The quantum limit of stock price prediction

The noise of stock price prediction

Quantum machine learning models can match the performance of deep learning for stock price prediction using only a fraction of the parameters, yet they still fail to beat market noise. Achieving this efficiency is difficult because financial time series are notoriously non-stationary and non-linear. This early-stage research, currently a preprint on Springer Science and Business Media LLC, suggests that even the most advanced quantum circuits struggle to find patterns where none exist.

Comparing classical and quantum architectures

Researchers evaluated several paradigms to identify any hidden advantage:

• Classical models: Ridge regression, Decision Trees, and Gradient Boosting.
• Deep learning: Artificial Neural Networks (ANN) and Long Short-Term Memory (LSTM) units.
• Quantum variants: Variational quantum circuits, including Quantum LSTMs and Quantum Deep Q-Networks (QDQN).

The study utilised over 200,000 one-minute bars across eleven liquid U.S. equities. On these intraday scales, the data resembled a random walk; no model—quantum or otherwise—bested a naive baseline. The team compared RMSE, MAE, and Sharpe ratios to ensure a rigorous evaluation. On daily horizons, deep learning models showed marginal gains, with quantum versions matching classical performance despite having a lower parameter count.

The reality of quantum advantage

The findings indicate that model exoticness is no substitute for rigorous experimental design. While the classical DQN agent managed positive average returns, the quantum agent and the standard buy-and-hold strategy lagged behind. This suggests that near-term quantum hardware may organise data efficiently but lacks a definitive profit-making edge. This research does not solve the fundamental issue of non-stationarity in financial time series. Future efforts must focus on signal quality rather than merely increasing computational complexity to find alpha in the noise.