Large-Scale Benchmarking of Quantum Neural Network Configurations for Financial Time Series Forecasting
This study presents a large-scale benchmarking of 1,368 Quantum Neural Network configurations for GBP/USD financial time series forecasting, identifying that specific gate selections and circuit architectures are more critical than parameter count for achieving high accuracy (outperforming classical BiLSTM) while also highlighting gate errors and decoherence as major barriers to practical deployment on near-term quantum hardware.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The world of finance relies heavily on predicting the future, specifically the movement of currency values, stock prices, and economic indicators. For decades, scientists have used classical computers to find patterns in these numbers, treating them like a complex puzzle where past data hints at what comes next. However, as data becomes more intricate and relationships more tangled, traditional computers sometimes struggle to see the full picture. This has led researchers to explore a different kind of machine: the quantum computer. Unlike standard computers that process information as simple on-or-off switches, quantum machines use the strange rules of physics to handle information in a way that allows them to exist in many states at once. When these machines are programmed to learn from data, they become quantum neural networks. These are not just faster versions of old models; they are fundamentally different tools that offer new pathways for processing high-dimensional data. The question facing scientists today is not just whether these machines can work, but how to build them correctly to solve real-world problems like predicting the price of money.
A team of researchers at Kingston University London and Brunel University London set out to answer this question by treating the design of a quantum neural network like a massive construction project. They wanted to know which specific parts of the machine worked best when trying to forecast the exchange rate between the British pound and the US dollar. To do this, they did not just build one model; they built and tested 1,368 different versions. Each version was a unique combination of different building blocks, such as how the data was fed into the machine, the specific arrangement of the internal logic gates, the number of quantum bits used, and the mathematical rules used to judge success. By running this grid search, the researchers could see exactly how changing one small piece of the puzzle affected the final result, providing a clear map of what works and what does not.
The study focused on a specific dataset of daily exchange rates spanning nearly twenty years, from 2005 to 2025. The researchers split this data into three parts: one for teaching the models, one for tuning them, and one for testing their final performance. They compared their quantum creations against a standard, highly sophisticated classical computer model known as a BiLSTM, which is currently considered a top performer for this type of time-based prediction. The goal was to see if the quantum models could not only match this classical benchmark but potentially surpass it. The results were striking. The single best-performing quantum configuration achieved a score of 0.985 in its ability to predict the data. While this slightly exceeded the score of a constrained classical model, it is important to note that the unconstrained classical baseline, which had access to a substantially longer lookback window and fully overlapping data, achieved a score of 0.9805. Nevertheless, the best quantum configuration demonstrated competitive and in some cases superior predictive accuracy compared to classical models operating under equivalent data constraints, providing encouraging evidence of the potential for QNNs in financial forecasting.
However, the path to this success was not a straight line, and the researchers discovered that bigger was not always better. A common assumption in the field is that adding more quantum bits or more layers of complexity would automatically lead to better results. The study found the opposite to be true. The most successful models used a relatively small number of quantum bits and a moderate number of layers. When the researchers increased the size of the model too much, the performance often dropped. This happened because larger models became harder to train and more prone to getting stuck in a state where they could not learn effectively. Furthermore, the researchers found that the specific choice of gates—the basic operations the machine performs—mattered far more than the total number of parameters or the sheer size of the circuit. It was not about having the most powerful engine, but about having the right engine for the job.
Another critical discovery concerned how the machine handles the "entanglement" of its parts. In quantum physics, entanglement is a phenomenon where particles become linked so that the state of one instantly influences the other, no matter the distance. The team found that this linking does not have to happen only in the main processing part of the machine. Instead, it is a property of the entire system. If the method used to feed data into the machine created enough entanglement, the main processing unit could be simpler and still perform well. Conversely, if the data feeding method did not create this link, a complex processing unit could not compensate for it. This means that the success of the model depends on the harmony of the whole system, not just the sophistication of its central component.
The researchers also tested how these models would behave on real, physical quantum hardware, rather than just on a computer simulation. They ran a selection of their best designs on a real quantum device called the IQM Emerald. The results here were sobering. While the models worked perfectly in the simulation, the real device introduced errors caused by the physical nature of the machine, such as noise and the loss of quantum states over time. These errors caused the prediction accuracy to drop significantly. The study showed that the depth of the circuit, or how many steps the machine had to take, was a major factor in how much the noise hurt the performance. Deeper circuits suffered more. Interestingly, the type of error that mattered most was not the reading of the final result, but the errors that happened during the calculation steps themselves. This suggests that for quantum computers to be useful in the real world right now, the designs must be kept shallow and simple to survive the noisy environment of current hardware.
Despite the challenges posed by real-world hardware, the study provides a clear blueprint for the future. The researchers demonstrated that it is possible to build a quantum neural network that competes with, and in specific configurations surpasses, classical models for financial time series, provided the architecture is chosen with care. The best designs balanced the number of components, the type of connections, and the method of data entry to create a machine that was both powerful and resilient. The findings indicate that while the technology is still in its early stages and faces significant hurdles related to hardware noise, the potential is real. The work establishes a baseline for how these machines should be built, offering practical guidance for engineers who are trying to turn quantum theory into a tool that can predict the future of money. As quantum hardware continues to improve, reducing the noise that currently limits performance, the configurations identified in this study could serve as the foundation for the next generation of financial forecasting tools.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.