Paper: arXiv 2609.25617

Authors: Hengyi Yang, Sida Lin, Yiyan Qi, Yankai Chen, Haohan Zhang, Xianhua Peng, Jian Guo

Abstract

Stock price forecasting is a long-standing challenge in computational finance, driven by the inherent randomness of markets and complex temporal patterns. While recent deep-learning models have raised forecasting accuracy by jointly modeling inter-stock and temporal price dynamics, they conflate inter-stock relationships with intra-stock temporal dependencies and focus solely on the univariate objective of price movement. To address these limitations, we propose LiMT, a Hierarchical Multi-Task Learning framework that integrates liquidity-aware signals for stock price forecasting. LiMT employs a Market Regime Encoder (MRE) module that first extracts contemporaneous cross-stock dependencies, then models each stock’s temporal dynamics, yielding a unified latent state. Building on this latent state, we introduce a Liquidity-Driven Learning (LDL) module, a mixture-of-experts architecture that features cross-task gating mechanisms to jointly predict price movement, volatility, and trading volume. We further design an Adaptive Portfolio Optimization (APO) mechanism that converts multi-task forecasts into executable portfolio weights under transaction-cost and liquidity constraints. Extensive experiments on the CSI300 and CSI500 benchmarks show that LiMT performs best among strong neural and tree-based baselines across the reported metrics. In realistic CSI300 backtests, APO improves annualized return from 3.99% to 10.01% and Sharpe ratio from 1.22 to 1.86 over equal weighting, showing that the multi-task forecasts translate into deployable portfolio gains.

Complexity vs Empirical Score

  • Math Complexity: 7.5/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: This paper presents a novel hierarchical multi-task learning framework with a strong emphasis on integrating liquidity-aware signals, which is a significant contribution. The empirical rigor is high, with extensive experiments and realistic backtests on established benchmarks. The methodology is clearly articulated, and the provision of code enhances reproducibility.

Research Flowchart

  flowchart TD
    A["Research Goal: Improve Stock Price Forecasting"] --> B{"Problem: Conflated Inter/Intra-stock dynamics, Univariate objective"};
    B --> C["Methodology: Hierarchical Multi-Task Learning (LiMT) with Liquidity-Aware Signals"];
    C --> D{"LiMT Components:"};
    D --> D1["Market Regime Encoder (MRE): Cross-stock dependencies & Temporal dynamics"];
    D --> D2["Liquidity-Driven Learning (LDL): Mixture-of-experts for Price, Volatility, Volume"];
    D1 & D2 --> E["Adaptive Portfolio Optimization (APO): Converts forecasts to portfolio weights"];
    E --> F["Outcomes: Improved Return & Sharpe Ratio in Backtests"];
    F --> G["Key Findings: LiMT outperforms baselines; APO generates deployable gains"];

    style A fill:#D4EDDA,stroke:#28A745,stroke-width:2px;
    style B fill:#F8D7DA,stroke:#DC3545,stroke-width:2px;
    style C fill:#D1ECF1,stroke:#17A2B8,stroke-width:2px;
    style D fill:#E2E3E5,stroke:#6C757D,stroke-width:2px;
    style D1 fill:#E2E3E5,stroke:#6C757D,stroke-width:1px;
    style D2 fill:#E2E3E5,stroke:#6C757D,stroke-width:1px;
    style E fill:#FFF3CD,stroke:#FFC107,stroke-width:2px;
    style F fill:#D4EDDA,stroke:#28A745,stroke-width:2px;
    style G fill:#D4EDDA,stroke:#28A745,stroke-width:2px;

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 LiMT, a hierarchical multi‑task learning framework that integrates liquidity‑aware signals for stock price forecasting.

Source sentence
we propose LiMT, a Hierarchical Multi-Task Learning framework that integrates liquidity-aware signals for stock price forecasting.

Assumes 3

  • Transaction costs are 10 basis points per trade and each stock’s traded value is capped at 2% of its daily turnover. execution
    Source sentence
    per trade, and cap the traded value of each stock at 2% of its daily turnover.
  • A participation‑rate cap of 2% is imposed on portfolio turnover. execution
    Source sentence
    α = 2% is the participation-rate cap, Ai,t is realized turnover, and C is the portfolio capital.
  • Data are split into training (2008‑01‑01 to 2014‑12‑31), validation (2015‑01‑01 to 2016‑12‑31) and test (2017‑01‑01 to 2020‑08‑01). data
    Source sentence
    training/validation/test splits of 2008-01-01 to 2014-12-31, 2015-01-01 to 2016-12-31, and 2017-01-01 to 2020-08-01.

Data

Backtest sample: 2017 – 2020 · CSI300 Chinese A‑share stocks · equities · daily

  • CSI300 benchmark public
    2008-2020 · largest and most liquid Chinese A‑share stocks
    Source sentence
    We evaluate LiMT on the Chinese A-share market using CSI300 and CSI500 constituents. CSI300 covers the largest and most liquid stocks, while CSI500 extends evaluation to a broader mid-cap universe. The data span 2008–2020
  • CSI500 benchmark public
    2008-2020 · broader mid‑cap Chinese A‑share stocks
    Source sentence
    We evaluate LiMT on the Chinese A-share market using CSI300 and CSI500 constituents. CSI300 covers the largest and most liquid stocks, while CSI500 extends evaluation to a broader mid-cap universe. The data span 2008–2020

Recognised sources: Qlib datasets

Costs

Modeled Costs are part of the reported headline results.

10 basis points per trade, with each stock’s traded value capped at 2% of its daily turnover.

Source sentence
per trade, and cap the traded value of each stock at 2% of its daily turnover.

Finds 2

Reported performance (LiMT with Adaptive Portfolio Optimization (APO)): Sharpe 1.86 (benchmark 1.22) · annual return 10.0%

  • APO improves annualized return from 3.99% to 10.01% and Sharpe ratio from 1.22 to 1.86 compared with equal‑weighting. performance
    Source sentence
    Relative to equal weighting, annualized return rises from 3.99% to 10.01%, OSR from 32.13% to 38.00%, and Sharpe ratio from 1.22 to 1.86
  • LiMT ranks first on all four metrics across both CSI300 and CSI500, achieving 56.22% and 51.98% gains in excess return and information ratio on CSI300 over the strongest baseline. performance
    Source sentence
    LiMT ranks first on all four metrics across both datasets. The gains are especially pronounced on the portfolio-oriented metrics: on CSI300, excess annualized return and excess information ratio improve by 56.22% and 51.98% over the strongest baseline

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 37 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 (75)

  1. Accurate multivariate stock movement prediction via data-axis transformer with multi-level contexts (2021)
  2. Stock Ranking with Multi-Task Learning (2022) — link
  3. Stock market forecasting using a multi-task approach integrating long short-term memory and the random forest framework (2022)
  4. Multiple Stock Time Series Jointly Forecasting with Multi-Task Learning (2020) — link
  5. Trading Volume Alpha (2024) — link
  6. Master: Market-guided stock transformer for stock price forecasting (2024)
  7. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling (2018)
  8. Tabnet: Attentive interpretable tabular learning (2021)
and 67 more
  1. On the Properties of Neural Machine Translation: Encoder-Decoder Approaches (2014)
  2. Long Short-term Memory (1997)
  3. Graph Attention Networks (2018)
  4. Localformer: Mitigating over-globalising in transformers on graphs with localised training (2025)
  5. Stock Price Prediction via Discovering Multi-Frequency Trading Patterns (2017)
  6. Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting (2021)
  7. A Dual-Stage Attention-Based Recurrent Neural Network for Time Series Prediction (2017)
  8. Efficient integration of multi-order dynamics and internal dynamics in stock movement prediction (2023)
  9. Temporal relational ranking for stock prediction (2019)
  10. Decision support system for stock trading using multiple indicators decision tree (2014)
  11. Xgboost: A scalable tree boosting system (2016)
  12. Lightgbm: A highly efficient gradient boosting decision tree (2017)
  13. Doubleensemble: A new ensemble method based on sample reweighting and feature selection for financial data analysis (2020)
  14. Short and long term stock trend prediction using decision tree (2017)
  15. A Multiple Indicators Model for Volatility Using Intra-Daily Data (2006) — link
  16. Predictive Regressions for Macroeconomic Data (2014) — link
  17. Macroeconomic Factors and Stock Exchange Return: A Statistical Analysis (2023) — link
  18. Forecasting Intraday Volume: Comparison of Two Early Models (2017) — link
  19. Ensemble methods: foundations and algorithms (2012)
  20. Stock price prediction using LSTM, RNN and CNN-sliding window model (2017)
  21. Adarnn: Adaptive learning and forecasting of time series (2021)
  22. Stock price prediction via discovering multi-frequency trading patterns (2017)
  23. Modeling and Forecasting Stock Return Volatility Using a Random Level Shift Model (2010) — link
  24. Cross- Sectional Variation of Intraday Liquidity , Cross-Impact , and Their Effect on Portfolio Execution (2022) — link
  25. Forecasting Intraday Volume Distributions (2015) — link
  26. Predicting Intraday Trading Volume and Volume Percentages (2018) — link
  27. Volatility Forecasting with Machine Learning and Intraday Commonality (2022) — link
  28. Semantic frames to predict stock price movement (2013)
  29. Stock selection via spatiotemporal hypergraph attention network: A learning to rank approach (2021)
  30. Spatiotemporal hypergraph convolution network for stock movement forecasting (2020)
  31. Hierarchical Adaptive Temporal-Relational Modeling for Stock Trend Prediction. (2021)
  32. Adaptive long-short pattern transformer for stock investment selection (2022)
  33. Temporal and Heterogeneous Graph Neural Network for Financial Time Series Prediction — → 2305.08740
  34. Hist: A graph-based framework for stock trend forecasting via mining concept-oriented shared information (2021) — link
  35. Stock market’s price movement prediction with LSTM neural networks (2017)
  36. Stock movement prediction from tweets and historical prices (2018)
  37. Enhancing stock movement prediction with adversarial training (2018) — link
  38. Qlib: An ai-oriented quantitative investment platform (2020) — link
  39. Attention is all you need (2017)
  40. Long short-term memory (2012)
  41. Layer normalization (2016) — link
  42. Reformer: The efficient transformer (2020) — link
  43. Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting (2022)
  44. CatBoost: Unbiased Boosting with Categorical Features (2018)
  45. A Distance Measure for Classifying Arima Models (1990) — link
  46. Stock price prediction using attention-based multi-input LSTM (2018)
  47. CNNpred: CNN-based stock market prediction using a diverse set of variables (2019)
  48. A CNN-BiLSTM-AM method for stock price prediction (2021)
  49. A Na "\i ve SVM-KNN based stock market trend reversal analysis for Indian benchmark indices (2015)
  50. Modeling the stock relation with graph network for overnight stock movement prediction (2021)
  51. Induction of decision trees (1986)
  52. C4. 5: programs for machine learning (2014)
  53. Space-time mixing attention for video transformer (2021)
  54. Cspo: Cross-market synergistic stock price movement forecasting with pseudo-volatility optimization (2025)
  55. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers (2022)
  56. Review of graph construction and graph learning in stock price prediction (2022)
  57. Spatial-temporal transformer for dynamic scene graph generation (2021)
  58. Lead–lag detection and network clustering for multivariate time series with an application to the US equity market (2022)
  59. Spatial-temporal transformer networks for traffic flow forecasting (2020) — link
  60. Tensor-based learning for predicting stock movements (2015)
  61. Transformer-based capsule network for stock movement prediction (2019)
  62. Hierarchical Multi-Scale Gaussian Transformer for Stock Movement Prediction. (2020)
  63. A distance measure for classifying ARIMA models (1990)
  64. Graph attention networks (2017) — link
  65. Learning phrase representations using RNN encoder-decoder for statistical machine translation (2014) — link
  66. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling (2018) — link
  67. Memory-Enhanced Transformer for Representation Learning on Temporal Heterogeneous Graphs (2023)

1 of 75 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.