Quantum machine learning models for sequential data face scalability challenges with complex multivariate signals. We introduce the Hybrid Quantum Temporal Convolutional Network (HQTCN), which combines classical temporal windowing with a quantum convolutional neural network core. By applying a shared quantum circuit across temporal windows, HQTCN captures long-range dependencies while achieving significant parameter reduction. Evaluated on synthetic NARMA sequences and high-dimensional EEG time-series, HQTCN performs competitively with classical baselines on univariate data and outperforms all baselines on multivariate tasks. The model demonstrates particular strength under data-limited conditions, maintaining high performance with substantially fewer parameters than conventional approaches. These results establish HQTCN as a parameter-efficient approach for multivariate time-series analysis.
Figures & tables
Fig. 1: Model Architectures. (a) A standalone QCNN with hierarchical quantum convolutional and pooling layers. (b) The proposed HQTCN, which applies a shared QCNN circuit over a temporal sliding window before aggregating the outputs.
Fig. 2: NARMA Time-Series Prediction Results. Comparison of HQTCN’s predictions (red) against the ground truth (gray) and other models on the NARMA test set. The zoomed-in view highlights the test-set performance.
Test Loss
# of Parameters
LSTM
0.0387 ± 0.0111
531
Transformer
0.0423 ± 0.0127
8,641
TCN
0.0380 ± 0.0032
10,785
QLSTM
0.0547 ± 0.0192
126
QCNN
0.0456 ± 0.0181
2,192
HQTCN
0.0487 ± 0.0193
312
TABLE I: NARMA Dataset Results. Comparison of mean test loss ( ± std. dev. over three seeds) and the number of trainable parameters for all models on the NARMA prediction task.
Performance-optimized
Parameter-constrained
Sample-constrained (N = 10)
Test AUROC
# of Parameters
Test AUROC
# of Parameters
Test AUROC
LSTM
0.7552 ± 0.0182
66,625
0.7657 ± 0.0178
5,265
0.6582 ± 0.0832
Transformer
0.7672 ± 0.0269
19,201
0.7618 ± 0.0261
7,729
0.6651 ± 0.0645
TCN
0.7815 ± 0.0296
49,985
0.7755 ± 0.0245
5,857
0.5875 ± 0.0528
QCNN
0.6649 ± 0.0342
127,760
0.6649 ± 0.0342
127,760
0.5112 ± 0.0072
HQTCN
0.7929 ± 0.0383
6,416
0.7929 ± 0.0383
6,416
0.6713 ± 0.0515
TABLE II: PhysioNet EEG Analysis Results. Test AUROC ( ± std. dev. over three seeds) and parameter counts for all models on the EEG classification task across performance-optimized, parameter-constrained, and sample-constrained (N=10) settings.
Dilation ( d )
Test AUROC
Receptive Field
d=1
0.7859±0.0509
12 time-steps
d=2
0.7882±0.0506
23 time-steps
d=3 (Baseline)
0.7929±0.0383
34 time-steps
d=4
0.7887±0.0479
45 time-steps
Depth ( L )
Test AUROC
Quantum Parameters
L=1
0.7888±0.0490
132
TABLE III: Ablation studies on PhysioNet EEG. We report Test AUROC ( ± std. dev. over three seeds) varying one hyperparameter while fixing others to baseline values ( d=3,L=2,n=8 ). Receptive field =d×(K−1)+1 , where K=12 .
Graduate Institute of Communication Engineering, National Taiwan University, Taipei, Taiwan · Brookhaven National Laboratory, AI & ML Department, Upton, NY, USA · Wells Fargo, New York, NY, USA