Paper: arXiv 2502.06830
Authors: Runyao Yu, Yuchen Tao, Fabian Leimgruber, Tara Esterl, Jochen Stiasny, Derek W. Bunn, Qingsong Wen, Hongye Guo, Jochen L. Cremer
Abstract
Probabilistic intraday electricity price forecasting is becoming increasingly important with the growth of renewable generation and the rise in demand-side engagement. Their uncertainties have increased the trading risks closer to delivery and the subsequent imbalance settlement costs. As a consequence, intraday trading has emerged to mitigate these risks. Unlike auction markets, intraday trading in many jurisdictions is characterized by the continuous posting of buy and sell orders on power exchange platforms. 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 full 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 the market price indices across high-liquidity (German) and low-liquidity (Austrian) markets demonstrate consistent improvements over conventional baselines, and ablation studies highlight the contributions of the main modeling components. The methodology is available at: https://runyao-yu.github.io/OrderFusion/.
Complexity vs Empirical Score
- Math Complexity: 7.5/10
- Empirical Rigor: 8.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper employs advanced deep learning architectures (e.g., attention-based encoders) and formal probabilistic modeling with quantile constraints, representing high mathematical complexity. It also demonstrates strong empirical rigor through extensive experiments on real market data (German and Austrian electricity markets), ablation studies, and references to a public GitHub repository for reproducibility.
Research Flowchart
flowchart TD A["Research Goal:<br>Probabilistic Intraday<br>Electricity Price Forecasting"] --> B["Input Data:<br>Orderbook Microstructure<br>Continuous Buy/Sell Orders"] B --> C["Methodology:<br>OrderFusion<br>End-to-End Encoder"] C --> D["Key Process:<br>Learn Interaction-Aware<br>Buy-Sell Dynamics"] D --> E["Computational Step:<br>Hierarchical Quantile<br>Estimation w/ Non-Crossing Constraints"] E --> F["Outcomes:<br>Probabilistic Forecasts<br>Reduced Trading Risk"] F --> G["Results:<br>Improvements over Baselines<br>High & Low Liquidity Markets"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes OrderFusion, an end-to-end and parameter‑efficient probabilistic forecasting model that explicitly encodes buy‑sell interactions in continuous intraday orderbooks and produces non‑crossing quantile forecasts.
Source sentence
We propose OrderFusion, an end-to-end and parameter‑efficient forecasting model that explicitly encodes buy‑sell interactions in CID orderbooks while producing non‑crossing probabilistic forecasts.
Assumes 4
- Recent trades carry the most predictive information under the market efficiency hypothesis. market
Source sentence
To reflect the prior that recent trades carry the most predictive information under the market efficiency hypothesis [33], [34], [35], we define a binary temporal mask D(s)_i...
- The CID market is weak‑form efficient, i.e., recent prices already reflect predictive information contained in historical orders. market
Source sentence
Under the Efficient Market Hypothesis (EMH), a market is weak-form efficient if recent prices already reflect predictive information contained in historical orders [28].
- Orderbook data can be represented as variable‑length 2‑D sequences of price, volume, and relative time delta. data
Source sentence
Each sample is represented as a variable-length 2D sequence: X(s)_i = [P(s)_t1 V(s)_t1 ∇t1
- Padding with a constant value (c = 10,000) does not appear in real data and can be used to align sequences. model
Source sentence
Padding values are set to a constant c = 10,000 to ensure they do not occur in the data.
Data
- German CID market order-book commercial
Source sentence
The German and Austrian CID markets are selected to evaluate the proposed method under different liquidity regimes [17].
- Austrian CID market order-book commercial
Source sentence
The German and Austrian CID markets are selected to evaluate the proposed method under different liquidity regimes [17].
Costs
Not applicable
Finds 5
- OrderFusion achieves the lowest AQL (5.64) and zero AQCR, outperforming all baselines on both German and Austrian markets. performance
Source sentence
OrderFusion consistently achieves superior probabilistic and pointwise forecasting accuracy
- Against 1D encoding baselines, OrderFusion improves AQL by 16.37% to 18.13%, showing raw orderbook information is more predictive than domain features. predictive-accuracy
Source sentence
OrderFusion achieves AQL improvements of 16.37%, 18.13%, and 14.86% over the 15‑min VWAP, last-price, and exhaustive-feature baselines, respectively.
- Removing the temporal mask degrades performance, increasing AQL by 30.51%, confirming the importance of recent trades under market efficiency. statistical
Source sentence
Results in Table III show that removing the mask leads to a 30.51% increase in AQL...
- OrderFusion eliminates quantile crossing, achieving an AQCR of 0.00% whereas 1D‑encoded methods exhibit non‑zero AQCR values. performance
Source sentence
By design, OrderFusion eliminates quantile crossing and consistently achieves an AQCR of 0.00%.
- The optimal temporal cutoff length varies by market (e.g., Austrian ID3 optimal cutoff = 1), indicating differing degrees of market efficiency. other
Source sentence
Austrian ID3 has an optimal cutoff of 1, implying near‑perfect market efficiency and indicating that traders rely on the latest pair of buy‑sell trades to form price expectations.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 35 related papers). Coverage is partial, so this is neither replication nor refutation.
7 papers in the archive : Improving Algorithmic Trading for BESS on Intraday … ; The Evolution of Probabilistic Price Forecasting Techniques: …
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (78)
- Orderbook Feature Learning and Asymmetric Generalization in Intraday Electricity Markets — → 2510.12685
- Price jumps on the intraday energy market - design and implementation of an alarm system with machine learning methods (2025) — link
- iTransformer: Inverted Transformers Are Effective for Time Series Forecasting (2023) — link
- A Time Series is Worth 64 Words: Long-term Forecasting with Transformers (2023)
- Aggregated residual transformations for deep neural networks (2017)
- Scaling Learning Algorithms Towards AI (2007)
- A Fast Learning Algorithm for Deep Belief Nets (2006)
- Deep learning (2016)
and 70 more
- Timemixer: Decomposable multiscale mixing for time series forecasting (2024) — link
- Measuring the information content of stock trades (1991)
- Empirical market microstructure: The institutions, economics, and econometrics of securities trading (2007)
- Hawkes processes in finance (2015)
- Continuous auctions and insider trading (1985)
- Unraveling in guessing games: An experimental study (1995)
- Efficient capital markets: A review of theory and empirical work (1970)
- Multi-scale cross-attention transformer encoder for event classification (2024)
- Nonparametric estimation of non-crossing quantile regression process with deep requ neural networks (2024)
- Generative quantile regression with variability penalty (2024)
- Non-stationary transformers: Exploring the stationarity in time series forecasting (2022)
- Ske2Grid: skeleton-to-grid representation learning for action recognition (2023)
- Learning Quantile Functions without Quantile Crossing for Distribution-free Time Series Forecasting (2022)
- A new approach for the vanishing gradient problem on sigmoid activation (2020)
- Searching for Activation Functions (2017) — link
- Forecasting day-ahead electricity prices: A review of state-of-the-art algorithms, best practices and an open-access benchmark (2021) — link
- The “German Paradox” in the balancing power markets (2017) — link
- Short-term electricity trading for system balancing: An empirical analysis of the role of intraday trading in balancing Germany’s electricity system (2019)
- Probabilistic price forecasting for day-ahead and intraday markets: Beyond the statistical model (2017)
- Forecasting the price distribution of continuous intraday electricity trading (2019)
- Ensemble forecasting for intraday electricity prices: Simulating trajectories (2020)
- Econometric modelling and forecasting of intraday electricity prices (2020)
- Multivariate probabilistic forecasting of intraday electricity prices using normalizing flows (2023)
- Trading on short-term path forecasts of intraday electricity prices (2022)
- Multivariate simulation-based forecasting for intraday power markets: Modeling cross-product price effects (2024)
- Simulation-based forecasting for intraday power markets: Modelling fundamental drivers for location, shape and scale of the price distribution (2024)
- Understanding intraday electricity markets: Variable selection and very short-term price forecasting using LASSO (2019) — link
- Short-term price forecasting models based on artificial neural networks for intraday sessions in the Iberian electricity market (2016)
- Towards the prediction of electricity prices at the intraday market using shallow and deep-learning methods (2021)
- Neural network based model comparison for intraday electricity price forecasting (2019)
- Intraday electricity price forecasting via LSTM and trading strategy for the power market: A case study of the West Denmark DK1 grid region (2024)
- Quantile and Probability Curves Without Crossing (2010) — link
- A hybrid model for GEFCom2014 probabilistic electricity price forecasting (2016) — link
- Construction of optimal prediction intervals for load forecasting problems (2010)
- Comprehensive review of neural network-based prediction intervals and new advances (2011)
- Going deeper with convolutions (2015)
- Pointpillars: Fast encoders for object detection from point clouds (2019)
- Classification of time-series images using deep convolutional neural networks (2018)
- Dicenet: Dimension-wise convolutions for efficient networks (2020)
- Beyond expectation: Deep joint mean and quantile regression for spatiotemporal problems (2020)
- Gqformer: A multi-quantile generative transformer for time series forecasting (2022)
- ChannelNets: Compact and Efficient Convolutional Neural Networks via Channel-Wise Convolutions (2018)
- Adam: A Method for Stochastic Optimization (2015)
- Deep residual learning for image recognition (2016)
- Temporal fusion transformers for interpretable multi-horizon time series forecasting (2021)
- Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting (2023)
- Crossvit: Cross-attention multi-scale vision transformer for image classification (2021)
- Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding (2022)
- DcTr: Noise-robust point cloud completion by dual-channel transformer with cross-attention (2023)
- Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting (2022)
- TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis (2023) — link
- TimeXer: Empowering Transformers for Time Series Forecasting with Exogenous Variables (2024)
- Comparing predictive accuracy (2002)
- Winner’s Curse? On Pace, Progress, and Empirical Rigor (2018) — link
- A Troubling Analysis of Reproducibility and Progress in Recommender Systems Research (2021) — link
- Sequential predictive conformal inference for time series (2023)
- Deeplob: Deep convolutional neural networks for limit order books (2019)
- Variance Stabilizing Transformations for Electricity Spot Price Forecasting (2018) — link
- Bayesian hierarchical probabilistic forecasting of intraday electricity prices (2025)
- The Role of Weather Predictions in Electricity Price Forecasting Beyond the Day-Ahead Horizon (2023) — link
- PriceFM: Foundation Model for Probabilistic Electricity Price Forecasting (2026) — link
- Probabilistic Electricity Price Forecast With Optimal Prediction Interval (2024) — link
- A Spike-Resilient Temporal-Adaptive Neural Framework for Day-Ahead Electricity Price Interval Forecasting (2025) — link
- Deep Learning for Electricity Price Forecasting: A Review of Day-Ahead, Intraday, and Balancing Electricity Markets — → 2602.10071
- Trading behaviour on the continuous intraday market Elbas (2016) — link
- Forecasting day-ahead electricity prices: A comparison of time series and neural network models taking external regressors into account (2022) — link
- Deep non-crossing quantiles through the partial derivative (2022)
- Monotonicity regularization: Improved penalties and novel applications to disentangled representation learning and robust classification (2022)
- An intra-day electricity price forecasting based on a probabilistic transformer neural network architecture (2023)
- Probabilistic intraday electricity price forecasting using generative machine learning (2025) — link
2 of 78 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.