The paper introduces a model-free visibility-graph asymmetry fingerprint computed from maximal forward and backward horizontal-visibility distances on absolute returns, summarized by KS distance, variance difference, entropy difference and extreme-span ratios . The Monte Carlo study shows these geometric features sharply separate clustered GARCH dynamics from i.i.d. Gaussian noise and from shuffled-GARCH controls, with a simple linear classifier achieving ≈90% in-sample accuracy . Applied to daily S&P500 absolute returns, the method reveals a pronounced forward–backward imbalance—most notably a variance difference ΔVar far exceeding simulated GARCH values— and this asymmetry vanishes after shuffling, indicating a temporal (clustering) origin rather than a pure marginal effect .
速读内容
Method overview: HVG on absolute returns and asymmetry fingerprint
Construct horizontal visibility graph (HVG) on x_t = |r_t| and for each node compute maximal forward L^+(t) and backward L^-(t) visibility distances .
Form fingerprint (D_KS, ΔVar, ΔH, Δ_max, R) comparing empirical distributions of L^+ and L^- to capture directional imbalance .
Monte Carlo design and key simulation finding
Model
D_KS
△Var
max
R
△H
Financial (GARCH)
0.01852
-192.99
0.5570
0.9619
-0.0600
Nonfinancial (Gaussian)
0.01018
-10.06
0.0000
1.0039
+0.0000
Shuffled GARCH
0.01037
+20.23
-0.1755
1.0034
-0.0001
Simulate n=5000 paths per regime (T=6000); HVG fingerprint values averaged per path show GARCH paths occupy a distinct region with larger D_KS and dramatically different ΔVar compared with Gaussian and shuffled-GARCH .
Entropy asymmetry ΔH is clearly negative for GARCH and centered near zero for Gaussian and shuffled controls, reinforcing that clustering (not marginal tails) drives the effect .
Shuffle-control and interpretation
Random permutation of GARCH series preserves marginal |r_t| distribution but destroys serial dependence; after shuffling, asymmetry metrics collapse to Gaussian levels, demonstrating the asymmetry is driven by temporal clustering rather than heavy tails alone .
Empirical S&P500 results
Model
D_KS
△Var
max
R
△H
Real S&P500
0.01069
+14007.84
0
1.2086
+0.05300
Shuffled S&P500
0.00268
-4665.74
0
0.9472
-0.00213
The real S&P500 shows a very large positive ΔVar (~1.4×10^4) and positive ΔH and R>1, indicating a strong forward–backward imbalance in visibility spans that disappears after shuffling .
The magnitude of ΔVar in real data exceeds stationary GARCH simulations by orders of magnitude, consistent with multi-scale nonstationarity, regime shifts, and long-range effects in equity markets .
Classification check (consistency)
A logistic regression using the five asymmetry features separates clustered vs non-clustered paths with ≈90% in-sample accuracy, confirming the fingerprint contains discriminative directional information though classification is only a sanity check here .
思维导图
深度阅读
元数据与概览(引言与报告概览)
报告标题为 “VISIBILITY-GRAPH ASYMMETRY AS A STRUCTURAL INDICATOR OF VOLATILITY CLUSTERING”,作者为 Michał Sikorski,隶属华沙大学跨学科数学与计算建模中心,发布日期为 2025-12-03。