This repository contains two variations of a hybrid CNN-LSTM deep learning model for predicting profitable trading opportunities in Infosys (INFY) stock on the National Stock Exchange (NSE) of India. The models combine Convolutional Neural Networks (CNN) for temporal feature extraction with Long Short-Term Memory (LSTM) networks for capturing long-term dependencies in financial time series data.

Target Variable: Binary classification predicting whether the next day's return will be positive.

• Label: 1 if next day's log return > 0, else 0
• Prediction Horizon: 1 day ahead

Use Case: Traditional directional trading strategy suitable for daily momentum-based approaches.

Target Variable: Binary classification predicting whether a profitable exit point exists within the next 6 days.

• Label: 1 if maximum return (close-to-high) in next 6 days exceeds 3%, else 0
• Forward Window: 6 days
• Profit Threshold: 3%

Use Case: Swing trading strategy focused on identifying entry points with high probability of reaching profit targets.

Both models use an identical CNN-LSTM hybrid architecture:

Input Layer (28 timesteps × 27 features)
    ↓
Conv1D Layer (64 filters, kernel=3, causal padding, ReLU)
    ↓
Batch Normalization
    ↓
MaxPooling1D (pool_size=2)
    ↓
Dropout (30%)
    ↓
LSTM Layer (64 units, dropout=30%, recurrent_dropout=10%)
    ↓
Batch Normalization
    ↓
Dropout (30%)
    ↓
Dense Output (1 unit, sigmoid activation)

1. Convolutional Layer (CNN):

   • Extracts local temporal patterns from price sequences
   • Uses causal padding to prevent look-ahead bias
   • 64 filters with kernel size 3 capture short-term patterns

2. LSTM Layer:

   • Models long-term dependencies and market regime changes
   • 64 hidden units maintain sequence memory
   • Dropout layers prevent overfitting

3. Regularization:

   • Batch normalization after CNN and LSTM layers
   • Multiple dropout layers (30% standard, 10% recurrent)
   • Early stopping and learning rate reduction callbacks

• log_returns: Daily logarithmic returns
• ret_3d, ret_5d, ret_10d: Multi-horizon returns (3, 5, 10 days)
• lag_1 to lag_7: Lagged returns for momentum patterns

• RSI_14: Relative Strength Index (14-day)
• MACD: MACD histogram (12,26,9 configuration)
• MACD_line, MACD_signal: MACD components
• OBV: On-Balance Volume

• volatility_10: 10-day rolling standard deviation
• vol_10d, vol_21d, vol_63d: Annualized volatility (10, 21, 63 days)
• vol_ratio: Short-term to long-term volatility ratio
• ATR_14: Average True Range (14-day)

• ret_skew_21: 21-day return skewness
• ret_kurt_21: 21-day return kurtosis

• vol_zscore_20: 20-day volume z-score

• Stock: Infosys (INFY) from NSE
• Period: January 2005 - December 2024 (20 years)
• Frequency: Daily OHLCV data
• Source: NSEPython library

Historical prices are adjusted for corporate actions (1:1 bonus issues):

• 2018-09-04
• 2015-06-15
• 2014-12-02
• 2006-07-13

• Sequence Length: 28 days (approximately one trading month)
• Rolling Windows: Each prediction uses the previous 28 days of normalized features
• Scaling: StandardScaler fitted on training data, applied to validation/test sets

• Strategy: TimeSeriesSplit with 6 folds
• Test Size: 252 days (1 trading year) per fold
• Purpose: Prevents data leakage and simulates walk-forward testing

• Optimizer: Adam (initial learning rate: 0.001)
• Loss Function: Binary crossentropy
• Batch Size: 64
• Max Epochs: 20 (CV), 50 (final model)
• Callbacks:
  • EarlyStopping (patience=5, monitor=val_loss)
  • ReduceLROnPlateau (factor=0.5, patience=3)

• Train Period: All data before 2022-01-01
• Test Period: 2022-01-01 onwards
• Purpose: Final model validation on completely unseen data

Predictions (probabilities) are converted to trading signals using dynamic thresholds:

• Long Signal: Prediction ≥ threshold + 0.02
• Short Signal: Prediction ≤ threshold - 0.02
• Neutral: Otherwise (if neutral band enabled)

A grid search over threshold values (mean ± 0.10) optimizes for maximum Sharpe ratio on cross-validation data.

1. Position Holding:

   • Minimum hold period: 0-1 days (model dependent)

2. Stop-Loss Mechanisms:

   • Daily Stop: Exit if single-day loss exceeds -2%
   • Rolling Stop: Exit if 10-day cumulative loss exceeds -5%

3. Volatility Targeting:

   • Target annual volatility: 12%
   • Maximum leverage: 2.0x
   • Position sizing scales inversely with realized volatility

4. Transaction Costs:

   • Cost: 5 basis points (0.05%) per turnover
   • Applied on position changes to simulate realistic execution

• Total Return & CAGR: Annualized compounded returns
• Sharpe Ratio: Risk-adjusted returns
• Maximum Drawdown: Largest peak-to-trough decline
• Hit Rate: Percentage of profitable trades
• Turnover: Number of position changes
• Comparison: All metrics computed for both strategy and buy-and-hold benchmark

Both notebooks generate:

1. Cross-validation performance by fold (accuracy, precision, recall, F1, ROC-AUC, PR-AUC)
2. Threshold sweep analysis with Sharpe ratio visualization
3. Out-of-sample backtest summary comparing strategy vs. buy-and-hold
4. Equity curve plots showing cumulative performance
5. Saved artifacts:
   • cv_predictions.npy: Cross-validation predictions
   • cv_labels.npy: True labels
   • threshold_sweep_cv.csv: Threshold optimization results
   • backtest_oos_summary.csv: Final OOS performance
   • Equity curve plots (PNG)

numpy
pandas
matplotlib
scikit-learn
tensorflow / keras
nsepython
pandas_ta
quantstats
tabulate

• Random Seed: 42 (for reproducibility)
• TensorFlow Determinism: Enabled via environment variables

1. Install dependencies:

   pip install numpy pandas matplotlib scikit-learn tensorflow nsepython pandas-ta quantstats tabulate

2. Run notebooks sequentially:

   • Model 1: Infy-CNN_LSTM.ipynb
   • Model 2: Infy-CNN_LSTM(1).ipynb

3. Outputs saved to: ./INFY_CNN_LSTM_Results/

• Multi-asset extension: Apply to portfolio of stocks
• Ensemble methods: Combine predictions from both models
• Feature engineering: Add market regime indicators, sentiment data
• Hyperparameter tuning: Grid search over CNN/LSTM architectures
• Alternative targets: Continuous value prediction (expected return)
• Real-time deployment: Integrate with live data feeds

These models are for educational and research purposes only. Past performance does not guarantee future results. Always conduct thorough due diligence and consider transaction costs, slippage, and market impact before deploying any trading strategy in live markets.

MIT License - Feel free to use and modify for your research and trading experiments.

Author: Akshat Jiwrajka
Last Updated: November 2025