ML-Driven Quantum Portfolio Optimization: Hybrid CNN-LSTM Architecture with Adaptive Zero-Noise Extrapolation on NISQ Devices

The quantum portfolio optimization can be hindered by intrinsic hardware imperfections for noisy intermediate-scale quantum (NISQ), which can impair the precision of expectation value calculations as well as constrain financial outcomes. Our study proposes a machine learning-based zero noise extrapolation (ML-ZNE) approach that can employ a hybrid convolutional neural network and long short-term memory (CNN-LSTM) structure, which can also have an adaptive noise scaling to reduce these errors. The method dynamically estimates the noise-free expectation values from a series of noise-scaled measurements and substitutes traditional static extrapolation methods with a data-driven framework. The CNN-LSTM model can identify the spatial and temporal relationships in noisy quantum circuit outputs that can result in more robust extrapolation than polynomial-fitting-based ZNE. Moreover, the approach includes the real-time hardware calibration data to refine the noise scaling parameters that can elevate the accuracy of quantum computations. Our experimental validation shows that ML-ZNE can decrease both the extrapolation error and improve the portfolio performance, which can quantify the Sharpe ratio while achieving better results than both static ZNE and unmitigated optimization. The proposed method improves the utility of realworld quantum finance by tackling key noise-induced constraints without compromising processing speed on NISQ devices. This work connects the quantum error correction alongside machine learning and presents a scalable approach for advanced quantum portfolio management.