Researchers Find No Quantum Advantage in Time-Series Forecasting Models

A quantum advantage in practical machine learning tasks remains unproven. Research from the Centre of Finance DHBW Stuttgart and DATEV eG reveals no systematic quantum advantage in time-series forecasting using four tested architectures.

Gerhard Hellstern and colleagues investigated whether quantum computers could improve time-series forecasting accuracy. Using four different forecasting architectures, including both classical and quantum models, they found no evidence of a quantum advantage. No quantum model consistently outperformed the best classical approach across Gaussian-process and NARMA-10 datasets. The team evaluated four forecasting architectures, including both classical and quantum models, using Gaussian-process and NARMA-10 datasets, but found no evidence to suggest a quantum advantage.

A key component of their approach involved a type of neural network called a CRBM, or Conditional Restricted Boltzmann Machine, which learns patterns by assessing how well different configurations fit together. The researchers rigorously tested these models, ensuring a fair comparison by optimising both classical and quantum parameters equally across thirteen experiments. Despite this thorough evaluation, no quantum model consistently outperformed its classical counterpart, raising questions about the current feasibility of quantum machine learning for this specific task.

Quantum and classical Restricted Boltzmann Machine performance equivalence under controlled

A hybrid quantum-classical QCRBM performed statistically indistinguishably from the strongest classical CRBM on both Gaussian-process and NARMA-10 datasets, a result previously unattainable given limitations in symmetric hyperparameter optimisation. Prior comparisons often favoured classical models due to unequal tuning, masking potential quantum benefits, and this study actively avoids that bias through a rigorous thirteen-experiment grid search. The researchers DATEV eG employed a Conditional Restricted Boltzmann Machine, a neural network that learns patterns by assessing how well different configurations fit together, in both classical and quantum forms.

With a sample size of twelve, a power analysis revealed that only medium to large effects are detectable, meaning subtle quantum advantages could remain hidden. An iso-parameter comparison, assessing models with equivalent computational budgets, showed the classical CRBM performed best at three out of four budget levels. Notably, no statistically significant difference emerged between any CRBM and the QCRBM at any tested budget. Experiment six directly characterised the onset of barren plateaus, measuring gradient variance across circuits with two to ten qubits and one to five layers, providing insight into trainability limits.

Symmetric hyperparameter optimisation of classical and quantum Boltzmann machine forecasting models

The team employed a rigorous hyperparameter optimisation process, ensuring a fair comparison between classical and quantum models, which involved a systematic grid search across thirteen structured experiments, carefully tuning parameters for both approaches. This ‘symmetric’ approach is important because it avoids artificially inflating the performance of either type of model through preferential treatment during training, a common pitfall in quantum machine learning evaluations. A Conditional Restricted Boltzmann Machine, or CRBM, served as the central component of this methodology; this type of neural network learns patterns in data by assigning energies to different configurations, similar to how a puzzle piece fits better with certain other pieces. Four forecasting models were evaluated: a classical Gaussian-Bernoulli CRBM, a hybrid quantum-classical QCRBM, a fully quantum QQRBM, and a lag-feature QFeatureQRBM. Testing involved both a Gaussian-process dataset, built using financial data, and the NARMA-10 benchmark, a standard test for nonlinear time-series prediction.

Rigorous optimisation standards reveal no current quantum benefit in energy-based models

The Centre of Finance DHBW Stuttgart and DATEV eG’s findings, while demonstrating no current quantum advantage, nevertheless establish a key benchmark for future work in energy-based models. Simply demonstrating a quantum calculation is possible does not equate to practical improvement, highlighting a persistent challenge in the field. The team’s power analysis reveals a limitation, detecting only medium to large effects with the current sample size, leaving open the possibility that subtle benefits remain elusive. Their work establishes a new standard for evaluating quantum machine learning models, moving beyond simply demonstrating quantum computation to assessing practical improvements, and their investigation into time-series forecasting revealed no systematic quantum advantage across Gaussian-process and NARMA-10 datasets.

The research demonstrated no systematic quantum advantage in energy-based forecasting models when comparing classical and quantum architectures. Using Gaussian-process data generated from financial sources and the NARMA-10 benchmark, four models, including a hybrid quantum-classical CRBM and fully quantum QQRBM, were rigorously tested with symmetric hyperparameter optimisation. This work establishes a new standard for evaluating quantum machine learning, focusing on demonstrable improvements rather than simply quantum computation itself.