Paper: arXiv 2409.14157

Authors: Kyungsub Lee

Abstract

This study explores the prediction of high-frequency price changes using deep learning models. Although state-of-the-art methods perform well, their complexity impedes the understanding of successful predictions. We found that an inadequately defined target price process may render predictions meaningless by incorporating past information. The commonly used three-class problem in asset price prediction can generally be divided into volatility and directional prediction. When relying solely on the price process, directional prediction performance is not substantial. However, volume imbalance improves directional prediction performance.

Complexity vs Empirical Score

  • Math Complexity: 4.5/10
  • Empirical Rigor: 7.0/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper employs standard statistical definitions and neural network architectures without deep theoretical derivations, but conducts a thorough empirical analysis using a large, recent dataset (AAPL Nasdaq ITCH) with detailed daily evaluation, cross-validation, and comparisons against naive baselines.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Predict High-Frequency Price Changes"] --> B["Methodology:<br>Deep Learning on Limit Order Book Data"]
  B --> C{"Key Inputs<br>Price Process vs. Volume Imbalance"}
  C --> D["Computational Process:<br>Three-Class Classification<br>Directional vs. Volatility Prediction"]
  D --> E{"Analysis of<br>Predictability Drivers"}
  E --> F["Outcome 1:<br>Price Process Alone<br>Lacks Directional Signal"]
  E --> G["Outcome 2:<br>Volume Imbalance<br>Significantly Improves Prediction"]

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 a deep learning model that predicts high‑frequency price changes using only Level 1 order‑book data, achieving comparable performance to models using the full LOB.

Source sentence
Using only Level 1 data produced results comparable to those of the entire LOB dataset.

Assumes 5

  • Analysis is restricted to standard trading hours (09:30‑16:00). data
    Source sentence
    We use the data only from standard trading hours, 09:30 to 16:00, which contain approximately…
  • The target variable is the modified return r_{k,k'}(t) defined without future information for r_{1,k}(t). model
    Source sentence
    The target variable is r1,k(t) = fk(t) −p1(t) / p1(t) which calculates the modified return based on a current price p1(t) as in Ntakaris et al. (2018).
  • Classes are balanced by choosing a constant α for the three‑class problem. model
    Source sentence
    Instances are classified as UP : y > α, DOWN : y < −α, STABLE : |y| ≤α for some constant α chosen to balance the class distribution.
  • Model training uses the preceding 20 days of data for each daily evaluation. execution
    Source sentence
    Model evaluation occurs on a daily basis throughout 2022. For each evaluation, the model is trained using data from the preceding 20 days.
Show 1 more
  • Volatility and directional predictability are evaluated separately using the STABLE/DIVERGE split. statistical
    Source sentence
    Volatility accuracy is the correct STABLE/DIVERGE prediction rate, while directional accuracy is the UP/DOWN hit ratio within correctly identified DIVERGE cases.

Data

  • AAPL 2022 LOB data · Nasdaq TotalView ITCH order-book commercial
    2022 · Apple Inc. (AAPL) stock
    Source sentence
    Our analysis uses highly active trading data for AAPL (Apple Inc.) stocks in 2022 from Nasdaq TotalView ITCH. These data provide detailed order book information, displaying the top 10 bid and ask price levels, share volume at each level, and timestamps for transactions.

Recognised sources: Nasdaq TotalView-ITCH

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • Overall classification accuracy using Level 1 data is 53.6%, matching the full‑LOB model. predictive-accuracy
    Source sentence
    When combined, the overall accuracy is 53.6%.
  • Volatility prediction accuracy reaches 69.4% on Level 1 data. predictive-accuracy
    Source sentence
    DIVERGE ... The average accuracy of the volatility prediction is 69.4% as in the first column in Table 7.
  • Directional prediction accuracy for DIVERGE cases is 71.1% using Level 1 data. predictive-accuracy
    Source sentence
    For DIVERGE, the procedure subsequently classifies UP or DOWN, achieving an average directional prediction accuracy of 71.1%.
  • Using only price information yields a directional accuracy of about 50%, supporting the efficient market hypothesis. predictive-accuracy
    Source sentence
    An interesting observation is that predictions based solely on price processes result in a consistent 50% directional accuracy, see Figure 3. Relying on price processes alone cannot forecast future price directions underscoring the efficient market hypothesis (Fama, 1970).
  • Incorporating volume imbalance markedly improves directional prediction performance. performance
    Source sentence
    The same analysis on AMZN, NVDA, and MSFT in 2022 yields similar results.

Checked by others

Supported by

  • Binary Tree Option Pricing Under Market Microstructure Effects: A Random Forest Approach (2025)
    On the question: Does order flow (volume) imbalance improve directional price prediction?
    Compare the findings

    This paper: Incorporating volume imbalance markedly improves directional prediction performance.

    That paper: Order flow imbalance is the most important predictor, contributing 43.2% to feature importance.

    Both find imbalance boosts prediction; A shows marked improvement, B assigns 43.2% feature importance.

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 (26)

  1. Deeplob: D eep convolutional neural networks for limit order books (2019)
  2. Benchmark dataset for mid-price forecasting of limit order book data with machine learning methods (2018)
  3. Forecasting stock prices from the limit order book using convolutional neural networks (2017)
  4. Efficient capital markets: A review of theory and empirical work (1970)
  5. Order imbalance and individual stock returns: T heory and evidence (2004)
  6. Deep learning with long short-term memory networks for financial market predictions (2018)
  7. Going Deeper with Convolutions (2014) — link
  8. Transformers for limit order books (2020) — link
and 18 more
  1. How informative is the order book beyond the best levels? Machine learning perspective (2022) — link
  2. Trading volume, bid-ask spread, and price volatility in futures markets (2000)
  3. Bid-ask spreads and volatility in the foreign exchange market: A n empirical analysis (1994)
  4. Transactions, volume, and volatility (1994)
  5. AE-ACG : A novel deep learning-based method for stock price movement prediction (2023)
  6. Forecasting directional movements of stock prices for intraday trading using LSTM and random forests (2022)
  7. A novel deep learning framework: Prediction and analysis of financial time series using CEEMD and LSTM (2020)
  8. Volatility Spillovers and Contagion During Major Crises: An Early Warning Approach Based on a Deep Learning Model (2023)
  9. Market volatility prediction and the efficiency of the S&P 100 index option market (1992)
  10. A comprehensive look at financial volatility prediction by economic variables (2012)
  11. Stock volatility prediction by hybrid neural network (2019)
  12. Stock returns, order imbalances, and commonality: E vidence on individual, institutional, and proprietary investors in China (2009)
  13. Recurrent neural network based parameter estimation of Hawkes model on high-frequency financial data — → 2304.11883
  14. Financial forecasting with machine learning: price vs return (2021)
  15. Predicting bitcoin price movements using sentiment analysis: a machine learning approach (2022)
  16. Forecasting cryptocurrencies’ price with the financial stress index: a graph neural network prediction strategy (2024)
  17. Reliability study of stock index forecasting in volatile and trending cities using public sentiment——based on word2 V ec and LSTM models (2023)
  18. A hybrid deep learning model for Bitcoin price prediction: data decomposition and feature selection (2023)

1 of 26 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.