Paper: arXiv 2408.08483
Authors: Zhoufan Zhu, Ke Zhu
Abstract
We open up the “black-box” to identify the predictive general price patterns in price chart images via the deep learning image analysis techniques. Our identified price patterns lead to the construction of image-induced importance (triple-I) weights, which are applied to weighted moving average the existing price trend trading signals according to their level of importance in predicting price movements. From an extensive empirical analysis on the Chinese stock market, we show that the triple-I weighting scheme can significantly enhance the price trend trading signals for proposing portfolios, with a thoughtful robustness study in terms of network specifications, image structures, and stock sizes. Moreover, we demonstrate that the triple-I weighting scheme is able to propose long-term portfolios from a time-scale transfer learning, enhance the news-based trading strategies through a non-technical transfer learning, and increase the overall strength of numerous trading rules for portfolio selection.
Complexity vs Empirical Score
- Math Complexity: 6.5/10
- Empirical Rigor: 6.0/10 (capped from 8.0: transaction costs not modelled; reported Sharpe ratio of 3.74 is implausibly high for daily or lower-frequency data; see penalty flags)
- Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 8.0)
Why this score: The paper employs advanced deep learning techniques (ResNet, Grad-CAM) and transfer learning theory, but is heavily grounded in extensive backtesting, dataset availability, and specific portfolio metrics (Sharpe ratios, turnover).
Research Flowchart
flowchart TD A["Research Goal<br>Identify predictive price patterns<br>from chart images"] --> B["Data Input<br>Chinese stock market data &<br>Price chart images"] B --> C["Methodology<br>Deep Learning Image Analysis"] C --> D["Process<br>Extract patterns &<br>Generate Image-Induced Importance Triple-I Weights"] D --> E["Application<br>Weighted Moving Average<br>on trading signals"] E --> F["Outcomes<br>Enhanced Trend Trading Strategies<br>& Portfolio Construction"] F --> G["Findings<br>Significant enhancement demonstrated<br>via transfer learning & robustness study"]
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 novel weighted moving average method that uses image‑induced importance (triple‑I) weights to enhance existing price trend trading signals.
Source sentence
we propose a novel weighted moving average method, aiming to enhance the performance of existing price trend trading strategies. Our method is based on the image-induced importance (III, or so-called triple-I) weights.
Assumes 4
- Analysis is performed on one‑day and one‑week holding periods (z = 1 and 5). execution
Source sentence
we focus our analysis on the performance of one-day and one-week (i.e., z = 1 and 5) holding periods.
- Excess return and Sharpe ratio are computed using the Treasury bill rate as the risk‑free rate. model
Source sentence
excess return and Sharpe ratio are computed using the Treasury bill 25 rate as a proxy for the risk-free return.
- First day closing price is normalized to one and all prices are scaled by daily returns. data
Source sentence
we normalize the first day closing price to one and calculate each subsequent daily closing price using the corresponding daily return in each image.
- Only the largest 1,800 A‑share stocks by market capitalization are used to avoid small‑cap anomalies. data
Source sentence
we primarily focus on the largest 1, 800 A-share stocks by market capitalization.
Data
Chinese stock market · equities · daily
- Wind A‑share daily stock data market-prices commercial
Source sentence
We consider daily stock data from the Wind database (https://www.wind.com.cn) for A-share stocks listed on the Shanghai and Shenzhen Stock Exchanges in China
Wind Yahoo Finance
Costs
Not modeled
Finds 4
Sharpe 3.74
- The RN34‑D5/R1 method achieves the highest Sharpe ratio, with at least 219% higher Sharpe and 27% lower turnover than the original method for MOM, STR, and WSTR classes. performance
Source sentence
in the classes of MOM, STR, and WSTR, the RN34-D5/R1 method demonstrates at least 219% higher values of Sharpe ratio and 27% lower values of turnover than the original method
- For one‑week horizons, the RN34‑D20/R5 method outperforms the original method, delivering 71%, 65%, 25% and 11% higher Sharpe ratios in MOM, STR, WSTR and TREND classes respectively. performance
Source sentence
the RN34-D20/R5 method performs the best, with 71%, 65%, 25%, and 11% higher values of Sharpe ratio than the original method in the classes of MOM, STR, WSTR, and TREND, respectively
- The ResNet “trader” achieves the highest Sharpe ratio of 3.74 with the RN34‑D20/R5 method among all traders for one‑week portfolios. performance
Source sentence
... as evidenced by the highest Sharpe ratio value of 3.74 achieved by the RN34-D20/R5 method
- Monthly turnover of one‑day long‑short portfolios selected by the ResNet trader exceeds 3200%, indicating high trading costs compared to traditional signals. performance
Source sentence
The monthly turnover values of the one-day long-short portfolios, selected by the ResNet “trader”, are in excess of 3200%, whereas that based on the MOM signal is only 281%.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 16 related papers). Coverage is partial, so this is neither replication nor refutation.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (50)
- (Re-)Imag(in)ing price trends (2023)
- Dividend yields and expected stock returns (1988)
- Market expectations in the cross-Section of present values (2013)
- Data-Snooping, technical trading rule performance, and the bootstrap (1999)
- The Extreme Value Method for Estimating the Variance of the Rate of Return (1980)
- Learning long-term dependencies with gradient descent is difficult (1994)
- Deep residual learning for image recognition (2016)
- Smooth grad-cam++: An enhanced inference level visualization technique for deep convolutional neural network models (2019) — link
and 42 more
- TorchCAM: class activation explorer (2020)
- Grad-cam: Visual explanations from deep networks via gradient-based localization (2017)
- Learning deep features for discriminative localization (2016)
- Score-CAM: Score-weighted visual explanations for convolutional neural networks (2020)
- Adam: A method for stochastic optimization (2015)
- Grad-CAM++: Generalized gradient-based visual explanations for deep convolutional networks (2018) — link
- A trend factor: Any economic gains from using information over investment horizons? (2016) — link
- 101 formulaic alphas (2016)
- Autoencoder asset pricing models (2021)
- Empirical asset pricing via machine learning (2020)
- Capturing volatility from large price moves: generalized range theory and applications (2007)
- Returns to buying winners and selling losers: Implications for stock market efficiency (1993)
- Evidence of predictable behavior of security returns (1990)
- Psychology-based models of asset prices and trading volume (2018)
- Deep learning (2016)
- Foundations of technical analysis: Computational algorithms, statistical inference, and empirical implementation (2000)
- The Heretics of Finance: Conversations with Leading Practitioners of Technical Analysis (2010) — link
- On technical analysis (1989)
- Trade and the revelation of information through prices and direct disclosure (1989)
- Market statistics and technical analysis: The role of volume (1994)
- A model of investor sentiment (1998)
- Technical analysis: An asset allocation perspective on the use of moving averages (2009)
- Forecasting the equity risk premium: the role of technical indicators (2014)
- Learning and predictability via technical analysis: Evidence from Bitcoin and stocks with hard-to-value fundamentals (2021)
- Charting by machines (2024)
- Technical trading revisited: False discoveries, persistence tests, and transaction Costs (2012)
- Forecasting stock prices with a feature fusion LSTM-CNN model using different representations of the same data (2019)
- Global Stock Market Prediction Based on Stock Chart Images Using Deep Q-Network (2019) — link
- CNNpred: CNN-based stock market prediction using a diverse set of variables (2019)
- Forecasting the overnight return direction of stock market index combining global market indices: A multiple-branch deep learning approach (2022) — link
- Effects of the validation set on stock returns forecasting (2020) — link
- When are contrarian profits due to stock market overreaction? (1990)
- Simple technical trading rules and the stochastic properties of stock returns (1992)
- Asset pricing via the conditional quantile variational autoencoder (2024)
- SmoothGrad: Removing noise by adding noise (2017)
- How much can machines learn finance from chinese text data? (2024) — link
- The adaptive lasso and its oracle properties (2006) — link
- Attention is All you Need (2017)
- An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (2021) — link
- Densely Connected Convolutional Networks (2017)
- Going Deeper With Convolutions (2015)
- A ConvNet for the 2020s (2022)
0 of 50 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.