Quantum Machine Learning Fails to Beat Classical Models in Financial Prediction Test

Quantum machine learning has been one of the most tantalizing promises in computational finance: the idea that qubits, superposition, and entanglement could somehow tease predictive patterns out of noisy market data that classical algorithms cannot reach. Now a large, systematic empirical test has delivered a sobering verdict. In a comprehensive evaluation published in Neural Computing and Applications, researchers led by Raed Abu Zitar of Liwa University in Abu Dhabi, working with collaborators across the United Arab Emirates and Spain, pitted a full arsenal of quantum models against classical machine learning methods on four real-world financial datasets — and classical models won every single time.

The study is notable for its sheer scope. Rather than cherry-picking a single quantum architecture, the team constructed a systematic experimental framework spanning four task categories: fully classical pipelines combining classical preprocessing with classical models, hybrid pipelines pairing classical preprocessing with quantum models, the reverse hybrid using quantum preprocessing with classical models, and fully quantum pipelines from preprocessing to prediction. For classification, they analyzed the Bank Marketing dataset, which predicts whether customers will subscribe to a term deposit, and the Credit Card Fraud Detection dataset, one of the most imbalanced and challenging problems in applied machine learning. For regression, they evaluated IBM stock price prediction and Bitcoin price forecasting, tasks where even small predictive edges translate into real money.

On the quantum side, the researchers deployed Quantum Support Vector Machines, Variational Quantum Classifiers and Regressors, Quantum Neural Networks, and Quantum Decision Trees. Two quantum feature encodings — the ZZFeatureMap and the PauliFeatureMap — were systematically compared, because how classical data gets encoded into quantum states is widely regarded as the decisive bottleneck in quantum machine learning. The team also varied the classical preprocessing layer, testing Principal Component Analysis for dimensionality reduction and Particle Swarm Optimization for hyperparameter tuning, generating dozens of distinct pipeline configurations across all four task categories.

The technical results were unambiguous. For Bank Marketing classification, a classical Decision Tree achieved 76.6 percent accuracy with an F1 score of 0.766, while the best quantum approach — a hybrid classical-plus-quantum neural network with ZZFeatureMap encoding — reached only 74.1 percent accuracy with an F1 of 0.739. Credit Card Fraud Detection showed the same pattern: classical Decision Trees hit 96.1 percent accuracy against 95.7 percent for their quantum counterparts. In regression, the gap widened dramatically. Classical models achieved near-perfect fits, with R-squared values above 0.999 for IBM stock prices and a perfect 1.000 for Bitcoin forecasting, while pure quantum pipelines struggled so badly that some configurations produced negative R-squared scores — meaning their predictions were worse than simply guessing the mean value of the training data.

That negative R-squared result deserves attention, because it reveals something fundamental about where current quantum models break down. Financial time series are dominated by trend and scale: a model that fails to capture the overall level of a price series produces residuals larger than the variance of the data itself. Quantum circuits with limited depth and noisy intermediate-scale quantum hardware constraints simply cannot represent the smooth, high-dimensional functions that classical neural networks learn effortlessly. The researchers found that Particle Swarm Optimization improved the stability of quantum model hyperparameters, but it never closed the performance gap with classical approaches — a finding that suggests the problem lies deeper than tuning.

Perhaps the most technically interesting finding concerns quantum preprocessing. When the team applied quantum feature maps to classical models — Task 3 in their framework — performance consistently underperformed the classical preprocessing alternatives. This is a striking result because quantum feature maps are theoretically motivated: they compute kernel functions in exponentially large Hilbert spaces that would be classically intractable. But the theory assumes the quantum kernel captures meaningful structure in the data. On real financial datasets with noisy, weakly correlated features, the quantum encodings appear to inject more distortion than useful structure, degrading the downstream classical learners rather than enhancing them.

Hybrid approaches — the classical-plus-quantum neural networks of Task 2 — occasionally came closest to classical performance, and this is where the most promising quantum results appeared. But the researchers concluded that the added computational complexity and resource requirements preclude practical adoption. Simulating quantum circuits on classical hardware is exponentially expensive, and running them on actual quantum processors introduces hardware noise, limited qubit counts, and circuit depth constraints that erode any theoretical advantage. For a production financial system processing millions of transactions daily, the overhead of a hybrid quantum pipeline is simply not justified by a two-and-a-half percentage point accuracy deficit.

The implications ripple across the quantum computing industry, which has invested heavily in quantum finance applications. Companies from IBM to specialized quantum startups have published papers claiming quantum advantages in portfolio optimization, fraud detection, and time series forecasting. This study does not refute the theoretical foundations of quantum machine learning — the Havlíček kernel methods and variational algorithms remain mathematically sound — but it does challenge the practical narrative. The researchers explicitly recommend that practitioners prioritize classical machine learning methods for production financial systems while monitoring QML developments for future potential. In an industry where marketing hype often outruns empirical evidence, that recommendation carries weight precisely because it comes from a comprehensive, systematic comparison rather than a single cherry-picked benchmark.

What would it take for quantum machine learning to become viable for finance? The authors point to fundamental challenges in quantum circuit design, feature encoding, and optimization that must be addressed before viable business applications emerge. Deeper circuits with better expressivity, error-corrected hardware with more qubits, and encoding schemes that actually capture the statistical structure of financial data are all prerequisites. Until then, the study stands as one of the most thorough empirical reality checks yet performed on quantum machine learning — a reminder that in the race between quantum promise and classical performance, the classical algorithms are still running laps around their quantum rivals.

Subject of Research: A comprehensive empirical evaluation of quantum machine learning for financial prediction: systematic comparison with classical approaches

Article Title: A comprehensive empirical evaluation of quantum machine learning for financial prediction: systematic comparison with classical approaches

Article References: Zitar, R. A., Loucif, S., Al-Rajab, M., & Nasayreh, A. (2026). A comprehensive empirical evaluation of quantum machine learning for financial prediction: systematic comparison with classical approaches. Neural Computing and Applications, 38(20), Article 817. https://doi.org/10.1007/s00521-026-12516-y

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12516-y

Keywords: comprehensive, empirical, evaluation, quantum, machine, learning, financial, prediction, systematic, comparison, classical, approaches