. 2, pp. 383–417, 1970

    work page 1970

  17. [18]

    Short-term price forecasting models based on artificial neural networks for intraday sessions in the iberian electricity market,

    C. Monteiro, I. J. Ramirez-Rosado, L. A. Fernandez- Jimenez, and P. Conde, “Short-term price forecasting models based on artificial neural networks for intraday sessions in the iberian electricity market,”Energies, vol. 9, no. 9, p. 721, 2016

    work page 2016

  18. [19]

    Prob- abilistic price forecasting for day-ahead and intraday markets: Beyond the statistical model,

    J. R. Andrade, J. Filipe, M. Reis, and R. J. Bessa, “Prob- abilistic price forecasting for day-ahead and intraday markets: Beyond the statistical model,”Sustainability, vol. 9, no. 11, p. 1990, 2017

    work page 1990

  19. [20]

    Forecasting the price dis- tribution of continuous intraday electricity trading,

    T. Janke and F. Steinke, “Forecasting the price dis- tribution of continuous intraday electricity trading,” Energies, vol. 12, no. 22, p. 4262, 2019

    work page 2019

  20. [21]

    Under- standing intraday electricity markets: Variable selec- tion and very short-term price forecasting using lasso,

    B. Uniejewski, G. Marcjasz, and R. Weron, “Under- standing intraday electricity markets: Variable selec- tion and very short-term price forecasting using lasso,” International Journal of Forecasting, vol. 35, no. 4, pp. 1533–1547, 2019,ISSN: 0169-2070.DOI: https:// doi.org/10.1016/j.ijforecast.2019.02.001

    work page doi:10.1016/j.ijforecast.2019.02.001 2019

  21. [22]

    Simulation-based forecasting for intraday power markets: Modelling fundamental drivers for location, shape and scale of the price distribution,

    S. Hirsch and F. Ziel, “Simulation-based forecasting for intraday power markets: Modelling fundamental drivers for location, shape and scale of the price distribution,” The Energy Journal, vol. 45, no. 3, pp. 107–144, 2024

    work page 2024

  22. [23]

    R. Yu, R. Wu, Y . Han, and J. L. Cremer,Orderbook feature learning and asymmetric generalization in in- traday electricity markets, 2025. arXiv: 2510 . 12685 [q-fin.CP]. [Online]. Available: https://arxiv.org/ abs/2510.12685

    work page arXiv 2025

  23. [24]

    Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting,

    T. Zhou, Z. Ma, Q. Wen, X. Wang, L. Sun, and R. Jin, “Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting,” inInter- national conference on machine learning, PMLR, 2022, pp. 27 268–27 286

    work page 2022

  24. [25]

    iTransformer: Inverted Transformers Are Effective for Time Series Forecasting

    Y . Liu et al., “Itransformer: Inverted transformers are effective for time series forecasting,”arXiv preprint arXiv:2310.06625, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  25. [26]

    A time series is worth 64 words: Long-term forecast- ing with transformers,

    Y . Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam, “A time series is worth 64 words: Long-term forecast- ing with transformers,” inInternational Conference on Learning Representations, 2023

    work page 2023

  26. [27]

    Timesnet: Temporal 2d-variation modeling for general time series analysis,

    H. Wu, T. Hu, Y . Liu, H. Zhou, J. Wang, and M. Long, “Timesnet: Temporal 2d-variation modeling for general time series analysis,” inThe Eleventh Interna- tional Conference on Learning Representations, 2023. [Online]. Available: https://openreview.net/forum?id= ju Uqw384Oq

    work page 2023

  27. [28]

    Timexer: Empowering transformers for time series forecasting with exogenous variables,

    Y . Wang et al., “Timexer: Empowering transformers for time series forecasting with exogenous variables,” inAdvances in Neural Information Processing Sys- tems, A. Globerson et al., Eds., vol. 37, Curran As- sociates, Inc., 2024, pp. 469–498. [Online]. Available: https : / / proceedings . neurips . cc / paper files / paper / 2024/file/0113ef4642264adc2e69...

    work page 2024

  28. [29]

    A hybrid model for gefcom2014 probabilistic electricity price forecasting,

    K. Maciejowska and J. Nowotarski, “A hybrid model for gefcom2014 probabilistic electricity price forecasting,” International Journal of Forecasting, vol. 32, no. 3, pp. 1051–1056, 2016,ISSN: 0169-2070.DOI: https:// doi.org/10.1016/j.ijforecast.2015.11.008

    work page doi:10.1016/j.ijforecast.2015.11.008 2016

  29. [30]

    Averag- ing predictive distributions across calibration windows for day-ahead electricity price forecasting,

    T. Serafin, B. Uniejewski, and R. Weron, “Averag- ing predictive distributions across calibration windows for day-ahead electricity price forecasting,”Energies, vol. 12, no. 13, 2019,ISSN: 1996-1073.DOI: 10.3390/ en12132561 10

    work page 2019

  30. [31]

    Measuring the information content of stock trades,

    J. Hasbrouck, “Measuring the information content of stock trades,”The Journal of Finance, vol. 46, no. 1, pp. 179–207, 1991

    work page 1991

  31. [32]

    Hasbrouck,Empirical market microstructure: The institutions, economics, and econometrics of securities trading

    J. Hasbrouck,Empirical market microstructure: The institutions, economics, and econometrics of securities trading. Oxford University Press, 2007

    work page 2007

  32. [33]

    Hawkes processes in finance,

    E. Bacry, I. Mastromatteo, and J.-F. Muzy, “Hawkes processes in finance,”Market Microstructure and Liq- uidity, vol. 1, no. 01, p. 1 550 005, 2015

    work page 2015

  33. [34]

    Aggregated residual transformations for deep neural networks,

    S. Xie, R. Girshick, P. Doll ´ar, Z. Tu, and K. He, “Aggregated residual transformations for deep neural networks,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 1492–1500

    work page 2017

  34. [35]

    Comparing predictive accuracy,

    F. X. Diebold and R. S. Mariano, “Comparing predictive accuracy,”Journal of Business & economic statistics, vol. 20, no. 1, pp. 134–144, 2002

    work page 2002

  35. [36]

    Quantile and probability curves without crossing,

    V . Chernozhukov, I. Fern ´andez-Val, and A. Galichon, “Quantile and probability curves without crossing,” Econometrica, vol. 78, no. 3, pp. 1093–1125, 2010. DOI: https://doi.org/10.3982/ECTA7880 eprint: https: //onlinelibrary.wiley.com/doi/pdf/10.3982/ECTA7880

    work page doi:10.3982/ecta7880 2010

  36. [37]

    Forecasting day-ahead electricity prices: A review of state-of-the-art algorithms, best practices and an open-access benchmark,

    J. Lago, G. Marcjasz, B. De Schutter, and R. Weron, “Forecasting day-ahead electricity prices: A review of state-of-the-art algorithms, best practices and an open-access benchmark,”Applied Energy, vol. 293, p. 116 983, 2021,ISSN: 0306-2619.DOI: https://doi.org/ 10.1016/j.apenergy.2021.116983 [Online]. Available: https : / / www. sciencedirect . com / scie...

    work page doi:10.1016/j.apenergy.2021.116983 2021

  37. [38]

    Adam: A method for stochastic optimization,

    D. Kingma and J. Ba, “Adam: A method for stochastic optimization,” inInternational Conference on Learning Representations (ICLR), San Diega, CA, USA, 2015

    work page 2015

  38. [39]

    Searching for Activation Functions

    P. Ramachandran, B. Zoph, and Q. V . Le,Searching for activation functions, 2017. arXiv: 1710 . 05941 [cs.NE]. [Online]. Available: https://arxiv.org/abs/ 1710.05941 APPENDIXA ENGINEERINGCONFIGURATION A. Data Scaling A ROBUSTSCALERis fitted strictly on the training data to avoid data leakage, and the fitted scaler is applied to the training, validation, a...

    work page internal anchor Pith review Pith/arXiv arXiv 2017