Paper: arXiv 2503.09988
Authors: Zijian Zhao, Xuming Zhang, Jiayu Wen, Mingwen Liu, Xiaoteng Ma
Abstract
In financial trading, return prediction is one of the foundation for a successful trading system. By the fast development of the deep learning in various areas such as graphical processing, natural language, it has also demonstrate significant edge in handling with financial data. While the success of the deep learning relies on huge amount of labeled sample, labeling each time/event as profitable or unprofitable, under the transaction cost, especially in the high-frequency trading world, suffers from serious label imbalance issue.In this paper, we adopts rigurious end-to-end deep learning framework with comprehensive label imbalance adjustment methods and succeed in predicting in high-frequency return in the Chinese future market. The code for our method is publicly available at https://github.com/RS2002/Label-Unbalance-in-High-Frequency-Trading .
Complexity vs Empirical Score
- Math Complexity: 3.0/10
- Empirical Rigor: 8.0/10
- Quadrant: Street Traders — empirical and applied, lower math complexity
Why this score: The paper focuses on practical deep learning methods and mentions public code and real market data (Chinese futures), indicating high empirical rigor, while the mathematical depth is limited to standard ML concepts like cost matrices and sampling techniques.
Research Flowchart
flowchart TD A["Research Goal: Predict HFT Returns<br>in Chinese Futures Market"] --> B["Data: Tick-level Futures Data"] B --> C["Problem: Severe Label Imbalance<br>Due to Transaction Costs"] C --> D["Method: End-to-End Deep Learning<br>with Label Imbalance Adjustment"] D --> E["Computational Process: Training<br>with Adjusted Loss Functions"] E --> F["Key Finding: High Prediction Accuracy<br>on Chinese Futures Market"] F --> G["Outcome: Public Code Release<br>on GitHub"]
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 an end‑to‑end deep learning framework with comprehensive label‑imbalance adjustment methods for high‑frequency return prediction in the Chinese futures market.
Source sentence
In this paper, we adopts rigurious end-to‑end deep learning framework with comprehensive label imbalance adjustment methods and succeed in predicting in high‑frequency return in the Chinese future market.
Assumes 4
- Labels are defined based on a fixed trading fee, with positive, negative, or neutral classes depending on whether the 1‑minute return exceeds the fee. data
Source sentence
yi = 1 if Ri > fee; -1 if Ri < -fee; 0 if |Ri| ≤ fee where: • Ri is the forward 1 minute return for instance i. • fee is the trading fee.
- The minority class (positive return) is assigned a higher misclassification cost than the majority class. model
Source sentence
The cost of misclassifying a sample to class 0 is 10, while the cost of misclassifying to class 1 is 5. The reason for the higher cost of misclassifying to class 0 is that our primary interest lies in the minority class, which we have defined as the positive class (1).
- Data are split chronologically (8:1:1) to prevent information leakage, rather than randomly. execution
Source sentence
The dataset is then divided into training, validation, and test sets in a 8:1:1 ratio. To prevent information leakage, we split the dataset chronologically rather than through random sampling.
- The forward 1‑minute return is the target horizon for classification. market
Source sentence
In particular, we consider the three‑class classification problem in the horizon 1 min, where yi = …
Data
- High‑frequency Chinese futures data (six varieties) market-prices unknown
Source sentence
Our data spans from May 4th, 2023, to May 29th, 2023, covering a total of 20 trading days of high‑frequency futures data, with a frequency of 0.5 seconds. The dataset includes six varieties: rebar (rb2310), silver (ag2308), fuel oil (fu2309), nickel (ni2306), tin (sn2306), and gold (au2308).
Costs
Unclear
Finds 4
- The dataset exhibits severe label imbalance: about 80% of samples belong to class 0, while each of the +1 and -1 classes accounts for roughly 10% of the data. statistical
Source sentence
we identified a significant label imbalance issue in the dataset, with approximately 80% of samples belonging to class 0, while classes +1 and -1 each account for only about 10%.
- Sensitive loss (Eq. 10) and loss weighting (Eq. 9) methods achieve better performance than resampling or focal loss, which sometimes underperform the benchmark with no imbalance handling. performance
Source sentence
For addressing the label imbalance problem, it appears that the sensitive loss (Eq. 10) and loss weighting (Eq. 9) methods yield better performance than others. In our task, both resampling and focal loss (Eq. 11) [20] methods sometimes performed worse than the benchmark, which employed no specific approach for the lab
- LSTM and Mamba models generally outperform MLP in most scenarios, likely due to better capture of temporal relationships. performance
Source sentence
The LSTM [11] and Mamaba [16] models generally outperform the MLP in most scenarios, likely because they are better suited to capture temporal relationships due to their structural design.
- Training on a single futures contract leads to overfitting, whereas using all six contracts improves model robustness. negative
Source sentence
Consequently, we utilized all six items to train the models and evaluated their performance across different items.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 49 related papers). Coverage is partial, so this is neither replication nor refutation.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (26)
- Learning from class-imbalanced data: Review of methods and applications (2017) — link
- An insight into classification with imbalanced data: Empirical results and current trends on using data intrinsic characteristics (2013) — link
- Study of the impact of resampling methods for contrast pattern based classifiers in imbalanced databases (2016)
- Performance of corporate bankruptcy prediction models on imbalanced dataset: The effect of sampling methods (2013)
- A review on ensembles for the class imbalance problem: bagging-, boosting-, and hybrid-based approaches (2011)
- Review of classification methods on unbalanced data sets (2021)
- Online cost-sensitive neural network classifiers for non-stationary and imbalanced data streams (2013)
- Covariate shift adaptation by importance weighted cross validation. (2007)
and 18 more
- An Introduction to Statistical Learning (2013)
- Finding the Missing Data: A BERT-inspired Approach Against Package Loss in Wireless Sensing (2024) — link
- Long short-term memory (1997)
- Mamba: Linear-time sequence modeling with selective state spaces (2023) — link
- Novel cost-sensitive approach to improve the multilayer perceptron performance on imbalanced data (2013)
- Focal Loss for Dense Object Detection (2017) — link
- Recurrent neural networks (2001)
- Learning phrase representations using RNN encoder-decoder for statistical machine translation (2014) — link
- Are transformers effective for time series forecasting? (2023)
- Attention is all you need (2017)
- Combining recurrent, convolutional, and continuous-time models with linear state space layers (2021)
- Efficiently modeling long sequences with structured state spaces (2021) — link
- Mining Limited Data Sufficiently: A BERT-inspired Approach for CSI Time Series Application in Wireless Communication and Sensing (2024) — link
- KNN-MMD: Cross Domain Wi-Fi Sensing Based on Local Distribution Alignment (2024) — link
- CrossFi: A Cross Domain Wi-Fi Sensing Framework Based on Siamese Network (2024) — link
- Power Consumption Modeling of 5G Multi-Carrier Base Stations: A Machine Learning Approach (2023)
- Adversarial-MidiBERT: Symbolic Music Understanding Model Based on Unbias Pre-training and Mask Fine-tuning (2024) — link
- An overview of domain-specific foundation model: key technologies, applications and challenges (2024) — link
0 of 26 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.