Quantum Recurrent Architectures for Financial Time Series Prediction
Organizations: Zentrum für Finanzwirtschaft, DHBW Stuttgart, Stuttgart, Germany · Zentrum für Digitale Innovationen, DHBW Ravensburg, Ravensburg, Germany · DATEV eG, Nürnberg, Germany
Abstract
This study evaluates quantum classical hybrid recurrent architectures for financial time-series forecasting: a Quantum Long Short-Term Memory (QLSTM) network and Quantum Reservoir Computing (QRC) with lag-based inputs. Classical lag vectors are encoded into quantum states via amplitude encoding, which compactly represents normalized observations but requires nontrivial state preparation; the QLSTM gates and the QRC reservoir are implemented with parameterized quantum circuits, while classical optimizers train the interface and readout layers. On a monthly revenue series extended to 1056 months by a Gaussian-process--hidden-Markov generator, with a fixed chronological 70/15/15 split and one-step-ahead forecasts, we compare both models against a full-size LSTM, a classical control matched in parameter budget and interface architecture, a small plain LSTM, and classical reservoir computing, over ten seeds and against persistence, seasonal-naive and SARIMA benchmark forecasts. The QLSTM, with 2,105 parameters, is statistically indistinguishable from an LSTM 126 times larger, but equally indistinguishable from the matched classical control and from a plain LSTM of the same budget, so its parameter efficiency cannot be attributed to the quantum circuits. The quantum reservoir is worse than a classical reservoir of equal feature dimension in all ten seeds and worse than SARIMA, whereas all recurrent models and both classical reservoirs improve on SARIMA by about 2% in RMSE. In a 24-month rolling-origin backtest the classical LSTMs degrade more slowly than the QLSTM. We find no evidence of a quantum accuracy or parameter-efficiency advantage in this low-dimensional setting.
Figures & tables
| Component | Purpose | Initialization | Constraint / limit |
|---|---|---|---|
| Rational quadratic | Slow trend and medium-scale variation | , | |
| Matérn (short-term) | Short-term variability | , fixed | |
| Matérn (long-term) | Smooth medium- to long-term variation | , fixed | |
| Periodic | Seasonality | , |
| Classical LSTM | QLSTM |
|---|---|
| Model | Quantum | Classical | Total |
|---|---|---|---|
| LSTM | 0 | 265,473 | 265,473 |
| QLSTM | 48 | 2,057 | 2,105 |
| cLSTM-S (matched control) | 0 | 2,097 | 2,097 |
| LSTM-S (small LSTM) | 0 | 2,038 | 2,038 |
| RC (linear readout) | 0 | 13 | 13 |
| RC (NN readout) | 0 | 225 | 225 |
| Model | Params | MSE | RMSE | MAE | MASE | Seeds SARIMA |
|---|---|---|---|---|---|---|
| LSTM | 265,473 | 0.00996 0.00002 | 0.0998 0.0001 | 0.0805 0.0003 | 0.625 0.002 | 10/10 |
| QLSTM | 2,105 | 0.01001 0.00015 | 0.1001 0.0007 | 0.0804 0.0007 | 0.624 0.005 | 9/10 |
| cLSTM-S (matched control) | 2,097 | 0.01000 0.00006 | 0.1000 0.0003 | 0.0802 0.0005 | 0.623 0.004 | 10/10 |
| LSTM-S (small LSTM) | 2,038 | 0.01001 0.00003 | 0.1000 0.0001 | 0.0807 0.0001 | 0.627 0.001 | 10/10 |
| Model | Params | MSE | RMSE | MAE | MASE | Seeds SARIMA |
|---|---|---|---|---|---|---|
| RC (linear readout) | 13 | 0.01002 0.00010 | 0.1001 0.0005 | 0.0797 0.0004 | 0.619 0.003 | 10/10 |
| RC (NN readout) | 225 | 0.00996 0.00018 | 0.0998 0.0009 | 0.0799 0.0008 | 0.621 0.006 | 10/10 |
| QRC (linear readout) | 13 | 0.01395 0.00287 | 0.1176 0.0116 | 0.0935 0.0079 | 0.726 0.061 | 0/10 |
| QRC (NN readout) | 225 | 0.01381 0.00272 | 0.1170 0.0110 | 0.0928 0.0079 | 0.721 0.061 | 0/10 |
| Model | MSE | RMSE | MAE | MASE |
|---|---|---|---|---|
| Persistence | 0.01830 | 0.1353 | 0.1099 | 0.854 |
| Seasonal-naive | 0.02784 | 0.1669 | 0.1331 | 1.034 |
| SARIMA | 0.01036 | 0.1018 | 0.0811 | 0.630 |
| A | B | A wins | (RMSE) | (MASE) | |
|---|---|---|---|---|---|
| LSTM | QLSTM | -0.0003 | 7/10 | 0.275 [0.551] | 0.275 |
| QLSTM | cLSTM-S | +0.0000 | 6/10 | 0.922 [0.922] | 0.625 |
| QRC-linear | RC-linear | +0.0175 | 0/10 | 0.002 [0.006] | 0.002 |
| QLSTM | LSTM-S | +0.0000 | 5/10 | 0.492 | 0.131 |
| LSTM | cLSTM-S | -0.0002 | 7/10 | 0.084 | 0.064 |
| QRC-nn | RC-nn | +0.0172 | 0/10 | 0.002 | 0.002 |
| Model | vs. SARIMA | Seeds better | (Wilcoxon, ) |
|---|---|---|---|
| LSTM | -0.0020 | 10/10 | 0.002 |
| QLSTM | -0.0018 | 9/10 | 0.004 |
| cLSTM-S (matched control) | -0.0018 | 10/10 | 0.002 |
| LSTM-S (small LSTM) | -0.0018 | 10/10 | 0.002 |
| RC (linear readout) | -0.0017 | 10/10 | 0.002 |
| RC (NN readout) | -0.0020 | 10/10 | 0.002 |
| Model | First 96 months | Months 1–739 | ||
|---|---|---|---|---|
| LSTM | 0.1005 0.0005 | 0.0998 0.0001 | -0.0007 | 0.002 |
| QLSTM | 0.1157 0.0148 | 0.1001 0.0007 | -0.0157 | 0.002 |
| cLSTM-S (matched control) | 0.1149 0.0147 | 0.1000 0.0003 | -0.0149 | 0.002 |
| LSTM-S (small LSTM) | 0.1006 0.0001 | 0.1000 0.0001 | -0.0005 | 0.002 |
| RC (linear readout) | 0.1025 0.0015 | 0.1001 0.0005 | -0.0025 | 0.004 |
| RC (NN readout) | 0.1005 0.0017 | 0.0998 0.0009 | -0.0007 | 0.027 |
| Model | – | Seeds SARIMA | |||||
|---|---|---|---|---|---|---|---|
| Persistence | 0.139 | 0.159 | 0.163 | 0.176 | 0.1652 | – | – |
| Seasonal-naive | 0.173 | 0.172 | 0.163 | 0.176 | 0.1780 | – | – |
| SARIMA | 0.104 | 0.134 | 0.166 | 0.176 | 0.1576 | – | – |
| LSTM | 0.102 | 0.122 | 0.135 | 0.130 | 0.1273 0.0014 | 10/10 | 0.002 |
| QLSTM | 0.102 | 0.123 | 0.139 | 0.139 | 0.1315 0.0027 | 10/10 | 0.002 |
| cLSTM-S (matched control) | 0.102 | 0.123 | 0.140 | 0.142 | 0.1331 0.0040 | 10/10 | 0.002 |
| Model | Epochs | s / epoch | Total s |
|---|---|---|---|
| LSTM | 63 | 0.57 | 35.0 |
| QLSTM | 87 | 2.44 | 208.5 |
| cLSTM-S (matched control) | 83 | 0.16 | 14.1 |
| LSTM-S (small LSTM) | 137 | 0.04 | 5.0 |
| RC (linear readout) | – | – | 0.01 |
| RC (NN readout) | – | – | 0.44 |
Time Series Forecasting
Time series forecasting predicts future values from ordered observations using statistical, neural, and probabilistic models to support decisions under uncertainty.
See all 930 papers