Simulation of Human Stock Traders' Chart Analysis Methods Using Deep Learning Models (Apr 8, 2024)

created At: 3/17/2025

Neutral

This analysis was written from a neutral perspective. We advise you to always make careful and well-informed investment decisions.

NONE

No Relevant Stock

Fact

Research Objective -Simulating human traders’ chart analysis using deep learning. -Leveraging market inefficiencies to optimize investment strategies. Methodology -Chart pattern learning: Using a ResNet-based deep learning model. -Long-term data analysis: Incorporating 600 days of price data. -Binarized return prediction: Forecasting whether prices will rise/fall by 10% or move sideways. -ResNet architecture optimization: Tailoring it for time-series data. -Softmax probability distribution: Determining the best trade entry points.

Opinion

This study demonstrates that deep learning-based chart analysis can generate meaningful excess returns. The model performed exceptionally well in the Korean market, reinforcing the idea that chart patterns are more useful in less efficient markets. In contrast, the lower performance in the U.S. market suggests that technical analysis alone may not be sufficient in highly efficient markets. However, fine-tuning Softmax logit values could improve strategy performance. Ultimately, this study highlights that deep learning-based chart analysis can be a viable investment approach in inefficient markets, and Softmax logit optimization is a key factor for success.

Core Sell Point

Deep learning-based chart analysis can yield high excess returns in the Korean market and, with strategy optimization, can enhance performance in the U.S. market as well.

This study aims to simulate the methods of human stock traders based on chart analysis using deep learning models. Despite the efficient market hypothesis, inefficiencies exist in the stock market, and deep learning can serve as a valuable tool to capitalize on these inefficiencies and outperform market returns.

1. Core Concept

Utilization of Chart Patterns:
Professional technical analysts identify specific patterns in historical price charts to predict future price movements. This study assumes that a deep learning model can learn and utilize such chart patterns.
Long-Term Data Usage:
The model analyzes stock price data from the past 600 days to capture long-term trends.
Binarized Returns:
The model predicts whether stock prices will rise or fall by 10% or 20% within a specified period (D days). Binarization reduces sensitivity to minor price fluctuations, simplifying model training.

2. Data Preprocessing

OHLCV Data Collection:
Daily Open, High, Low, Close, and Volume (OHLCV) data from South Korea (KOSPI, KOSDAQ) and the U.S. (NYSE, NASDAQ, AMEX) markets are collected.
Log Price Transformation:
To prevent sharp gradient explosions, logarithmic price data (base 10) is used instead of daily returns.
Label Assignment:
The model classifies stock price movements into three categories:
1. 10% (or 20%) Rise: If the highest price within D days exceeds the current closing price by 10% (or 20%).
2. 10% (or 20%) Fall: If the lowest price within D days is at least 10% (or 20%) below the current closing price.
3. Sideways Movement: If neither threshold is met.

3. Model Architecture

ResNet-Based Model:
The model employs skip connections from ResNet to effectively capture long-term trends. A modified ResNet architecture optimized for time-series data, as proposed by Wang et al. (2017), is used.
Input:
A 5×600 tensor, where 5 represents OHLCV features and 600 corresponds to past days of stock data.
Layers:

Five ResNet blocks (each with convolutional layers of kernel sizes 7×1, 5×1, and 3×1).
Batch normalization and ReLU activation functions applied.
Global Average Pooling (GAP) layer to extract key feature representations.
Fully connected (FC) layer followed by Softmax activation to output probabilities for price movement categories (10% rise, 10% fall, sideways movement).

4. Model Training

Loss Function: Binary Cross Entropy (BCE).
Optimization Algorithm: Not specified.
Epochs: 50.
Batch Size: Not specified.
Precision: 16-bit floating-point precision.
Training Duration: 1.5 days for Korean stock data, 5 days for U.S. stock data

5. Key Experimental Results

South Korea Market Performance:

Annualized return: 75.36%
Sharpe ratio: 1.57
Outperformed simple market investments and achieved significantly higher risk-adjusted returns.

U.S. Market Performance:

Annualized return: 27.17%
Sharpe ratio: 0.61
Outperformed NASDAQ, S&P 500, and Dow Jones benchmarks but underperformed AMEX.
Still demonstrated superior performance compared to major indices.

Impact of Softmax Logit Thresholds:

Adjusting the Softmax logit threshold significantly influenced backtesting results.
Optimized threshold selection enhanced accuracy and reduced trading risks.

These findings suggest that deep learning-based chart analysis can be an effective strategy for outperforming market returns. The model's superior performance in the Korean market suggests that chart patterns may be more useful in relatively inefficient markets.

In contrast, the model’s lower performance in the U.S. market may be due to higher market efficiency, making it harder to extract excess returns solely from chart analysis. However, the study highlights the potential for improving performance by fine-tuning Softmax logit thresholds, making this an essential aspect of trading strategy optimization.

[Compliance Note]

All posts by Sellsmart are for informational purposes only. Final investment decisions should be made with careful judgment and at the investor’s own risk.
The content of this post may be inaccurate, and any profits or losses resulting from trades are solely the responsibility of the investor.
Core16 may hold positions in the stocks mentioned in this post and may buy or sell them at any time.

Comments

Please leave a comment first

Home Articles Service Price About

Menu