Paper: arXiv 2508.06914

Authors: Ying Peng, Yifan Zhang, Xin Wang

Abstract

In this paper, we mainly focus on the prediction of short-term average return directions in China’s high-frequency futures market. As minor fluctuations with limited amplitude and short duration are typically regarded as random noise, only price movements of sufficient magnitude qualify as statistically significant signals. Therefore data imbalance emerges as a key problem during predictive modeling. From the view of data distribution imbalance, we employee the mean-uncertainty logistic regression (mean-uncertainty LR) classification method under the sublinear expectation (SLE) framework, and further propose the mean-uncertainty support vector machines (mean-uncertainty SVM) method for the prediction. Corresponding investment strategies are developed based on the prediction results. For data selection, we utilize trading data and limit order book data of the top 15 liquid products among the most active contracts in China’s future market. Empirical results demonstrate that comparing with conventional LR-related and SVM-related imbalanced data classification methods, the two mean-uncertainty approaches yields significant advantages in both classification metrics and average returns per trade.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 6.0/10 (capped from 7.5: transaction costs not modelled; see penalty flags)
  • Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 7.5)

Why this score: The paper employs advanced nonlinear expectation theory (sublinear expectation) and develops novel mean-uncertainty LR/SVM methods with theoretical proofs, indicating high math complexity. Empirically, it uses real high-frequency futures data (top 15 products) for backtesting investment strategies and reports performance metrics against benchmarks, showing strong data/implementation rigor.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Predict Short-Term Avg.<br>Return Directions in<br>China's Futures Market"] --> B["Data & Problem<br>High-Freq. Futures & Limit Order Book Data<br>Data Imbalance Issue"]
  B --> C["Methodology<br>Mean-Uncertainty Logistic Regression<br>Under SLE Framework"]
  B --> D["Methodology<br>Mean-Uncertainty Support Vector Machines<br>Under SLE Framework"]
  C & D --> E["Investment Strategy<br>Based on Prediction Results"]
  E --> F["Findings<br>Superior Classification Metrics &<br>Higher Avg. Returns vs. Conventional Methods"]

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 mean‑uncertainty logistic regression and mean‑uncertainty support vector machine classifiers, and develops corresponding investment strategies for high‑frequency futures.

Source sentence
we employee the mean-uncertainty logistic regression (mean-uncertainty LR) classification method under the sublinear expectation (SLE) framework, and further propose the mean-uncertainty support vector machines (mean-uncertainty SVM) method for the prediction.

Assumes 5

  • Disturbance term δi follows δi = εi+Mi with Mi ~ M[µ,µ] and εi standard logistic. model
    Source sentence
    Assumption 1. Assume that the disturbance term δi satisfies δi = εi+Mi, where {εi, Mi}n i=1 is an independent and identically distributed (i.i.d.) sequence under the SLE. Mi follows the maximum distribution M[µ,µ], and εi is nonlinearly independent of Mi. In particular, εi follows the standard logistic distribution.
  • Minor fluctuations with limited amplitude and short duration are treated as market noise and not used as signals. market
    Source sentence
    Essentially, minor fluctuations with limited amplitude and short durations are treated as market noise under our framework, as they fail to capture authentic market dynamics.
  • Data from the five‑minute pre‑opening and pre‑closing periods are excluded to avoid abnormal noise. data
    Source sentence
    Data of the five‑minute pre‑opening and pre‑closing periods are excluded.
  • A rolling time window of 30 seconds is used, with the first 25 seconds for independent variables and the next 5 seconds for the dependent variable. execution
    Source sentence
    a rolling time window with a duration of 30 seconds is employed within each trading day, where the data of the first 25 seconds is used to compute the independent variables, and that of the subsequent 5 second are used to calculate the dependent variable.
Show 1 more
  • Only observations exceeding the 95th percentile (or below the 5th percentile) of short‑term average returns are labeled as the minority class. data
    Source sentence
    we construct classification datasets based on the 95th percentile threshold of short-term average returns for the two distinct classification tasks: for the "upward movement" versus "non‑upward" prediction task, observations exceeding this critical value are labeled "1" (minority class), while others receive "0" (major

Data

  • Top 15 futures products dataset · CTP-API order-book unknown
    2024-10-01 to 2024-10-31 · Gold, Copper, Tin, Nickel, Corn, Zinc, Bottle Chips, Rebar, Silver, Aluminum, Hogs, Cotton
    Source sentence
    The selected dataset consist of trading data and limit order book data which are obtained from the CTP-API and collected every 0.5 seconds. The time span ranges from October 1st 2024 to October 31st 2024, covering a total of 18 trading days.

Costs

Not modeled A trading or execution result is reported without costs.

Finds 4

  • Both mean‑uncertainty LR and mean‑uncertainty SVM outperform conventional sampling‑based classifiers in predictive performance. performance
    Source sentence
    Empirical results demonstrate that both the two mean-uncertainty methods outperform conventional sampling-based
  • The mean‑uncertainty methods achieve higher average returns per trade on 80% of futures products for upward/non‑upward predictions. performance
    Source sentence
    both methods surpass conventional sampling-based classifiers on 80% of futures products for "upward / non-upward" predictions
  • For downward/non‑downward predictions, mean‑uncertainty LR yields higher returns on 80% of products, while mean‑uncertainty SVM does so on 67% of products. performance
    Source sentence
    for "downward/non-downward" predictions, the mean-uncertainty LR achieves higher returns on 80% of products, and the mean-uncertainty SVM shows superior performance on 67% of…
  • Mean‑uncertainty SVM shows superior recall improvement over all three conventional SVM‑related methods, indicating better recognition of minority class samples. predictive-accuracy
    Source sentence
    This demonstrates that mean-uncertainty SVM is more effective in recognizing minority class samples.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 25 related papers). Coverage is partial, so this is neither replication nor refutation.

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 (30)

  1. How and When are High-Frequency Stock Returns Predictable? (2022)
  2. Short-term stock price trend prediction with imaging high frequency limit order book data (2024)
  3. Attention-based CNN–LSTM for high-frequency multiple cryptocurrency trend prediction (2024)
  4. Do high-frequency financial data help forecast oil prices? The MIDAS touch at work (2015)
  5. Stock market volatility forecasting: Do we need high-frequency data? (2021)
  6. Correcting the errors: Volatility forecast evaluation using high-frequency data and realized volatilities (2005)
  7. Modelling and forecasting high-frequency data with jumps based on a hybrid nonparametric regression and LSTM model (2024)
  8. SMOTE: synthetic minority over-sampling technique (2002)
and 22 more
  1. Addressing the curse of imbalanced training sets: one-sided selection (1997)
  2. Efficient hybrid oversampling and intelligent undersampling for imbalanced big data classification (2024)
  3. Nearest neighbors and density-based undersampling for imbalanced data classification with class overlap (2024)
  4. Undersampling method based on minority class density for imbalanced data (2024)
  5. Daily volume forecasting using high frequency predictors (2010)
  6. Cost-sensitive continuous ensemble kernel learning for imbalanced data streams with concept drift (2024)
  7. A hybrid multi-criteria meta-learner based classifier for imbalanced data (2024)
  8. Adaptive SV-Borderline SMOTE-SVM algorithm for imbalanced data classification (2024)
  9. Backward SDE and related g-expectation (1997)
  10. Filtration consistent nonlinear expectations and evaluations of contingent claims (2004)
  11. Nonlinear expectations and nonlinear markov chains (2005)
  12. G-expectation, G-Brownian motion and related stochastic calculus of It ^o type (2007)
  13. Stochastic maximum principle for stochastic recursive optimal control problem under volatility ambiguity (2016)
  14. Dynamic programming principle for stochastic recursive optimal control problem driven by a G-Brownian motion (2017)
  15. Nonlinear regression without iid assumption (2019)
  16. Linear regression under model uncertainty (2023)
  17. Improving value-at-risk prediction under model uncertainty (2023)
  18. Nonlinear Expectations and Stochastic Calculus under Uncertainty: with Robust CLT and G-Brownian Motion (2019)
  19. Imbalanced binary classification under distribution uncertainty (2023)
  20. Optimal unbiased estimation for maximal distribution (2021)
  21. Inferring trade direction from intraday data (1991)
  22. Foundations of the theory of probability: Second English Edition (2018)

0 of 30 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.