Can quantum machine learning offer a forecasting advantage for complex financial data. An evaluation of Quantum Long Short-Term Memory networks and Quantum Reservoir Computing finds no evidence that these models outperform their classical counterparts in predicting monthly revenue series; despite utilising quantum circuits, both achieved similar accuracy to standard methods. Specifically, the QLSTM with over two thousand parameters performed statistically identically to a much larger LSTM containing more than two hundred and sixty-five thousand parameters.
Despite increasing complexity in quantum computing models, recent work evaluating Quantum Long Short-Term Memory networks and Quantum Reservoir Computing found no performance benefit over traditional forecasting methods when applied to monthly revenue data. Forecasting financial time series is notoriously difficult due to short histories and unpredictable shifts; even simple statistical methods often prove competitive with complex machine learning models.
Researchers at DHBW Stuttgart and DATEV eG have been investigating whether quantum computing can offer an edge in this field, specifically by evaluating Quantum Long Short-Term Memory networks, a specialised type of computer memory designed to remember patterns in sequences but built using both classical and quantum processing units, alongside Quantum Reservoir Computing techniques. To assess performance, the team created extended monthly revenue data spanning over eight years using a method that simulates rolling dice based on past results, ensuring realistic market fluctuations.
Quantum LSTM shows parity without scaling advantages for time series prediction
A Quantum Long Short-Term Memory network achieved statistical parity with a classical LSTM boasting over 265,473 parameters despite containing just 2,105; this demonstrates that comparable performance to established methods doesn’t necessarily require complex quantum circuits. Previously, achieving such accuracy demanded sharply more computational resources from traditional models, a substantial barrier given financial data’s inherent complexity and limited historical depth.
The team at DHBW Stuttgart and DATEV eG found no evidence of an advantage in parameter efficiency or forecasting ability when using these hybrid quantum-classical recurrent neural networks compared to their purely classical counterparts, even considering the added intricacy of encoding information into quantum states via amplitude encoding. They extended their initial dataset of eighty years (96 months) by combining a Gaussian process with a hidden Markov Model, generating 960 additional synthetic months for stronger testing; this brought the total time-series length to 1,056 months.
This expanded series was then split chronologically: 739 months were used for training (70%), 148 for validation (15%) and another 159 served as a test set (15%). Above all, all classical scalers were fitted only on the original observed revenue data within the training window before consistent application across both real and simulated periods.
Rigorous benchmarking reveals limited near-term gains for hybrid quantum forecasting
The models offered no clear advantage over classical counterparts despite demonstrating modest gains over simple SARIMA forecasting methods; this raises questions about whether current hybrid architectures are optimally designed to capture financial time-series dynamics or if fundamental breakthroughs in quantum algorithms are needed. The team acknowledges limitations stemming from testing only “low-dimensional settings,” leaving unclear how these findings translate to more complex datasets with numerous influencing factors, a common reality within real-world finance. Still, failing to demonstrate an immediate advantage doesn’t diminish the value of these tests; it carefully benchmarks early quantum models against classical counterparts using rigorous financial data generation techniques.
This detailed comparison examining parameter efficiency and degradation rates establishes a key baseline for future work in the field of quantum finance. While Quantum Reservoir Computing was explored alongside the Quantum Long Short-Term Memory (QLSTM), performance remained comparable to standard techniques like LSTM and traditional reservoir computing. The implications extend beyond simply confirming parity, highlighting the need for further investigation into how these architectures perform with more complex datasets, potentially revealing scenarios where their unique capabilities could shine through or pinpointing specific areas requiring algorithmic refinement before practical benefits materialize.
This research compared hybrid quantum-classical recurrent neural networks, specifically a Quantum Long Short-Term Memory network and Quantum Reservoir Computing, to established classical models in forecasting monthly revenue data extended to 1056 months. While both quantum approaches improved upon SARIMA forecasts by approximately 2% in root mean squared error (RMSE), they did not outperform equivalently sized classical LSTM or reservoir computing models; even the QLSTM’s parameter efficiency was not demonstrated given its size of 2,105 parameters. The study establishes a baseline for future work, suggesting that further investigation is needed to determine if these architectures can offer advantages with more complex financial datasets.
👉 More information
🗞 Quantum Recurrent Architectures for Financial Time Series Prediction
✍️ Danyal Maheshwari and Gerhard Hellstern (DHBW Stuttgart); Martin Zaefferer (DHBW Ravensburg); Martin Braun and Tanja Döhler (Germany*Correspondence: Danyal Maheshwari)
🧠 ArXiv: https://arxiv.org/abs/2611.15298




See today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals.
