Paper: arXiv 2508.14656

Authors: Yuqi Luan

Abstract

This study proposes a behaviorally-informed multi-factor stock selection framework that integrates short-cycle technical alpha signals with deep learning. We design a dual-task multilayer perceptron (MLP) that jointly predicts five-day future returns and directional price movements, thereby capturing nonlinear market behaviors such as volume-price divergence, momentum-driven herding, and bottom reversals. The model is trained on 40 carefully constructed factors derived from price-volume patterns and behavioral finance insights. Empirical evaluation demonstrates that the dual-task MLP achieves superior and stable performance across both predictive accuracy and economic relevance, as measured by information coefficient (IC), information ratio (IR), and portfolio backtesting results. Comparative experiments further show that deep learning methods outperform linear baselines by effectively capturing structural interactions between factors. This work highlights the potential of structure-aware deep learning in enhancing multi-factor modeling and provides a practical framework for short-horizon quantitative investment strategies.

Complexity vs Empirical Score

  • Math Complexity: 6.5/10
  • Empirical Rigor: 6.0/10 (capped from 7.0: transaction costs not modelled; see penalty flags)
  • Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 7.0)

Why this score: The paper presents advanced deep learning architectures with detailed mathematical formulations (MLP, CNN, loss functions), indicating high math complexity. It includes a comprehensive empirical evaluation using 40 factors, specific metrics (IC, IR), and portfolio backtesting results, making it backtest-ready with substantial data implementation.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Forecast Short-Term Equity Trends<br>using Behavioral Factors"] --> B["Data: 40 Price-Volume &<br>Behavioral Finance Factors"]
  B --> C["Model: Dual-Task Deep MLP"]
  C --> D["Task 1: Predict 5-Day<br>Future Returns"]
  C --> E["Task 2: Predict Directional<br>Price Movement"]
  D & E --> F["Outcome: Superior Performance<br>High IC/IR & Stable Returns"]
  F --> G["Conclusion:<br>Structure-Aware Deep Learning<br>Outperforms Linear Baselines"]

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 behaviorally-informed multi-factor stock selection framework using a dual‑task multilayer perceptron that jointly predicts five‑day returns and directional movements.

Source sentence
This study proposes a behaviorally-informed multi-factor stock selection framework that integrates short-cycle technical alpha signals with deep learning.

Assumes 5

  • A time‑based split is used, with data before 2023 for training and data from 2023 onward for validation. data
    Source sentence
    We use a time-based split: data before 2023 is used for training, and data from 2023 onward is reserved for validation.
  • Future 5‑day returns are clipped to the [5%, 95%] quantile range to suppress extreme outliers. data
    Source sentence
    The target variable future return 5 is clipped to the [5%, 95%] quantile range to suppress the impact of extreme outliers.
  • Features are standardized each day using cross‑sectional z‑score normalization to remove scale differences across assets. data
    Source sentence
    using cross-sectional z-score normalization per day to remove scale differences across assets.
  • Transaction costs, liquidity constraints, and turnover effects are not incorporated into the optimization process. execution
    Source sentence
    the current experimental design abstracts away from trading frictions, as transaction costs, liquidity constraints, and turnover
Show 1 more
  • The dual‑task architecture assumes that jointly learning return magnitude and direction improves generalization and signal extraction. model
    Source sentence
    The shared feature representation allows both heads to benefit from a more expressive latent structure, leading to better generalization and stronger signal extraction.

Data

Backtest sample: US equities · equities · daily

  • 40 technical signals derived from price‑volume patterns and behavioral finance insights market-prices unknown
    Source sentence
    The model is trained on 40 carefully constructed factors derived from price-volume patterns and behavioral finance insights.

Costs

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

Source sentence
the current experimental design abstracts away from trading frictions, as transaction costs, liquidity constraints, and turnover

Finds 5

Reported performance (dual-task MLP): Sharpe 1.61 · annual return 53.5%

  • The dual‑task MLP achieves the highest average daily IC of 0.0340, IC Std of 0.2097 and IR of 0.1621, outperforming CNN and SVM. performance
    Source sentence
    It achieves the highest average daily IC of 0.0340, with a relatively low IC Std of 0.2097, resulting in an IR of 0.1621, the strongest among the three models.
  • The MLP’s annualized Sharpe Ratio is 1.6075, compared with 1.1487 for CNN and 0.7709 for SVM. performance
    Source sentence
    MLP achieves the highest Sharpe Ratio (1.6075), followed by the CNN (1.1487), while the SVM lags significantly behind (0.7709).
  • Cumulative 5‑year return of the MLP long‑short strategy exceeds 800%, with smoother upward trend and minimal drawdowns. performance
    Source sentence
    The Top 5 cumulative return in 5 years for MLP model reaches above 800%.
  • During the sharp market downturn in January 2025, the MLP shows more resilience than CNN and SVM, which suffered severe drawdowns. performance
    Source sentence
    the CNN and SVM strategies suffered severe drawdowns, while the MLP model showed more resilience.
  • MLP outperforms CNN and SVM in both IC magnitude and consistency, indicating stronger and more stable predictive signals. predictive-accuracy
    Source sentence
    model significantly outperforms both CNN and SVM in terms of IC magnitude and consistency.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 48 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.