qadqn-trading

Quantum Attention Deep Q-Network (QADQN) for financial market prediction and optimal trading strategy development. Variational quantum circuit inside deep Q-learning framework, achieving superior risk

SKILL.md

GitHub

QADQN Trading

Core Methodology

Quantum Attention Deep Q-Network (QADQN) — hybrid quantum-classical reinforcement learning for financial market prediction and optimal trading strategy. Embeds variational quantum circuits within a traditional deep Q-learning framework to leverage quantum advantages in decision-making.

Architecture

Market State → Feature Encoding → Quantum Attention Layer → VQC Q-Value Network → Trading Action
                     ↑                      ↑
              Historical Data         Experience Replay
              (OHLCV, indicators)    + Target Network

Components

  1. State Representation: Market features (prices, technical indicators) encoded as quantum state via amplitude or angle encoding
  2. Quantum Attention Layer: Variational quantum circuit computes attention weights over feature dimensions, identifying market-relevant features
  3. Q-Value Network: Hybrid quantum-classical network outputs Q-values for actions (buy/hold/sell)
  4. Experience Replay: Standard DQN mechanism with prioritized sampling for stable training
  5. Target Network: Stabilizes learning via periodic weight copying

Performance Metrics (S&P 500)

  • Sortino Ratio: 1.28 (non-overlapping test), 1.19 (overlapping test)
  • Downside Risk Management: Superior risk-adjusted returns vs classical baselines
  • Transaction Costs: Validated with fixed cost per trade — realistic market conditions
  • Published: IEEE QCE 2024 (Quantum Computing and Engineering)

Implementation Pattern

Step 1: Data Preparation

  • Collect historical market data (OHLCV, volume, technical indicators)
  • Split into non-overlapping and overlapping train/test sets
  • Normalize features to [-1, 1] for quantum encoding

Step 2: Quantum Feature Encoding

  • Use angle encoding: map each feature to a rotation angle on a qubit
  • Number of qubits = number of features (or use amplitude encoding for compression)

Step 3: Quantum Attention Circuit

  • Design VQC with trainable rotation gates
  • Compute attention scores via expectation values of Pauli-Z observables
  • Softmax attention weights to weight feature contributions

Step 4: Q-Value Computation

  • Pass weighted features through classical layers
  • Output Q-values for discrete action space (buy/hold/sell)
  • Apply Double DQN or Dueling DQN extensions for stability

Step 5: Training

  • Loss: MSE between predicted Q-values and target Q-values
  • Target: r + γ * max_a' Q_target(s', a')
  • Gradient descent on both quantum and classical parameters

When to Use

  • Algorithmic trading strategy development with quantum-enhanced decision making
  • Risk-aware portfolio management requiring downside protection
  • Market regimes where attention over features improves prediction
  • NISQ-era hybrid quantum-classical implementations

Error Handling

  • Quantum noise: Use noise-aware training, error mitigation (readout correction)
  • Market regime shift: Retrain periodically with recent data, use rolling window
  • Circuit barren plateaus: Use layer-wise training, proper initialization, shallow circuits
  • Transaction cost sensitivity: Include costs in reward function during training

Source

arXiv: 2408.03088 — "QADQN: Quantum Attention Deep Q-Network for Financial Market Prediction" by Siddhant Dutta, Nouhaila Innan, Alberto Marchisio, Sadok Ben Yahia, Muhammad Shafique. IEEE QCE 2024.

Activation: QADQN trading, quantum attention deep q network, quantum RL trading, quantum market prediction, quantum reinforcement learning finance, variational quantum circuit trading