Quantum Attention Deep Q-Network for Financial Market Prediction
By Amrit, Abhiraj
Financial markets exhibit nonstationary dynamics, regime shifts, and complex interactions that challenge standard supervised learning predictors. Reinforcement learning (RL) is well-suited to decision-making under uncertainty because it optimizes sequential actions directly toward long-term reward (e.g., risk-adjusted return) [1]. This project aims to develops a hybrid quantum–classical Deep Q-learning agent that uses a Long Short-Term Memory (LSTM) encoder for temporal feature extraction and quantum attention/post-net layers built from Variational Quantum Circuits (VQCs) for expressive, compact representations. The quantum layers exploit superposition and entanglement to generate rich nonlinear features, and their outputs are measured to produce action values. A modification from the existing QADQN architecture by Dutta et al. [2] augments the daily feature vector of OHLC (Open, High, Low, Close) data with widely used technical indicators (RSI, MACD, Bollinger-band distances, and Volume), providing informative signals. The end-to-end system learns a trading policy (sit/buy/sell) from historical market data under realistic backtesting (including transaction costs).
The intellectual merit of this project lies in its step-by-step integration of quantum computing techniques into a well-established deep reinforcement learning pipeline for financial time series. By combining a classical LSTM for temporal pattern recognition with a quantum self-attention module via variational quantum circuits, the design can potentially uncover richer representations than classical counterparts and create benefits in computation time [2, 3]. The project hopes to enable controlled investigation of these quantum mechanisms in a benchmarked backtesting environment. Augmenting the input feature set with domain-standard technical indicators allows for an unbiased evaluation of whether quantum hybrid models respond to realistic information. This offers insights into both the representational power of variational circuits and the value of increased market context.
The broader impacts of this research extend beyond academic advancement in Quantum Machine Learning (QML). The integration of technical indicators with quantum algorithms could revolutionize algorithmic trading by providing institutional investors and fund managers with more robust, risk-adjusted trading strategies. This work contributes to the emerging field of quantum finance [4,5], potentially influencing the development of quantum-enhanced financial products and services. Moreover, this research could accelerate the adoption of quantum technologies in the financial sector, driving innovation in quantum hardware development and hybrid quantum-classical algorithms. The educational impact includes training the next generation of researchers in the intersection of quantum computing and finance, fostering collaboration between physics, computer science, and economics departments.