Convolutional
Neural Networks
for Investment Signal Generation

Exploring visual data types, research trends, and methodologies in financial machine learning

Financial Forecasting Deep Learning Computer Vision

99.3%

Maximum accuracy achieved with CNN candlestick analysis

Technical indicators used in S&P 500 prediction study

53.8%

Minimum accuracy range observed across studies

CNN-LSTM

Most common hybrid architecture for financial data

Primary Visual Data Types

CNNs primarily analyze candlestick charts, financial time-series plots, and technical indicator visualizations for investment signal generation

Candlestick Charts

Originating from 17th-century Japanese rice trading, candlestick charts provide rich visual representations of price movements (OHLC) with color-coding indicating market sentiment.

Key Finding:

Research achieved 99.3% accuracy using CNNs on Japanese candlestick patterns for Forex market prediction [146]

Rich informational content with body and wick patterns
Established practice in technical analysis
Effective for both classification and regression tasks

Financial Time-Series Plots

Line charts, bar charts, and specialized encodings like Gramian Angular Fields (GAF) transform numerical time-series data into visual patterns for CNN analysis.

Research Insight:

Image-based models generally outperform numerical input models for predicting stock movement over 5-day periods [167]

Encoding Methods:

• Gramian Angular Field (GAF)
• Markov Transition Field (MTF)
• Bar chart image conversion
• Line chart representations

Technical Indicator Plots

Visual representations of indicators like Moving Averages, RSI, MACD, and Bollinger Bands provide valuable signals for CNN pattern recognition.

Performance Boost:

Charts augmented with technical indicators generally achieve higher accuracy than baseline candlestick charts alone [153]

Indicator Types

• Momentum (RSI)
• Trend (MA, MACD)
• Volatility (Bollinger)
• Volume (OBV)

LSTM Priorities

• HLC3 [43]
• SMA5
• TEMA
• SMA50

Emerging and Niche Visual Data

Beyond traditional charts, researchers explore visualized financial statements, text-to-image conversions, and multi-modal integration

Visualized Financial Statements

Converting balance sheets and income statements into visual heatmaps or graphical representations for CNN pattern recognition.

Challenge: Designing effective visualizations that encode multi-dimensional financial data

Text-to-Image Conversion

Converting news articles, social media sentiment, and financial reports into visual plots processable by CNNs.

Application: Sentiment trend graphs, word cloud visualizations, news impact charts

Multi-modal Integration

Combining visual data from charts with numerical data, sentiment scores, and news embeddings for holistic analysis.

Promise: Most actively researched area for robust prediction models

Integrating fundamental analysis tools, including financial and political news, annual reports, companies' product lifecycles, or their financial horizon alongside chart data is more promising than relying solely on historical prices

— Research finding from [13]

Key Research Directions

Four primary research streams drive innovation in CNN-based financial forecasting

Direct Price Movement Prediction

CNNs are applied to predict future stock prices (regression) or price direction (classification) using visual and numerical inputs.

S&P 500 Prediction Results

0.014

MAE

0.008

MAPE

0.045

RMSE

Achieved by LSTM model using 88 technical indicators [43]

Trend Identification and Forecasting

CNNs classify market states (upward, downward, sideways) and predict future trend directions by recognizing visual chart patterns.

Upward

Downward

Sideways

Double Top

Rounded Bottom

Novel CNN Architectures and Hybrid Models

Researchers develop attention mechanisms, dilated CNNs, and hybrid models combining CNNs with LSTMs for enhanced performance.

CNN-LSTM Hybrid Architecture

graph TD A["Candlestick Chart Images"] --> B["CNN Feature Extraction"] C["OHLC Price Data"] --> D["LSTM Temporal Modeling"] B --> E["Feature Vector"] D --> E E --> F["Fully Connected Layers"] F --> G["Price/Trend Prediction"]

CNN with Attention

Enhanced performance in capturing volume data and color information

Improved accuracy over B&W charts

CNN-LSTM Hybrid

Most common architecture combining spatial and temporal modeling

Adjusted R²: 0.9968

Dilated CNN-LSTM

Wider receptive fields for capturing long-range dependencies

Enhanced temporal modeling

Comparative Studies of Visual Inputs

Key Finding: Pattern Annotations

Explicitly detected candlestick patterns (using YOLOv8) did not improve model performance over raw chart images alone, suggesting CNNs learn relevant features directly from pixel data [140]

Image-Based vs Numerical Models

Image-based models generally outperformed numerical input models for 5-day stock movement prediction, especially when enhanced with post-hoc calibration [167]

Performance Comparison

Raw Chart Images

95%

GAF Encoded Images

88%

Technical Indicator Augmented

98%

Data Preprocessing and Augmentation

Critical techniques for handling financial data characteristics and improving model robustness

Normalization & Standardization

Scaling pixel values to standard ranges ([0,1] or [-1,1]) ensures equal feature contribution and faster convergence.

Application: Essential for LSTM components in hybrid models

Handling Non-Stationarity

Addressing changing statistical properties in financial time-series through differencing, detrending, or robust architecture design.

Challenge: Market regime shifts and economic crises affect pattern persistence

Data Augmentation

Artificially increasing training dataset size through rotations, shifts, and noise addition to reduce overfitting.

Need: Datasets 2-3 orders of magnitude larger required for robust models

Performance Evaluation and Challenges

Assessing CNN performance in financial forecasting requires robust metrics while addressing inherent challenges

Evaluation Metrics

Regression Metrics

MAE

Mean Absolute Error

RMSE

Root Mean Squared Error

MAPE

Mean Absolute Percentage Error

R²

Coefficient of Determination

Classification Metrics

Accuracy

Overall correctness

F1-Score

Precision-Recall balance

AUC-ROC

Area Under Curve

Sharpe

Risk-adjusted returns

Key Challenges

Overfitting and Generalization

Financial markets' dynamic nature and limited relevant historical data increase overfitting risk.

Mitigation: Dimensionality reduction (PCA), regularization, walk-forward validation, outlier removal

Model Interpretability

CNN "black box" nature poses challenges for trust, risk management, and regulatory compliance.

Solutions: SHAP values, permutation importance, feature importance analysis

Temporal Bias

Models trained on past data may not adapt to new market regimes (e.g., post-COVID dynamics).

Example: 1950-2023 training data vs. post-2020 market conditions

Future Research Directions

Enhanced Interpretability

Developing methods to understand how specific indicators influence predictions

Multi-modal Fusion

Advanced integration of visual, numerical, and textual data sources

Adaptive Architectures

Models that automatically adapt to changing market conditions

Robustness Enhancement

Improved handling of non-stationarity and market shocks

Convolutional Neural Networks for Investment Signal Generation