Quantum Time-Series Analysis Boosts Forecasting

A new quantum-inspired ARIMA methodology improves the accuracy of time series analysis, a key technique used across diverse fields from environmental modelling to industrial forecasting. Nishikanta Mohanty and colleagues at University of Technology Sydney present a method that uses quantum-assisted techniques for identifying relevant patterns and estimating parameters within complex datasets. The approach integrates quantum autocorrelation and partial autocorrelation functions with variational quantum circuits, offering a potential reduction in meta-optimisation overhead and providing clear insight into where quantum effects enhance order discovery, lag refinement, and parameter estimation. Through rolling-origin evaluations against classical ARIMA models, the team demonstrates the potential of their method to improve forecasting accuracy, as measured by metrics including mean squared error and mean absolute percentage error.

Quantum enhancements deliver substantial gains in time series forecasting accuracy

A 14% decrease in mean squared error was observed when comparing the quantum-inspired ARIMA methodology to automated classical ARIMA models across environmental and industrial datasets. Previously, such gains required extensive manual tuning and were highly sensitive to data characteristics, often failing to achieve consistent reductions below 5% in similar forecasting scenarios. The methodology’s seven quantum contributions, differencing selection, quantum autocorrelation (QACF), quantum partial autocorrelation (QPACF), swap-test primitives with delayed-matrix construction, VQC-AR, VQC weak-lag refinement, and VQC-MA, collectively minimise meta-optimisation and pinpoint where quantum effects refine order discovery, lag refinement, and parameter estimation. Time series analysis, at its core, aims to predict future values based on historical data, and the ARIMA (Autoregressive Integrated Moving Average) model is a cornerstone of this field. Classical ARIMA models rely on identifying correlations within the time series data to extrapolate future trends, but this process can be computationally expensive and often requires significant expertise to tune effectively. The quantum-inspired approach seeks to alleviate these limitations by leveraging principles from quantum mechanics to accelerate and improve the accuracy of these crucial steps.

Rolling-origin evaluations, utilising metrics like mean squared error, mean absolute percentage error and the Diebold, Mariano test, confirm the performance of these enhancements across environmental and industrial datasets. Fixed-configuration variational quantum circuits are used for parameter estimation and weak-lag refinement within the moving-average component. Candidate autoregressive and moving-average orders are selected based on quantum autocorrelation and quantum partial autocorrelation, utilising swap tests and phase-corrected projections to align quantum measurements to time-domain regressors, followed by information-criterion parsimony; a lightweight variational quantum circuit weak-lag refinement then re-weights autoregressive lags without altering the selected orders. The swap test, a key quantum primitive employed, allows for the efficient comparison of quantum states without directly measuring them, which is crucial for calculating autocorrelation and partial autocorrelation functions. The delayed-matrix construction is a technique used to map the time-series data into a quantum-accessible format, enabling the application of quantum algorithms. Information-criterion parsimony, such as the Akaike information criterion (AIC) or Bayesian information criterion (BIC), is then used to select the most parsimonious model, that is, the model with the fewest parameters that still adequately explains the data. This prevents overfitting and improves generalisation performance.

Quantum-assisted lag discovery integrates with fixed-configuration variational quantum circuits for parameter estimation and weak-lag refinement in this approach. Swap-test-driven quantum autocorrelation and quantum partial autocorrelation identify differencing and candidate lags, utilising a delayed-matrix construction to align quantum projections to time-domain regressors, followed by information-criterion parsimony. After screening orders, a fixed variational quantum circuit ansatz, optimizer, and training budget are retained, and the circuit is deployed for autoregressive and moving-average coefficient estimation; a lightweight weak-lag refinement re-weights AR lags without altering the selected orders. The autoregressive (AR) component models the dependence of a value on its own past values, while the moving-average (MA) component models the dependence on past forecast errors. By refining both these components, the methodology aims to capture the underlying dynamics of the time series more accurately. The use of a fixed variational quantum circuit is significant; these circuits are parameterised quantum algorithms designed to be optimised for a specific task. The ‘ansatz’ refers to the specific structure of the quantum circuit, and the ‘optimizer’ is the algorithm used to adjust the circuit’s parameters to minimise a cost function, in this case, the forecasting error. Maintaining a fixed configuration simplifies the optimisation process and reduces the computational resources required, although it may limit the potential for achieving the absolute best possible performance.

Fixed variational circuits balance clarity against potential forecasting gains

The current reliance on “fixed-configuration” variational quantum circuits (VQCs) presents a notable constraint, despite demonstrable improvements in forecasting accuracy against traditional automated models. This deliberate design choice, intended to minimise extraneous variables and maintain clarity, potentially limits the ultimate predictive power achievable. Researchers acknowledge this trade-off, suggesting that a fully flexible VQC could sharply outperform the current implementation. The choice of a fixed VQC is a pragmatic one, balancing the potential for increased accuracy with the complexities of training more flexible models. Fully flexible VQCs, while potentially more powerful, require significantly more computational resources and are more susceptible to issues like vanishing gradients, making them difficult to train effectively. The current approach prioritises a stable and interpretable solution, even if it means sacrificing some potential performance gains.

Despite the potential for improvement through more complex quantum circuits, this represents a strong step forward in applied quantum forecasting, with gains achieved across diverse datasets including Australian beer sales and Sydney temperature readings. This establishes a pathway for integrating quantum principles into time series analysis, moving beyond purely classical approaches. By combining established statistical modelling with quantum-assisted techniques, improvements in parameter estimation and lag refinement within the ARIMA framework were achieved, utilising quantum autocorrelation and partial autocorrelation to identify key patterns. The successful application of this methodology to real-world datasets, such as the Australian beer sales data (a classic time series benchmark) and Sydney temperature readings, demonstrates its practical viability. Future research could explore the use of more sophisticated quantum algorithms, such as quantum machine learning techniques, to further enhance the performance of the ARIMA model. Furthermore, investigating the scalability of this approach to handle larger and more complex datasets is crucial for its widespread adoption. The potential for reducing the computational burden of time series analysis, particularly in applications involving high-frequency data or long historical records, is a significant benefit of this quantum-inspired methodology.

The researchers developed a new time series forecasting method combining quantum-inspired techniques with the ARIMA statistical model. This approach utilises quantum autocorrelation and partial autocorrelation to improve the identification of relevant patterns and refine parameter estimation. Results from evaluations on datasets including Australian beer sales and Sydney temperature readings demonstrated a reduction in forecasting errors compared to classical ARIMA methods. The authors suggest future work may focus on exploring more advanced quantum algorithms to further enhance model performance and scalability.