AI-based Trading Strategies: Comparing Deep Learning Architectures

This repository presents a structured framework for developing and evaluating AI-driven trading strategies, with a focus on systematic risk premia. The project integrates cutting-edge deep learning architectures with robust backtesting and interpretability tools — all within a modular and extensible research pipeline.

Data Source

This project utilizes historical market data obtained through the EODHD APIs. Their high-quality financial data has been instrumental in developing and testing AI-based trading strategies.

Description:

This project deploys a systematic, AI-driven framework to develop and evaluate quantitative trading strategies. Our modular pipeline integrates robust API-based data acquisition with advanced feature engineering—including regime detection via Hidden Markov Models—to produce high-quality inputs for our models. We implement a suite of deep learning architectures, including a GRU, a standalone LSTM model, and an LSTM-CNN hybrid, all validated through rolling cross-validation and realistic backtesting that incorporates transaction cost adjustments. Performance is benchmarked against a buy-and-hold baseline, and the framework supports rapid experimentation with alternative architectures such as LSTM with attention mechanisms and transformers.

Clean Strategy Comparison

This project builds an end-to-end pipeline for evaluating deep learning models in financial markets. It includes:

• API-driven data ingestion
• Feature engineering (including regime detection using Hidden Markov Models)
• Multi-architecture model experimentation (GRU, LSTM, CNN-LSTM, Attention LSTM, Transformers)
• Rolling-window validation and realistic backtesting
• Transaction cost sensitivity, risk-adjusted metrics, and hybrid overlays
• Explainability via attention weights and strategy diagnostics

The final deliverable includes a strategy-ready version and a hybrid attention-based allocation prototype for future AI+XAI asset management models.

Neural Networks for Trading Strategies

Neural networks (NNs) have been widely adopted in trading strategies due to their ability to model nonlinear relationships, detect patterns, and adapt to changing market conditions. The evolution of deep learning architectures, including convolutional neural networks (CNNs), recurrent neural networks (RNNs), and transformer-based models, has further enhanced their predictive power. NNs have revolutionized trading by enhancing prediction accuracy, optimizing execution, and identifying alpha signals in ways that traditional methods cannot. However, challenges such as interpretability, overfitting, and computational cost must be addressed for robust deployment.

Long Short-Term Memory (LSTM) networks are a specialized type of RNNs designed to handle sequential data and long-term dependencies—making them particularly well-suited for financial time-series forecasting and trading strategies. LSTMs are a powerful tool for time-series forecasting and trading, but they require careful tuning and robust validation. With recent advancements in deep learning, transformer architectures have proven highly effective in time-series forecasting, outperforming LSTMs and traditional statistical models.

Transformer vs SPY Benchmark

Deep learning models enable systematic strategies to:

• Capture non-linear relationships
• Detect regime shifts and volatility clusters
• Generate adaptive position sizing signals

This project uses:

• LSTM for sequential modeling
• GRU as a compact alternative
• CNN-LSTM for local-global patterns
• Attention-LSTM for interpretable dynamics
• Transformers for end-to-end forecasting

LSTM with Attention-Based Trading Strategy

Our approach leverages a state-of-the-art deep learning architecture that combines the sequential modeling strengths of LSTM networks with an attention mechanism designed to isolate and weigh the most predictive segments of historical market data. This attention-based framework also opens promising avenues for systematic exploration in asset allocation strategies, enabling dynamic weighting of assets informed by adaptive market regime detection.

Attention-LSTM vs SPY Benchmark

Strategy performance is benchmarked with vectorbt and adjusted for transaction costs. Attention mechanisms are further analyzed for signal interpretability and allocation logic.

Project Structure

AI-based-Trading-Strategies/
│
├── envs/
│   ├── env_models_simple.yml        # LSTM, GRU, CNN-LSTM
│   ├── env_models_tf.yml            # Attention LSTM, Transformer
│   └── env_comparison.yml           # vectorbt and reporting
│
├── notebooks/
│   ├── 01_LSTM_model_template.ipynb                 # Baseline LSTM model
│   ├── 02_GRU_model_template.ipynb                  # GRU architecture
│   ├── 03_CNN_LSTM_model_template.ipynb             # CNN-LSTM hybrid
│   ├── 04_ATT_LSTM_model_template.ipynb             # LSTM with attention layer
│   ├── 04_ATT_LSTM_model_attention_weights.ipynb    # Attention score extraction
│   ├── 05_Transformer_model_template.ipynb          # Transformer-based forecaster
│   ├── 06_Full_Strategy_Clean_Baseline.ipynb        # vectorbt backtesting pipeline
│   ├── 07_Strategy_Enhancements_Sensitivity.ipynb   # Realism overlays: min-hold, vol sizing, stop-loss
│   └── 08_Hybrid_Attention_Allocation.ipynb         # Attention-driven dynamic asset allocation
│
├── data/
│   ├── df_lstm.csv
│   ├── df_gru.csv
│   ├── df_cnn.csv
│   ├── df_att.csv
│   ├── df_trans.csv
│   └── GSPC_fixed.csv
│
├── README.md
├── requirements.txt
└── .gitignore

Highlight: Attention-Based Allocation

This framework goes beyond signal generation — it shows how attention weights can be used for interpretable, dynamic asset allocation, bypassing traditional optimizers (e.g., covariance matrix inversion). It paves the way for explainable AI (XAI) in portfolio construction.

Motivation

This research aims to bridge AI interpretability and portfolio realism, contributing to the next generation of systematic strategies:

• Enhanced forecasting in noisy environments
• Smarter, cost-aware execution
• Modular testing of AI-based allocation models

Methodology

📌 Methodology Overview

1. Data Collection & Preprocessing

   • EODHD APIs for historical prices
   • Normalization and cleaning
   • Regime-based labeling

2. Model Development

   • Model training (LSTM, GRU, CNN-LSTM, etc.)
   • Cross-validation and performance tracking

3. Strategy Construction

   • Convert predictions to trading signals
   • Apply overlays (stop-loss, volatility sizing, hybrid constraints)

4. Backtesting & Evaluation

   • Use vectorbt, QuantStats, and Sharpe sensitivity
   • Visualize performance: cumulative return, drawdown, signal frequency

Installation & Requirements

The project is built using Python 3.8+ and relies on several key libraries. To set up the environment, clone the repo and install dependencies:

Clone the repository:

git clone https://github.com/FranQuant/AI-based-Trading-Strategies.git
cd AI-based-Trading-Strategies
pip install -r requirements.txt

Contributing

Contributions are welcome! If you have ideas for improvements, please feel free to open an issue or submit a pull request. When contributing, please adhere to the repository’s coding style and document any changes for clarity.

License

This project is licensed under the MIT License.

DISCLAIMER

This project is intended solely for educational and research purposes.

• It is not designed for real trading or investment use.
• No warranties or guarantees are provided.
• The creator bears no responsibility for any financial losses. By using this software, you acknowledge and agree that it is for learning purposes only.