Paper: arXiv 2506.05764

Authors: Haochuan Wang

Abstract

Cryptocurrency price dynamics are driven largely by microstructural supply demand imbalances in the limit order book (LOB), yet the highly noisy nature of LOB data complicates the signal extraction process. Prior research has demonstrated that deep-learning architectures can yield promising predictive performance on pre-processed equity and futures LOB data, but they often treat model complexity as an unqualified virtue. In this paper, we aim to examine whether adding extra hidden layers or parameters to “blackbox ish” neural networks genuinely enhances short term price forecasting, or if gains are primarily attributable to data preprocessing and feature engineering. We benchmark a spectrum of models from interpretable baselines, logistic regression, XGBoost to deep architectures (DeepLOB, Conv1D+LSTM) on BTC/USDT LOB snapshots sampled at 100 ms to multi second intervals using publicly available Bybit data. We introduce two data filtering pipelines (Kalman, Savitzky Golay) and evaluate both binary (up/down) and ternary (up/flat/down) labeling schemes. Our analysis compares models on out of sample accuracy, latency, and robustness to noise. Results reveal that, with data preprocessing and hyperparameter tuning, simpler models can match and even exceed the performance of more complex networks, offering faster inference and greater interpretability.

Complexity vs Empirical Score

  • Math Complexity: 5.5/10
  • Empirical Rigor: 8.5/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs moderately advanced math in features and neural architectures but is anchored in a rigorous empirical setup with specific public data, multiple models, and direct performance comparisons. This balances theoretical depth with strong backtest-ready implementation, fitting the Holy Grail quadrant.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Assess if model complexity<br>or data preprocessing drives<br>LOB forecasting gains?"] --> B["Data Acquisition<br>Bybit BTC/USDT LOB snapshots<br>100ms - multi-second intervals"]
  B --> C{"Data Preprocessing Pipelines"}
  C --> C1["Kalman Filter"]
  C --> C2["Savitzky-Golay Filter"]
  C --> C3["Raw / Baseline"]
  C --> D["Labeling Schemes<br>Binary (Up/Down)<br>Ternary (Up/Flat/Down)"]
  D --> E["Model Training<br>Linear: Logistic Regression<br>Tree-based: XGBoost<br>Deep: DeepLOB, Conv1D+LSTM"]
  E --> F["Out-of-Sample Evaluation<br>Accuracy | Latency | Robustness"]
  F --> G["Key Findings"]
  G --> G1["Simpler models (LR/XGBoost)<br>match/exceed deep nets<br>with proper preprocessing"]
  G --> G2["Data filtering reduces noise<br>significantly more than<br>adding hidden layers"]
  G --> G3["Trade-off: Speed & Interpretability<br>vs. Complexity"]

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 that data preprocessing and hyperparameter tuning allow simpler models to match or exceed the performance of more complex neural networks, offering faster inference and greater interpretability.

Source sentence
Results reveal that, with data preprocessing and hyperparameter tuning, simpler models can match and even exceed the performance of more complex networks, offering faster inference and greater interpretability.

Assumes 5

  • All observable order‑book information is instantaneously incorporated into asset prices, leading to a 50:50 split in high‑frequency mid‑price moves. market
    Source sentence
    Under the Efficient Market Theorem, all observable order‑book information is instantaneously incorporated into asset prices, producing the empirically observed 50:50 split in high‑frequency mid‑price moves, and people will theoretically not be able to make profits, as the market has already incorporated past informatio
  • The latent state of each feature follows a simple random walk in the Kalman filter. model
    Source sentence
    The model assumes a simple random walk for the latent state, where Q and R denote the variances of the process and observation noise.
  • The empirical analysis uses a single trading day of Bybit BTC/USDT LOB snapshots sampled at 100 ms intervals. data
    Source sentence
    All LOB data in this study were obtained from Bybit’s public historical data portal
  • Epsilon for ternary classification is tuned to achieve roughly equal class frequencies and inverse‑frequency weighting is applied in the loss. statistical
    Source sentence
    we will tune ε to achieve roughly equal class frequencies when we do ternary split and apply inverse-frequency weighting in the loss so that rare “up” and “down” events incur proportionally greater penalty
Show 1 more
  • Low latency is crucial for reducing slippage in high‑frequency trading. execution
    Source sentence
    low latency—crucial for reducing slippage in high‑frequency trading.

Data

Backtest sample: BTC/USDT · crypto · intraday

  • Bybit BTC/USDT LOB snapshots · Bybit public historical data portal order-book public
    2025-01-30 · BTC/USDT
    Source sentence
    All LOB data in this study were obtained from Bybit’s public historical data portal

Recognised sources: Bybit FI-2010

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • Simpler models can match or exceed the performance of more complex networks when data preprocessing and hyperparameter tuning are applied. performance
    Source sentence
    Results reveal that, with data preprocessing and hyperparameter tuning, simpler models can match and even exceed the performance of more complex networks, offering faster inference and greater interpretability.
  • Simpler models such as XGBoost and logistic regression marginally outperform the more complex neural networks by 1–2%. performance
    Source sentence
    simpler models such as XGBoost and logistic regression marginally outperform our more complex neural networks by 1–2%.
  • Adding an extra CNN layer to a CNN + LSTM architecture did not yield the substantial accuracy gains expected. negative
    Source sentence
    CNN + LSTM based architecture did not yield the substantial accuracy gains we expected
  • All models consistently perform best after Savitzky–Golay smoothing, while Kalman filter results can be worse than raw data. performance
    Source sentence
    all models consistently perform best after Savitzky–Golay smoothing. Note that the Kalman filter results are sometimes worse than using the raw data itself.
  • Binary and ternary prediction accuracy ranges from 0.42 to 0.71 across different prediction horizons and LOB depths. predictive-accuracy
    Source sentence
    binary and ternary prediction accuracy ranges from 0.42 to 0.71—spanning 100ms predictions with 5-level LOB data to 1000ms predictions with 40-level LOB data.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 45 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 (24)

  1. TLOB: A Novel Transformer Model with Dual Attention for Stock Price Trend Prediction with Limit Order Book Data — → 2502.15757
  2. Limit Order Book Event Stream Prediction with Diffusion Model — → 2412.09631
  3. Modeling of limit order book data with ordered fuzzy numbers (2024) — link
  4. Deep limit order book forecasting: a microstructural guide — → 2403.09267
  5. A generative model of a limit order book using recurrent neural networks (2023) — link
  6. Forecasting stock market for an efficient portfolio by combining XGBoost and Hilbert-Huang​ transform (2023) — link
  7. Axial-LOB: High-Frequency Trading with Axial Attention — → 2212.01807
  8. The Price Impact of Generalized Order Flow Imbalance (2021) — link
and 16 more
  1. Order book mid-price movement inference by CatBoost classifier from convolutional feature maps (2021) — link
  2. How Robust are Limit Order Book Representations under Data Perturbation? (2021) — link
  3. Deep Learning modeling of Limit Order Book: a comparative perspective (2020) — link
  4. Using Deep Learning for price prediction by exploiting stationary limit order book features (2018) — link
  5. Multi-Level Order-Flow Imbalance in a Limit Order Book (2018) — link
  6. DeepLOB: Deep Convolutional Neural Networks for Limit Order Books (2018) — link
  7. Universal features of price formation in financial markets: perspectives from deep learning (2018) — link
  8. Forecasting Stock Prices from the Limit Order Book Using Convolutional Neural Networks (2017) — link
  9. Benchmark dataset for mid-price forecasting of limit order book data with machine learning methods (2017) — link
  10. Modelling high-frequency limit order book dynamics with support vector machines (2015) — link
  11. Price jump prediction in a limit order book (2012) — link
  12. Does Algorithmic Trading Improve Liquidity (2011) — link
  13. Order Imbalance, Liquidity, and Market Returns (2003) — link
  14. Smoothing and differentiation of data by simplified least square procedure. (1964) — link
  15. Optimizing Order Types in Indian Futures Market (2022) — link
  16. Ensemble Methods in Machine Learning (2007)

4 of 24 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.