Paper: arXiv 2411.13555
Authors: Junjie Guo
Abstract
This paper provides an empirical study explores the application of deep learning algorithms-Multilayer Perceptron (MLP), Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and Transformer-in constructing long-short stock portfolios. Two datasets comprising randomly selected stocks from the S&P500 and NASDAQ indices, each spanning a decade of daily data, are utilized. The models predict daily stock returns based on historical features such as past returns,Relative Strength Index (RSI), trading volume, and volatility. Portfolios are dynamically adjusted by longing stocks with positive predicted returns and shorting those with negative predictions, with equal asset weights. Performance is evaluated over a two-year testing period, focusing on return, Sharpe ratio, and maximum drawdown metrics. The results demonstrate the efficacy of deep learning models in enhancing long-short stock portfolio performance.
Complexity vs Empirical Score
- Math Complexity: 3.5/10
- Empirical Rigor: 6.0/10 (capped from 7.0: transaction costs not modelled; see penalty flags)
- Quadrant: Street Traders — empirical and applied, lower math complexity (effective empirical rigor capped from 7.0)
Why this score: The paper uses established deep learning architectures (MLP, CNN, LSTM, Transformer) applied to financial time-series prediction, requiring some advanced math for model architecture but no novel theoretical derivations. Its empirical rigor is high due to a detailed backtesting framework spanning 8 years of training and 2 years of testing, using specific metrics (Sharpe, MDD) on real market data from S&P500 and NASDAQ.
Research Flowchart
flowchart TD A["Research Goal<br>Assess Deep Learning Models for<br>Long-Short Stock Portfolios"] --> B["Data Preparation<br>S&P 500 & NASDAQ Decade-Long Data"] B --> C["Feature Engineering<br>RSI, Volume, Volatility, Past Returns"] C --> D["Model Training<br>MLP, CNN, LSTM, Transformers"] D --> E["Portfolio Construction<br>Long Positive Predictions, Short Negative"] E --> F["Performance Evaluation<br>Test Period: 2 Years"] F --> G["Outcomes<br>Deep Learning Enhances<br>Portfolio Performance (Sharpe, Returns)"]
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 applying four deep learning models (MLP, CNN, LSTM, Transformer) to construct long‑short stock portfolios using daily return predictions.
Source sentence
This paper provides an empirical study explores the application of deep learning algorithms—Multilayer Perceptron (MLP), Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and Transformer—in constructing long-short stock portfolios.
Assumes 5
- Transaction costs are negligible. execution
Source sentence
We assume transaction costs are negligible.
- The portfolio is fully invested each day. execution
Source sentence
Assume we fully invest the portfolio.
- All stocks receive equal weight in the portfolio. execution
Source sentence
The weights of investing in each stock are equally assigned.
- Predictions are made for the next‑day return and trades are executed based on those predictions. market
Source sentence
At each day of test set, we predict the next day’s return. We long those stocks with positive return and short those stocks with negative return.
Show 1 more
- Ten randomly selected stocks from each index represent the market segment. data
Source sentence
For each group, 10 stocks were randomly chosen, ensuring each had 10 years of continuous daily stock price data
Data
2022-10 – 2022-10 · S&P 500 stocks · equities · daily
- S&P 500 selected stocks market-prices public
Source sentence
... 10 stocks were randomly chosen, ensuring each had 10 years of continuous daily stock price data from October 1, 2014, to October 25, 2024.
- NASDAQ selected stocks market-prices public
Source sentence
... another from the NASDAQ index. For each group, 10 stocks were randomly chosen, ensuring each had 10 years of continuous daily stock price data from October 1, 2014, to October 25, 2024.
Costs
Not modeled
Source sentence
We assume transaction costs are negligible.
Finds 5
Sharpe 2.27 · total return 79.2% · max drawdown 13.7%
- On the S&P 500, the Transformer model achieved the highest total return (0.792059), highest Sharpe ratio (2.270424), and lowest max drawdown (0.137142) among all models. performance
Source sentence
Transformer 0.792059 2.270424 0.137142
- On the NASDAQ, the LSTM model slightly outperformed the Transformer in total return (0.609962 vs 0.605086), Sharpe ratio (1.351239 vs 1.305926), and had the lowest max drawdown (0.155910). performance
Source sentence
LSTM 0.609962 1.351239 0.155910
- Returns of stocks in both groups increased steadily over time, indicating stable performance of the deep learning models. performance
Source sentence
From observation, it is noticed that the returns of stocks in both groups increased steadily over time though with some fluctuations at certain period.
- CNN and MLP models generally underperformed compared to Transformer and LSTM on both indices, with lower returns and higher drawdowns. performance
Source sentence
The CNN and MLP models generally underperformed compared to the Transformer and LSTM on both indices, with lower returns and higher drawdowns.
- The study concludes that LSTM and Transformer architectures demonstrate superior capability in predicting daily stock returns compared to traditional models. predictive-accuracy
Source sentence
Our findings indicate that deep learning models, particularly LSTM and Transformer architectures, demonstrate a superior capability in predicting daily stock returns
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 25 related papers). Coverage is partial, so this is neither replication nor refutation.
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.