Paper: arXiv 2502.17493
Authors: Ruoyu Guo, Haochen Qiu, Xuelun Hou
Abstract
Making consistently profitable financial decisions in a continuously evolving and volatile stock market has always been a difficult task. Professionals from different disciplines have developed foundational theories to anticipate price movement and evaluate securities such as the famed Capital Asset Pricing Model (CAPM). In recent years, the role of artificial intelligence (AI) in asset pricing has been growing. Although the black-box nature of deep learning models lacks interpretability, they have continued to solidify their position in the financial industry. We aim to further enhance AI’s potential and utility by introducing a return-weighted loss function that will drive top growth while providing the ML models a limited amount of information. Using only publicly accessible stock data (open/close/high/low, trading volume, sector information) and several technical indicators constructed from them, we propose an efficient daily trading system that detects top growth opportunities. Our best models achieve 61.73% annual return on daily rebalancing with an annualized Sharpe Ratio of 1.18 over 1340 testing days from 2019 to 2024, and 37.61% annual return with an annualized Sharpe Ratio of 0.97 over 1360 testing days from 2005 to 2010. The main drivers for success, especially independent of any domain knowledge, are the novel return-weighted loss function, the integration of categorical and continuous data, and the ML model architecture. We also demonstrate the superiority of our novel loss function over traditional loss functions via several performance metrics and statistical evidence.
Complexity vs Empirical Score
- Math Complexity: 3.5/10
- Empirical Rigor: 6.0/10 (capped from 7.5: transaction costs not modelled; see penalty flags)
- Quadrant: Street Traders — empirical and applied, lower math complexity (effective empirical rigor capped from 7.5)
Why this score: The mathematical complexity is moderate, primarily involving standard ML components like cross-entropy and embedding layers, without heavy advanced theory or derivations; the empirical rigor is high, as the paper presents detailed backtesting results over two distinct time periods (2019-2024 and 2005-2010) with specific performance metrics (annual return, Sharpe ratio) using publicly available data, demonstrating a practical, implementation-focused trading system.
Research Flowchart
flowchart TD A["Research Goal: <br>Develop High-Return Daily Stock Trading System"] --> B["Data Input: <br>OHLCV, Sector Info, Technical Indicators"] B --> C["Methodology: <br>Novel Return-Weighted Loss Function"] C --> D["Computational Process: <br>Deep Learning Model Training"] D --> E["Key Finding 1: <br>61.73% Annual Return (2019-2024)"] D --> F["Key Finding 2: <br>37.61% Annual Return (2005-2010)"] E --> G["Outcome: <br>Superior Performance vs Traditional Loss Functions"] 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 a machine‑learning based daily trading system that uses a novel return‑weighted loss function to capture top growth opportunities.
Source sentence
We propose a ML-based efficient trading system equipped with a novel return-weighted loss function that is specifically adept at capturing top growth opportunities.
Assumes 6
- All trades are executed at the market open with full capital invested at all times. execution
Source sentence
We assume all trades are conducted at market open and that all assets are invested at all times.
- The starting investment amount is normalized to 1 for performance calculation. statistical
Source sentence
we assume the starting investment amount is 1, and the final value is calculated by investing in the top 10 stocks the model recommends daily.
- Sector information is represented as a one‑hot vector in 12 dimensions. data
Source sentence
There are 11 sectors in the dataset, and each stock’s sector information is stored as a one-hot vector in 12 dimensions, where the extra dimension is for those without sector information in Yahoo Finance (such as ETFs).
- Micro‑cap stocks and low‑volume stocks are removed from the sample. data
Source sentence
We process stocks by removing microcap stocks and those with low dollar trading volume.
Show 2 more
- Features are calculated using opening prices because simulated trades occur at market open. model
Source sentence
Since our simulated trades all occur at market open, all features are calculated with the opening prices instead of closing prices by default.
- The trading strategy sells holdings not in the next‑day recommendation, holds overlapping stocks, and buys new recommendations in equal dollar amounts. execution
Source sentence
The strategy sells the current holdings that are not in the next-day recommendations, holds the current holdings that are in the next-day recommendations, and buys the remaining recommended stocks in equal dollar amounts.
Data
2019 – 2024 · US stocks · equities · daily
- Yahoo Finance daily stock data market-prices public
Source sentence
daily stock price (open, close, high, low) data, volume (number of shares traded) data, and sector information from Yahoo Finance using its Python API.
Costs
Not modeled
Finds 5
Sharpe 1.18 (benchmark 0.68) · annual return 61.7% · max drawdown 31.0%
- The best models achieve a 61.73% annual return with an annualized Sharpe Ratio of 1.18 over 1340 testing days (2019‑2024). performance
Source sentence
Our best models achieve 61.73% annual return on daily rebalancing with an annualized Sharpe Ratio of 1.18 over 1340 testing days from 2019 to 2024
- In the earlier period, the models achieve a 37.61% annual return with a Sharpe Ratio of 0.97 over 1360 testing days (2005‑2010). performance
Source sentence
and 37.61% annual return with an annualized Sharpe Ratio of 0.97 over 1360 testing days from 2005 to 2010
- Linear regression models using all 28 features in the 2019‑2024 period yield an annual return of about 40% and a top‑10 Sharpe Ratio near 0.85‑0.87. performance
Source sentence
... Annual Return 40.16% 38.17% 39.98% Top 10 SR 0.82 0.80 0.87
- The novel return‑weighted loss function outperforms traditional loss functions in both return and risk‑adjusted metrics, especially in the 2019‑2024 period. performance
Source sentence
In terms of return and risk-adjusted return (SR), our new loss function outperforms other traditional loss functions, especially in the later period of 2019-2024.
- The new loss function shows a smaller maximum drawdown and faster recovery time compared to baselines. performance
Source sentence
The new loss function also shows a smaller maximum drawdown and faster recovery time comparatively
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 27 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 (46)
- Attention is All you Need (2017) — link
- Empirical asset pricing via machine learning (2020)
- Gradient-based learning applied to document recognition (1998)
- U-net: Convolutional networks for biomedical image segmentation (2015)
- Capital asset prices: A theory of market equilibrium under conditions of risk (1964)
- Efficient capital markets (1970)
- Common risk factors in the returns on stocks and bonds (1993)
- On persistence in mutual fund performance (1997)
and 38 more
- A five-factor asset pricing model (2015)
- Glove: Global vectors for word representation (2014)
- Bert: Pre-training of deep bidirectional transformers for language understanding (2018) — link
- Profitability context and the cross-section of stock returns (2024)
- Time series analysis: forecasting and control (2015)
- CNNpred: CNN-based stock market prediction using a diverse set of variables (2019)
- Technical analysis and sentiment embeddings for market trend prediction (2019)
- Deep Learning techniques for stock market forecasting: Recent trends and challenges (2023)
- News-based intelligent prediction of financial markets using text mining and machine learning: A systematic literature review (2023) — link
- A novel trading system for the stock market using Deep Q-Network action and instance selection (2024)
- Stock price prediction using CNN and LSTM-based deep learning models (2020)
- Stock price prediction using CNN-BiLSTM-Attention model (2023)
- A mathematical theory of communication (1948)
- Technical Analysis Library in Python (2018) — link
- Deep transformer-based asset price and direction prediction (2024)
- Transformer-based attention network for stock movement prediction (2022)
- Factors affecting text mining based stock prediction: Text feature representations, machine learning models, and news platforms (2022)
- Stock trend prediction using deep learning approach on technical indicator and industrial specific information (2021)
- Open source cross-sectional asset pricing (2021)
- Deep learning using risk-reward function for stock market prediction (2018)
- Generalized loss-based cnn-bilstm for stock market prediction (2024)
- A CNN-LSTM-based model to forecast stock prices (2020)
- Short-term stock correlation forecasting based on CNN-BiLSTM enhanced by attention mechanism (2024)
- A Stock Prediction Model Based on CNN-BiLSTM and Multiple Attention Mechanisms (2023)
- Adam: A Method for Stochastic Optimization (2017) — link
- Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift (2015) — link
- Dropout: a simple way to prevent neural networks from overfitting (2014)
- Cryptocurrency portfolio management with deep reinforcement learning (2017) — link
- Predicting Stock Prices Based on Machine Learning to Build Self-adaptive Trading Strategy (2025) — link
- Machine learning in business and finance: a literature review and research opportunities (2024) — link
- A convolutional neural network based approach to financial time series prediction (2022) — link
- Financial Time Series Forecasting using CNN and Transformer — → 2304.04912
- Financial Time Series Forecasting – A Deep Learning Approach (2017) — link
- Deep residual learning for image recognition (2016)
- Enhancing neural network predictions with finetuned numeric embeddings for stock trend forecasting (2025)
- A comprehensive survey of loss functions and metrics in deep learning (2025)
- A comprehensive survey of loss functions in machine learning (2022)
- Focal loss for dense object detection (2017)
1 of 46 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.