How Do 1,368 Quantum Neural Network Configurations Compare on Financial Forecasting?

The Scale of the Search

Four authors submitted the paper on 8 Oct 2026. They built a grid search that spans encoding methods, ansatz designs, qubit counts, layer depths, and cost functions. The combinatorial space yields 1,368 distinct model configurations. Each configuration receives a full evaluation on prediction accuracy, computational cost, and convergence behaviour. The case study uses the GBP/USD spot exchange rate. No other dataset appears in the benchmark.

Classical baselines exist for comparison. A bidirectional LSTM serves as the reference model. The quantum circuits run in simulation first. Later a subset moves to physical hardware. The paper does not report how many of the 1,368 configurations were executed on the device.

What the Best Configuration Looks Like

The top-performing QNN reaches an $R^2$ score of 0.985. That number beats the classical BiLSTM baseline.

The authors highlight two structural findings. Gate selection and arrangement outweigh raw parameter count. Entanglement behaves as a system-level property of the full circuit rather than a feature isolated to the ansatz block. These statements come from the aggregate analysis across the full grid. The text does not name the specific encoding, ansatz, qubit count, depth, or cost function that produced the 0.985 result.

When the Circuits Meet Real Hardware

Noise assessment happens on the IQM Emerald device. Gate errors and decoherence appear as a significant barrier to practical deployment. Gate selection and circuit depth emerge as key determinants of hardware noise resilience. The paper does not quantify the performance drop between simulation and hardware for the best configuration. It does not state how many qubits the Emerald device provides. It does not report error rates for single-qubit or two-qubit gates on that device.

Where the Guidance Stops

The study establishes a baseline characterisation of QNN noise sensitivity on near-term quantum devices. It offers practical architectural guidance for QNN design. The scope remains limited to regression on a single financial time series. Classification tasks, other asset classes, and multivariate forecasting fall outside the reported work. The authors do not discuss training time per configuration. They do not compare against classical models beyond the BiLSTM. They do not address data loading or state preparation overhead.