Large-Scale Benchmarking of Quantum Neural Network Configurations for Financial Time Series Forecasting

01 ABSTRACT

Using the GBP/USD spot exchange rate as a case study, the study performs a grid search over encoding methods, ansatz designs, qubit counts, layer depths, and cost functions, yielding 1,368 QNN configurations. The authors report that the best configuration achieves an R² of 0.985, outperforming a classical BiLSTM baseline, and note that gate selection and arrangement are more critical than parameter count, while entanglement is a system-level property of the full circuit. Hardware noise is assessed via the IQM Emerald device, revealing gate errors and decoherence as significant barriers to practical deployment.

02 KEY FINDINGS

1. Grid search covers encoding methods, ansatz designs, qubit counts, layer depths, and cost functions, evaluating 1,368 QNN configurations.
2. The best QNN configuration achieves R²=0.985 on GBP/USD forecasting, outperforming a classical BiLSTM baseline.
3. Gate selection and arrangement are more critical than raw parameter count; entanglement is a system-level property of the full circuit; gate errors and decoherence are key hardware noise barriers.