Volatility heredity concerns the information that past fluctuations in one series may carry about another after accounting for the target's own persistence. QS-TCN compares this dependence in both directions to estimate an order of temporal precedence. For each series, quasi-maximum likelihood estimation yields a persistence coefficient, and the matrix pencil method extracts temporal modes from the autocovariances of squared observations. These estimates define filters in a temporal convolutional network. The network models the target's squared observations from their own past and scores the association between the residuals and the filtered history of a candidate source. Differences between the two directional scores are aggregated into a global order. In an exploratory simulation with 100 systems in each of four volatility regimes, an ensemble of ten models trained with fixed seeds achieved higher mean ordering quality than CCF-AUC, a lead--lag method based on distance correlation, in every regime. Against CCF-AUC applied to squared standardized observations, the mean gains were 8.69 percentage points under positive linear coupling and 7.81 points when nonlinear coupling depended on the target's state. In the other two regimes, paired 95% bootstrap intervals for the gains included zero. Mean agreement with reference constraints was 71.55% across nine Finance systems and 54.02% across 27 simulated functional magnetic resonance imaging systems, compared with 50% expected under random ordering. In a panel of daily currency and gold returns, the estimated order placed the dollar--franc and dollar--yen exchange rates before gold.
Figures & tables
Data setting
Domain
Systems / assets
Length
Reference structure
Finance [ 24 , 25 ]
Simulated financial systems
9 datasets, 25 variables
T=4000
Known
fMRI [ 24 , 26 ]
Simulated BOLD systems
27 evaluated datasets, 5–50 nodes
T=50 – 5000
Known
Forex/Gold
Observed financial markets
7 assets
T=2384
No complete GT
Cryptocurrencies [ 36 , 38 ]
Observed crypto markets
3 or 6 assets
Study-dependent
Published directional relations
Table 1: Datasets used in the empirical evaluation. A known reference structure allows direct computation of ordering quality. Published market relations are used only for cross-study comparison and are not interpreted as complete causal ground truth.
Dataset
Pairs
QS-TCN A
VAR-LiNGAM
Mean OQ
SD
Min–Max
manyinputs
6
1.000
1.000
1.000
0.000
1.000–1.000
20_1A
18
0.778
0.500
0.806
0.054
0.722–0.889
20_1B
20
0.650
0.650
0.590
0.052
0.500–0.650
20_1C
19
0.526
0.421
0.532
0.017
0.526–0.579
20_1D
19
0.684
0.474
0.705
0.037
0.632–0.737
20_1E
17
0.824
0.765
0.782
0.088
0.647–0.882
Table 2: Finance ordering quality by dataset. QS-TCN A is the historical frozen order; VAR-LiNGAM is one deterministic order fitted on the same estimation prefix. Mean, standard deviation and range for QS-TCN are computed across the ten bootstrap seed schemes A–J.
Dataset
Pairs
QS-TCN A
VAR-LiNGAM
Mean OQ
SD
Min–Max
1
5
0.400
0.400
0.460
0.165
0.200–0.600
2
11
0.545
0.727
0.618
0.120
0.455–0.818
3
18
0.444
0.444
0.439
0.032
0.389–0.500
4
61
0.475
0.475
0.503
0.021
0.475–0.525
5
5
0.400
0.200
0.380
0.063
0.200–0.400
6
11
0.273
0.273
0.318
0.048
0.273–0.364
Table 3: fMRI ordering quality by dataset. QS-TCN A is the historical frozen order; VAR-LiNGAM is one deterministic order fitted on the same estimation prefix. Mean, standard deviation and range for QS-TCN are computed across the ten bootstrap seed schemes A–J.
Regime
QS-TCN
Native dCor
Squared-input dCor
Positive linear volatility
0.8595
0.5740
0.7726
Signed linear volatility
0.7428
0.5475
0.7202
Original SV mechanism
0.5729
0.4643
0.5414
Target-state nonlinear volatility
0.7772
0.5633
0.6991
Linear lead-lag control
1.0000
1.0000
1.0000
Table 4: Mean ordering quality on synthetic systems. The linear lead-lag control is scored against factor-delay precedence.
Regime
Comparator
Difference
95% interval
W/T/L
Positive linear
Native dCor
+0.2855
[+0.2351,+0.3351]
78/16/6
Squared-input dCor
+0.0869
[+0.0535,+0.1206]
45/42/13
Signed linear
Native dCor
+0.1953
[+0.1441,+0.2462]
67/14/19
Squared-input dCor
+0.0226
[−0.0291,+0.0758]
35/33/32
Original SV
Native dCor
+0.1086
[+0.0457,+0.1714]
53/16/31
Squared-input dCor
+0.0314
[−0.0229,+0.0857]
44/23/33
Table 5: Paired OQ differences (QS-TCN minus comparator) over the same 100 systems, with descriptive 95% system-bootstrap intervals. W/T/L counts the systems in which QS-TCN scores higher, equal or lower.
Directional relation
Cij
Rij
GBPUSD → USDCAD
0.0892
0.847
GOLD → GBPUSD
0.0691
0.775
GOLD → USDCAD
0.0958
0.759
USDCHF → USDCAD
1.2944
0.621
USDCHF → EURUSD
0.3230
0.589
USDJPY → GOLD
0.1116
0.450
Table 6: Selected directional relations from the frozen Forex/Gold QS-TCN analysis. Cij denotes the aggregated antisymmetric pairwise evidence and Rij its associated reported reliability.
Study
Published comparison object
QS-TCN result
Agreement
Ji et al. [ 36 ]
15 pairwise volatility-spillover directions and a separate net-spillover ranking
DASH ≺ BTC ≺ LTC ≺ XRP ≺ XLM ≺ ETH
Pairwise: 10/15≈0.667 , pperm=0.2347 ; net ranking: τ=0.60
Pre-COVID: ETH ≺ LTC ≺ BTC; COVID: BTC ≺ ETH ≺ LTC
2/5 pre-COVID; 2/3 COVID; combined agreement =4/8
Table 7: Comparison of QS-TCN cryptocurrency orderings with directional structures reported in independent studies. Concordance is computed only for relations that map unambiguously onto the corresponding asset set.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
A
B
C
D
E
F
G
H
I
J
Mean
manyinputs
1.000
1.000
1.000
1.000
1.000
1.000
1.000
1.000
1.000
1.000
1.000
20_1A
.778
.889
.833
.833
.778
.889
.722
.778
.778
.778
.806
20_1B
.650
.550
.650
.600
.550
.650
.550
.500
.600
.600
.590
20_1C
.526
.526
.526
.526
.579
.526
.526
.526
.526
.526
.532
20_1D
.684
.737
.737
.684
.737
.684
.737
.632
.684
.737
.705
20_1E
.824
.824
.647
.765
.824
.882
.647
.706
.882
.824
.782
Appendix
Table 8: Finance ordering quality for all ten frozen bootstrap seed schemes.
DS
A
B
C
D
E
F
G
H
I
J
Mean
SD
1
.400
.200
.600
.400
.200
.600
.600
.600
.600
.400
.460
.165
2
.545
.636
.455
.545
.636
.727
.818
.727
.636
.455
.618
.120
3
.444
.389
.444
.444
.444
.444
.444
.444
.389
.500
.439
.032
4
.475
.492
.525
.525
.475
.525
.492
.492
.525
.508
.503
.021
5
.400
.400
.400
.400
.400
.200
.400
.400
.400
.400
.380
.063
6
.273
.273
.273
.364
.273
.364
.364
.364
.364
.273
.318
.048
Appendix
Table 9: fMRI ordering quality for every frozen bootstrap seed scheme. The table gives the ten recorded seed-scheme scores underlying the fMRI summaries in Table 3 .
Period
Direction
Coefficient
p -value
Pre-COVID
BTC → ETH
−0.144
0.000
Pre-COVID
BTC → LTC
−0.129
0.000
Pre-COVID
ETH → BTC
−0.058
0.032
Pre-COVID
ETH → LTC
−0.269
0.000
Pre-COVID
LTC → ETH
1.028
0.000
COVID
BTC → ETH
0.037
0.000
Appendix
Table 10: Retained Yousaf–Ali coefficients under the direction convention stated in their Table 3 note.
Setting
Memory Cell
Directional module
Lag horizon
8
8
Free filters
12
12
Parameters
601
4,959
Optimizer
AdamW
AdamW
Learning rate
2×10−3
1.5×10−3
Weight decay
10−4
2×10−4
Appendix
Table 11: Architecture and optimization settings for the synthetic-comparison checkpoints. Memory parameters are frozen before directional training.
Faculty of Science and Technology, Beijing Normal-Hong Kong Baptist University, Zhuhai, China · Guangdong Provincial Key Lab of Interdisciplinary Research and Application for Data Science (BNBU), Zhuhai, China