Abstract
Quantum machine learning (QML) is increasingly proposed for financial forecasting and risk analysis, yet comparisons with classical methods almost always stop at accuracy and wall-clock time. As quantum systems approach deployment, their resource footprint — cryogenic cooling, repeated circuit executions, calibration and the classical optimiser loops surrounding them — becomes a first-order concern. This paper introduces an energy-aware benchmarking framework for hybrid quantum–classical ML pipelines and applies it to three financial tasks: credit-risk scoring (UCI German Credit) and next-day direction prediction for the S&P 500 and NIFTY 50 indices. Variational quantum classifiers (VQC) and quantum-kernel support vector machines (QSVM) at 4–8 qubits are compared with class-weighted logistic regression, RBF-SVM, random forest, XGBoost and a GPU-trained LSTM on identical data subsets. Classical energy is measured directly by a 10 Hz power sampler; quantum energy combines the measured optimisation-loop energy with a per-shot hardware cost model spanning optimistic to pessimistic scenarios. On credit risk the best quantum model (QSVM, 8 qubits) trails logistic regression by 7.8 balanced-accuracy points while consuming 7 × 1010 times its energy. On index direction all models, classical and quantum, perform at chance, so the 109× quantum energy premium purchases nothing. The gap is insensitive to the hardware assumption: break-even would require roughly 10-10 J per shot, ten orders of magnitude below current superconducting systems. Once the statistical uncertainty of the accuracy estimates is taken into account, no quantum model.

This publication has 19 references indexed in Scilit: