Paper: arXiv 2504.17313

Authors: Zhuohang Zhu, Haodong Chen, Qiang Qu, Xiaoming Chen, Vera Chung

Abstract

Effective stock price forecasting (estimating future prices) and prediction (estimating future price changes) are pivotal for investors, regulatory agencies, and policymakers. These tasks enable informed decision-making, risk management, strategic planning, and superior portfolio returns. Despite their importance, forecasting and prediction are challenging due to the dynamic nature of stock price data, which exhibit significant temporal variations in distribution and statistical properties. Additionally, while both forecasting and prediction targets are derived from the same dataset, their statistical characteristics differ significantly. Forecasting targets typically follow a log-normal distribution, characterized by significant shifts in mean and variance over time, whereas prediction targets adhere to a normal distribution. Furthermore, although multi-step forecasting and prediction offer a broader perspective and richer information compared to single-step approaches, it is much more challenging due to factors such as cumulative errors and long-term temporal variance. As a result, many previous works have tackled either single-step stock price forecasting or prediction instead. To address these issues, we introduce a novel model, termed Patched Channel Integration Encoder (PCIE), to tackle both stock price forecasting and prediction. In this model, we utilize multiple stock channels that cover both historical prices and price changes, and design a novel tokenization method to effectively embed these channels in a cross-channel and temporally efficient manner. Specifically, the tokenization process involves univariate patching and temporal learning with a channel-mixing encoder to reduce cumulative errors. Comprehensive experiments validate that PCIE outperforms current state-of-the-art models in forecast and prediction tasks.

Complexity vs Empirical Score

  • Math Complexity: 7.0/10
  • Empirical Rigor: 6.5/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper introduces a novel neural network architecture (PCIE) with technical details on tokenization and multi-head attention, indicating substantial math/ML complexity, and it includes a comprehensive experimental section with multiple baselines and performance metrics, showing strong empirical backing.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Multi-Step Stock Price Forecast & Prediction"] --> B["Data Input<br>Multi-Channel Historical Prices"]
  B --> C["Methodology: PCIE Model<br>Univariate Patching & Tokenization"]
  C --> D["Computation<br>Channel-Mixing Encoder & Temporal Learning"]
  D --> E["Outcomes<br>Reduced Cumulative Error"]
  D --> F["Outcomes<br>Multi-Step Prediction"]
  E --> G["Validation<br>Outperforms SOTA Models"]
  F --> G

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 the Patched Channel Integration Encoder (PCIE), a novel model for multi-step stock price forecasting and prediction that uses tokenization of price and price change channels.

Source sentence
we introduce a novel model, termed Patched Channel Integration Encoder (PCIE), to tackle both stock price forecasting and prediction.

Assumes 6

  • Stock price data are non‑stationary with shifting means and variances over time. statistical
    Source sentence
    the non‑stationary distribution of stock prices, characterized by shifting means and variances over time, exacerbates the difficulty of accurate multi‑step forecasts.
  • Forecasting targets follow a log‑normal distribution while prediction targets follow a normal distribution. statistical
    Source sentence
    Forecasting targets typically follow a log‑normal distribution, characterized by significant shifts in mean and variance over time, whereas prediction targets adhere to a normal distribution.
  • Data are split into training, validation and testing sets with a 7:1:2 ratio. data
    Source sentence
    We split all data into training, validation and testing with a ratio of 7:1:2.
  • Daily sampling interval is used for all price series. data
    Source sentence
    We collect the historical price from 2016/01/04 to 2023/12/29 with a daily sampling interval.
Show 2 more
  • Instance normalization can mitigate distribution shift in stock price data. model
    Source sentence
    ... this technique coupled with our temporal learning block and patching techniques addresses the problem of distribution shift in stock price data.
  • The benchmark models are tuned according to the recommendations of their original papers for a fair comparison. model
    Source sentence
    We tuned all baseline models according to the recommendations of their corresponding papers for a fair comparison.

Data

  • US_71 · Yahoo Finance market-prices public
    2016/01/04 to 2023/12/29 · 71 high trade volume U.S. stocks across 9 major industries
    Source sentence
    For US_71, it contains the historical price of 71 high trade volume stocks from the U.S. stock market, representing the top 6-9 stocks in capital size and trade volume across the 9 major industries.
  • US_14L · Yahoo Finance market-prices public
    2005/01/04 to 2023/12/29 · 14 high market‑capitalization, high‑volume U.S. stocks
    Source sentence
    For US_14L, we selected 14 stocks characterized by high market capitalization and trading volume. We extended the time frame for data collection, capturing historical prices from 2005/01/04 to 2023/12/29, on a daily basis.

Recognised sources: Yahoo Finance

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • PCIE achieves the lowest MSE and MAE across all forecast horizons on both datasets, outperforming PatchTST, D‑Va, Autoformer, and Informer. performance
    Source sentence
    The results obtained from our experiments are detailed in Table 2. The Patched Channel Integration
  • The novel data preprocessing (combining price and price change) improves overall performance across all models by 0.6%–4%. performance
    Source sentence
    the performance improvement detailed in Table 3 demonstrates that our novel data preprocessing method improves the performance across different models.
  • Ablation study shows that tokenization reduces MSE (e.g., from 0.0838 to 0.0690 for US_71 forecast horizon 10). performance
    Source sentence
    Results summarized in Table 4 illustrate the critical role of tokenization in enhancing the model’s performance.
  • Models that ignore inter‑channel correlations (e.g., PatchTST) perform poorly on stock price forecasting. negative
    Source sentence
    models like PatchTST, which convert the input into univariate series, fail to perform well because they omit these essential inter‑channel correlations.
  • Methods assuming stationary distributions (e.g., Autoformer) underperform on forecasting tasks where data distribution evolves. negative
    Source sentence
    Autoformer employs a learned Fast Fourier Transform (FFT) attention mechanism, which assumes a stationary distribution. This assumption holds less effectively for forecasting tasks, where the data distribution evolves, resulting in subpar performance.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 32 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 1 paper in the archive

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 (40)

  1. Computing Machinery and Intelligence (1950)
  2. Learning Representations by Back-Propagating Errors (1986)
  3. Planning as Satisfiability (1992)
  4. Collaborative Plans for Complex Group Action (1996)
  5. The Entropy Formula for the R icci Flow and its Geometric Applications (2002)
  6. Causality (2009)
  7. Incorporating Corporation Relationship via Graph Convolutional Neural Networks for Stock Price Prediction (2018) — link
  8. Deep Learning for Stock Prediction Using Numerical and Textual Information (2016) — link
and 32 more
  1. A Hybrid ARIMA and Support Vector Machines Model in Stock Price Forecasting (2005)
  2. Stock Price Prediction Using the ARIMA Model (2014)
  3. A Machine Learning Model for Stock Market Prediction (2014) — link
  4. The Impact of Investor’s Attention on Global Stock Market: Statistical Review of Literature (2024)
  5. Diffusion Variational Autoencoder for Tackling Stochasticity in Multi-Step Regression Stock Price Prediction — → 2309.00073
  6. Stock Price Prediction via Discovering Multi-Frequency Trading Patterns (2017) — link
  7. Noise Trading and Stock Market Volatility (2007)
  8. Attention Is All You Need (2017)
  9. A Transformer-Based Framework for Multivariate Time Series Representation Learning (2021) — link
  10. A Time Series Is Worth 64 Words: Long-Term Forecasting with Transformers (2022)
  11. Are Transformers Effective for Time Series Forecasting? (2023)
  12. Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting (2021)
  13. Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting (2021)
  14. Stock Market Index Prediction Using Deep Transformer Model (2022)
  15. Reversible Instance Normalization for Accurate Time-Series Forecasting Against Distribution Shift (2021) — link
  16. Enhancing Stock Movement Prediction with Adversarial Training (2019) — link
  17. Stock Movement Prediction from Tweets and Historical Prices (2018) — link
  18. Deep Learning-Based Stock Price Prediction Using LSTM and Bi-Directional LSTM Model (2020) — link
  19. Stock Market Autoregressive Dynamics: A Multinational Comparative Study with Quantile Regression (2016)
  20. Listening to Chaotic Whispers: A Deep Learning Framework for News-Oriented Stock Trend Prediction (2017)
  21. Stock Movement Prediction via Attention-Aware Multi-Order Relation Graph Neural Network (2023) — link
  22. Deep Attentive Learning for Stock Movement Prediction From Social Media Text and Company Correlations (2020) — link
  23. Hierarchical Multi-Scale Gaussian Transformer for Stock Movement Prediction (2020) — link
  24. A Novel Differentiable Rank Learning Method Towards Stock Movement Quantile Forecasting (2023) — link
  25. Asset Price and Direction Prediction via Deep 2D Transformer and Convolutional Neural Networks (2022) — link
  26. Stock Movement Prediction Based on Bi-typed Hybrid-relational Market Knowledge Graph via Dual Attention Networks
  27. Time Horizon-Aware Modeling of Financial Texts for Stock Price Prediction (2021) — link
  28. Temporal Relational Ranking for Stock Prediction (2019) — link
  29. Learning Multiple Stock Trading Patterns with Temporal Routing Adaptor and Optimal Transport (2021)
  30. NumHTML: Numeric-Oriented Hierarchical Transformer Model for Multi-task Financial Forecasting (2022) — link
  31. The Longer-Horizon Predictability of German Stock Market Volatility (2006) — link
  32. Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing (2018) — link

1 of 40 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.