CNN-based Stock Price Prediction with Instantaneous Sharpe Ratio

This project implements a CNN-LSTM model to predict the risk-adjusted stock return, represented as an instantaneous Sharpe ratio. The prediction target quantifies the expected k-bar return normalized by its inherent volatility, allowing the model to focus on high signal-to-noise opportunities.

Prediction Target: Instantaneous Sharpe Ratio

The model predicts s_t^(k), defined as:

[ s_t^{(k)} = \frac{C_{t+k} - C_t}{C_t} \Bigg/ \sqrt{\frac{1}{k}\sum_{i=1}^{k}r_{t+i}^2} ]

Where:

• C_t is the closing price at time t
• r_{t+i} are the future returns over the next k periods
• The denominator represents the square root of the mean of squared future returns (RMS volatility)

Interpretation

This ratio represents an instantaneous Sharpe for a k-bar trade, where:

• s = 1 means the expected return equals one standard deviation of future noise
• s = 2 means two standard deviations, etc.

This standardized measure helps the model identify opportunities with attractive risk-reward profiles rather than just high absolute returns.

Model Architecture

The model employs a sophisticated CNN-LSTM architecture with attention mechanisms:

1. Feature Importance Layer:

   • Learns to weight input features with a trainable parameter vector
   • Uses softmax activation to create a probability distribution over features

2. Convolutional Blocks:

   • Gated 1D convolutional layers (similar to GLU) for adaptive feature extraction
   • Batch normalization and ReLU activation

3. Dual Attention Mechanism:

   • Channel attention: Captures feature-wise importance
   • Spatial attention: Focuses on important time steps

4. Temporal Processing:

   • Temporal attention to aggregate features across time
   • Bidirectional LSTM with 2 layers to capture sequential patterns

5. Output Layers:

   • Residual blocks with skip connections
   • Trainable output scaling with tanh activation for bounded predictions

Implementation Details

Data Preprocessing

• Raw OHLCV data is transformed into 426+ technical indicators
• For "raw mode," only 9 base features are used: open, high, low, close, volume, ret, ma5, ma10, ma20
• Risk-adjusted returns are calculated using the exact formula shown above
• Data is normalized feature-wise to zero mean and unit variance

Training Process

• Uses sliding window approach for time series data (default window = 30 bars)

• Loss function combines multiple components:

  • Primary loss: Smooth L1 (Huber) loss
  • Correlation penalty: Encourages alignment between prediction and target trends
  • R²-like term: Normalizes MSE by target variance
  • Sparsity regularization: L1 penalty on feature importance
  • Group LASSO: Encourages feature group selection

• Supports balanced sampling across magnitude bins to handle class imbalance

• Implements early stopping with patience and learning rate scheduling

Evaluation Metrics

• MSE/RMSE: Basic prediction error
• MAE: Absolute magnitude of errors
• Directional Accuracy: % of correct direction predictions
• R²: Proportion of variance explained

Backtesting

• Converts model predictions to trading signals (+1/-1/0)
• Computes equity curves, returns, Sharpe ratios, and max drawdowns
• Compares performance against buy-and-hold benchmark

Usage

Data Preprocessing

python data_preprocessing.py data/stock_data.xlsx --out processed_data/output.csv --k-bar 1 --vol-window 20

Training

python train.py processed_data/output.csv --win 30 --target-shift 1 --epochs 20 --raw-only

Evaluation

python eval.py processed_data/output.csv model_checkpoint/model.pt --win 30

Backtesting

python backtest.py eval_out/predictions.csv

Key Findings

1. The risk-adjusted target provides more stable training signals
2. Short-term predictions (target_shift=1) significantly outperform longer horizons
3. Using only raw features often yields better results than all engineered features
4. Balanced sampling improves training when target distribution is highly skewed

Implementation Notes

• The CNN extracts local patterns from price action
• Attention mechanisms help focus on relevant parts of the time series
• LSTM captures temporal dependencies and sequence information
• Feature importance learning identifies the most predictive indicators
• Output scaling helps maintain prediction stability

Project Structure

The codebase follows a modular design with four main components:

1. Data Preprocessing (data_preprocessing.py):

   • Loads raw OHLCV data from Excel files
   • Computes technical indicators (~426 features)
   • Calculates risk-adjusted returns
   • Saves processed data as CSV files

2. Model Training (train.py):

   • Builds and trains CNN-LSTM models
   • Uses sliding window approach for time series data
   • Implements various loss functions (MSE, correlation, R²-like)
   • Offers class-balanced sampling options

3. Model Evaluation (eval.py):

   • Evaluates model performance on test data
   • Calculates metrics (MSE, MAE, directional accuracy, R²)
   • Generates prediction files for backtesting

4. Backtesting (backtest.py):

   • Simulates trading strategies based on model predictions
   • Calculates performance metrics (return, Sharpe ratio, drawdown)
   • Compares to buy-and-hold benchmark