Trained, Frozen, or Absent: Identifying What a Quantum Layer Contributes to Cross-Sectional Equity Ranking
Authors: Chi-Sheng Chen, Xinyu Zhang, En-Jui Kuo, Rong Fu, Qiuzhe Xie, Fan Zhang
Organizations: Beth Israel Deaconess Medical Center & Harvard Medical School, Boston, MA 02115, USA · Brain Research Center, National Tsing Hua University, Hsinchu 300044, Taiwan · Luddy School of Informatics, Computing, and Engineering, Indiana University Bloomington, Bloomington, IN 47405 · Department of Electrophysics, National Yang Ming Chiao Tung University, Hsinchu 30010, Taiwan · Institute of Electronics Engineering, National Taiwan University, Taipei 106319, Taiwan · Department of Mathematics, Boise State University, Boise, ID 83702, USA
Hybrid quantumclassical models are increasingly proposed for nancial prediction, but reported gains are rarely separated from the classical components around them. We study a hybrid quantum temporal convolutional network for daily cross-sectional ranking of 2000 Tokyo Stock Exchange equities under a protocol built to make the quantum contribution identiable: six conditions (circuit trained, frozen or absent, a parameter-matched classical surrogate, and the circuit alone, trained or frozen) through one code path; four walk-forward folds with ve seeds; bootstrap intervals, Deated Sharpe Ratios, backtest-overtting probabilities and family-wise error control; ve classical baselines including a gradient-boosted tree; and a locked holdout. On 500 out-of-sample days two one-sided tests place the trained-versus-frozen dierence inside 0.03 annualised Sharpe (p = 0.0055, minimum detectable eect 0.02): the circuit is equivalent to a frozen one, not merely indistinguishable from it. With the circuit as the head's only input the two are equivalent inside 0.25; against the network with no circuit the design resolves only 0.50. No condition is distinguishable from zero (best De-ated Sharpe Ratio 0.17). The holdout reproduces the ablation on a protable window, and 2000 US equities reproduce it where every condition loses money.
Figures & tables
Adjustment
∣r∣>30%
∣r∣>50%
max∣r∣
None (raw closes)
718
565
20.09
Forward cumulative product
1418
1091
23.43
Backward cumulative product
16
3
1.12
Table 1: Artefactual daily moves under three adjustment conventions.
Group
Definition
n
Momentum ( momn )
P~t/P~t−n−1
1, 2, 5, 10, 20
Volatility ( voln )
rolling standard deviation of r
5, 10, 20
Standardised return ( ret_zn )
(rt−rˉn)/σn(r)
5, 10, 20
Log volume
log(1+Vt)
–
Log dollar volume
log(1+Dt)
–
Dollar-volume z -score ( dv_zn )
rolling z -score of log(1+Dt)
5, 10, 20
Table 2: The 21 input features. All rolling statistics use a window of exactly n trading days and are undefined until n observations exist.
Batch
Per-sample loop (s)
Batched (s)
Speed-up
max∣Δ∣
64
0.765
0.023
33.1 ×
0
256
3.218
0.050
64.2 ×
0
512
8.635
0.120
71.7 ×
0
Table 3: Batched against per-sample circuit evaluation at n=8 , L=3 , 18 parameters. Median of three repeats after warm-up; the machine was under scheduler control. Outputs agree exactly in double precision.
Figure 1: Annualised Sharpe ratio of each condition on the pooled 500-day walk-forward track, with 95% stationary-bootstrap confidence intervals. The four hybrid conditions differ only in what occupies the quantum position.
Condition
Sharpe
95% CI
p
DSR
Seed sd
Max DD
Transformer baseline
+0.107
[-1.09, +1.27]
0.85
0.17
0.31
-0.128
QTCNN, no circuit
-0.006
[-1.21, +1.13]
0.99
0.14
0.20
-0.164
TCN baseline
-0.085
[-1.28, +1.11]
0.88
0.12
0.47
-0.163
QTCNN, trained circuit
-0.088
[-1.27, +1.04]
0.87
0.11
0.14
-0.176
QTCNN, frozen circuit
-0.096
[-1.29, +1.03]
0.86
0.11
0.15
-0.178
QTCNN, circuit only
-0.124
[-1.34, +0.95]
0.82
0.11
0.22
-0.148
Table 4: Out-of-sample performance on the pooled walk-forward track ( 500 trading days, five seeds per condition). Confidence intervals are 95% stationary bootstrap; DSR is the Deflated Sharpe Ratio deflated by the 55 configurations evaluated.
Condition
Sharpe
95% CI
p
IC ( t )
Δ (s.e.)
TOST
No circuit
−0.006
[ −1.21 , +1.13 ]
0.99
0.0082 (0.99)
–
–
Frozen circuit
−0.190
[ −1.48 , +0.97 ]
0.75
0.0084 (1.04)
−0.184 (0.160)
–
Trained circuit
−0.222
[ −1.48 , +0.95 ]
0.71
0.0077 (0.97)
−0.032 (0.091)
0.25
Table 5: Second ansatz family (data re-uploading, CNOT ring, mean of Z over all wires) on the Tokyo walk-forward track, 500 days, five seeds; the no-circuit control is the one of Table 4 . Bootstrap 95% intervals and p against zero; information coefficient with its Newey–West t ; Δ is the pooled Sharpe difference from the row above it with its paired standard error; TOST gives the smallest of the margins 0.03 , 0.10 , 0.25 at which equivalence holds at 5% .
Figure 2: Per-cell Sharpe ratio, trained circuit against frozen circuit. The dashed line is the identity.
JPX
US
Condition
IC
tIC
Mono.
IC
tIC
Mono.
QTCNN, trained circuit
0.0083
1.01
0.054
0.0033
0.30
-0.376
QTCNN, frozen circuit
0.0083
1.01
0.048
0.0033
0.30
-0.379
QTCNN, classical surrogate
0.0073
0.91
0.080
0.0023
0.24
-0.432
QTCNN, no circuit
0.0082
0.99
0.114
0.0033
0.32
-0.336
QTCNN, circuit only
0.0075
0.91
0.115
–
–
–
Table 6: Ranking skill on both panels: the mean daily information coefficient, its Newey–West t -statistic, and the Spearman correlation between score decile and realised mean forward return as a monotonicity measure. Means over the twenty (JPX) and twenty-five (US) cells per condition.
Figure 3: Sharpe ratio per 125-day test block, one point per seed, for the trained-circuit model and the no-circuit control. Bars mark seed means.
Figure 4: Annualised Sharpe net of round-trip transaction costs. Net Sharpe is linear in cost at fixed turnover; dotted verticals mark break-even.
Condition
Sharpe
95% CI
p
DSR
Seed sd
Max DD
LSTM baseline
-0.655
[-1.83, +0.50]
0.28
0.00
0.55
-0.402
LightGBM baseline
-0.672
[-2.00, +0.58]
0.32
0.00
0.32
-0.229
Transformer baseline
-1.158
[-2.34, +0.09]
0.087
0.00
0.42
-0.503
QTCNN, no circuit
-1.441
[-2.75, -0.05]
0.052
0.00
0.16
-0.532
QTCNN, frozen circuit
-1.463
[-2.75, -0.09]
0.046
0.00
0.23
-0.535
QTCNN, trained circuit
-1.464
[-2.74, -0.09]
0.043
0.00
0.22
-0.535
Table 7: Out-of-sample performance on the pooled walk-forward track for the US panel ( 625 trading days, five seeds per condition). Confidence intervals are 95% stationary bootstrap; DSR is the Deflated Sharpe Ratio deflated by the 55 configurations evaluated.
Model
Sharpe
95% CI
p
DSR
IC
tIC
B/E (bps)
QTCNN (reference)
1.80
[-0.03, 3.79]
0.07
0.54
0.0016
0.42
7.4
Transformer
0.46
[-0.88, 2.04]
0.53
0.11
0.0025
0.65
2.7
QLSTM
0.13
[-1.42, 1.75]
0.85
0.06
0.0132
2.37
0.8
LightGBM ranker
-0.32
[-2.34, 1.51]
0.78
0.02
0.0060
1.52
–
QNN
-0.43
[-2.31, 1.42]
0.66
0.02
0.0148
2.10
–
MLP
-0.56
[-2.24, 1.39]
0.60
0.01
0.0134
2.57
–
Table 8: The reference implementation’s nine models, re-scored on 241 out-of-sample days (2020-12-10 to 2021-12-03) with annualised metrics, 95% stationary-bootstrap intervals, Deflated Sharpe Ratios and Newey–West information-coefficient t -statistics. Break-even is omitted where the gross return is not positive. Nothing else about the pipeline was changed.
Figure 5: Circuit-level diagnostics at eight qubits and three layers. (a) Gradient variance against register width. (b) Rank preservation against shot budget at three input distributions; faint lines are the twenty trained cells.
Random angles
Initialised encoder
Trained cells
Shots per security
(sd 0.443 )
(sd 0.018 )
median (sd 0.10 )
range
256
0.991
0.56
0.921
0.33–0.990
1 024
0.997
0.79
0.972
0.53–0.997
8 192
0.9996
0.96
0.995
0.83–0.9995
32 768
–
0.989
0.998
0.94–0.9998
131 072
–
0.997
0.9995
0.98–0.9999
Table 9: Rank preservation against shot budget at three input distributions: uniform random angles; real feature windows through an initialised encoder (12 dates up to 8192 shots and 6 further dates above, 300 securities each); and real windows through each of the twenty trained cells on dates from its own test window (3 dates per cell, 300 securities each). Entries are within-date Spearman correlations between finite-shot and analytic scores; for the trained cells the median over cells and the range across cells are given.
Condition
Sharpe
Seed sd
p
pRW
DSR
Break-even (bps)
QTCNN, no circuit
+1.83
0.84
0.07
–
0.42
3.7
QTCNN, classical surrogate
+1.57
1.08
0.18
0.95
0.34
5.5
QTCNN, frozen circuit
+1.51
1.05
0.17
0.95
0.33
4.4
QTCNN, trained circuit
+1.49
1.05
0.18
0.96
0.32
4.4
TCN baseline
+1.37
0.77
0.22
0.95
0.29
3.0
Transformer baseline
+0.68
0.87
0.47
0.54
0.15
2.1
Table 10: Locked holdout, 135 trading days. Uncorrected bootstrap p against zero; Romano–Wolf p across the eight contrasts with the no-circuit control on this window (the two circuit-only conditions and the second ansatz family were not); DSR deflated by the 55 configurations (eleven models, five seeds) the study evaluated, using the empirical dispersion of their Sharpe ratios on this window (sd 0.92 , expected maximum under the null 2.12 ); break-even computed on the pooled track, undefined where the gross return is not positive.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Condition
September
November
Cell corr.
JPX walk-forward
trained circuit
−0.07
−0.09
0.78
frozen circuit
−0.07
−0.10
0.78
classical surrogate
−0.04
−0.22
0.79
no circuit
+0.01
−0.01
0.44
circuit only
−0.25
−0.12
0.82
Appendix
Table 11: The first (September 2026) and the reproducible (November 2026) run of every network condition that was re-run: pooled-track annualised Sharpe ratio of each, and the correlation between the two runs’ per-cell Sharpe ratios (not given for the five-cell holdout). The last line of each panel is the trained-minus-frozen pooled Sharpe difference with its paired standard error and the two one-sided test against a 0.03 margin.
Hybrid quantum-classical machine learning combines quantum circuits or annealers with classical models for learning, prediction, generation, and optimization.
Financial time-series forecasting is still difficult owing to the non-stationary nature of the market, regime changes, and catastrophic forgetting, which makes conventional deep learning models hard to adapt in the long term. To achieve uncertainty-aware robust financial forecasting in the face of changing market conditions, this study proposes a Hybrid Quantum-Classical Neural Network with Continual Learning (HQNN-CL). The proposed framework combines Gaussian Hidden Markov Model (HMM) based regime detection, BiLSTM temporal encoding, adaptive-depth parameterized quantum circuits, Elastic Weight Consolidation (EWC), experience replay, and probabilistic forecasting in a single architecture. The framework was tested on four real-time stock datasets: Microsoft, Systems Limited, Teradata and Tata Consultancy Services, and compared with seven state-of-the-art forecasting models, including recurrent and Transformer-based models. The experimental results show that HQNN-CL has outperformed other methods on the Systems Limited (RMSE = 3.0149) and Teradata (RMSE = 0.8909) datasets and is within 0.32% and 0.40% of the best baseline method for Microsoft and Tata Consultancy Services, respectively. The proposed framework also generates well-calibrated probabilistic forecasts within the range of 89.5% to 96.5% empirical coverage rate, has zero catastrophic forgetting (BWT = 0.0000) for all sequential learning tasks, and consistently shows positive forward transfer. Ablation and statistical significance analyses also validate the performance of adaptive quantum feature learning and continual adaptation under heterogeneous market conditions, regime-wise.
Muhammad Atif Saeed, Akhtar Jamil, Subhan Ullah
Department of Artificial Intelligence & Data Science, National University of Computer & Emerging Sciences, Islamabad, Pakistan · Department of Computer Science, National University of Computer & Emerging Sciences, Islamabad, Pakistan · Department of Cyber Security, National University of Computer & Emerging Sciences, Islamabad, Pakistan
Quantum machine learning, and quantum neural networks (QNNs) in particular, are advancing fields with growing potential. Although systematic comparisons of QNN configurations have been explored primarily for classification tasks, comparatively little attention has been given to regression problems, particularly financial time series forecasting. This study presents a large-scale systematic comparative evaluation of QNN component configurations for financial time series forecasting, using the GBP/USD spot exchange rate as a case study. A grid search across encoding methods, ansatz designs, qubit counts, layer depths, and cost functions yields 1,368 distinct model configurations, each evaluated in terms of prediction accuracy, computational cost, and convergence behaviour. The results reveal unique insights into how the choice of methods influences performance, such as that gate selection and arrangement are more critical to model success than raw parameter count, and that entanglement is a system-level property of the full circuit rather than solely at the ansatz level. The best-performing QNN configuration achieves an R2 score of 0.985, outperforming a classical BiLSTM baseline. Additionally, the impact of real quantum hardware noise is assessed through execution on the IQM Emerald device, revealing that gate errors and decoherence represent a significant barrier to practical deployment, with gate selection and circuit depth identified as key determinants of hardware noise resilience. Overall, the findings provide practical architectural guidance for QNN design and establish a baseline characterisation of QNN noise sensitivity on near-term quantum devices.
Jack Waller, Xing Liang, Dimitrios Makris +1
School of Computer Science and Mathematics, Kingston University London, London, KT1 2EE, U.K.. · Department of Electronics and Electrical Engineering, Brunel University of London, London, UB8 3PH, U.K..
Accurate financial volatility forecasting is crucial but challenged by the non-linear, highly correlated nature of market data. Recently, quantum computing has emerged as a promising paradigm for solving complex high-dimensional sampling problems. To harness this, we propose a novel hybrid framework combining the temporal representation power of classical neural networks with the distribution-learning capabilities of quantum models. Specifically, we integrate a Long Short-Term Memory (LSTM) network with a Quantum Circuit Born Machine (QCBM). The LSTM extracts dynamic features, while the QCBM acts as a learnable generative prior modeling complex market distributions to guide forecasting. Evaluated on 5-minute high-frequency data from the SSE Composite and CSI 300 indices, our model significantly outperforms a classical LSTM baseline across MSE, RMSE, and QLIKE metrics. Furthermore, by introducing a stochastic Drop-Prior" mechanism during training, the LSTM implicitly distills structured information from the quantum prior. This establishes a pragmatic paradigm of quantum-assisted training with classical-efficient inference", whereby the model retains its quantum-enhanced accuracy even when the quantum module is entirely disabled during deployment. This demonstrates a practical pathway for leveraging quantum computing to enhance classical models without real-time quantum inference latency.