Systemic-risk monitoring frameworks are largely built on absolute Pearson correlation networks: Assets are linked if their returns co-move on average, and "systemic hubs" are defined by high degree. Such approaches implicitly assume (ⅰ) a single contagion timescale and (ⅱ) stability of dependence, even though crises typically unfold in layers: A fast equity/volatility unwind, followed by slower stress in funding, FX, rates, and commodities. We proposed a stress-aware, multiscale alternative, and constructed the multiscale coherence–entropy centrality (MCEC) network in which (a) an edge between two assets exists only if their wavelet coherence is statistically significant and persistent across adjacent frequency bands, and (b) node importance is an entropy-weighted multi-horizon strength that is high only if an asset is strongly connected and active across time scales. We then generated a synthetic stressed panel by shocking all assets with a common heavy-tailed t-copula draw scaled by GARCH(1,1) volatilities, and compared MCEC to a traditional absolute-correlation backbone using 2021–2024 data. We reported three findings that are directly relevant for macroprudential supervision. First, under stress, the MCEC network reallocated centrality toward canonical stress transmitters (U.S. equity benchmarks, implied volatility (VIX), dollar/FX, long-term yields, crude oil, and gold), while ordinary correlation networks continued to present a single equity-dominated block. Second, MCEC delivered higher ex-ante classification performance (AUC) in identifying those transmitters even before the stressed regime was applied, indicating early-warning value. Third, MCEC made the stress-driven rewiring of cross-market spillover channels explicit across horizons rather than treating dependence as static.

We introduce uniform, one-sided conformal prediction bands for forward realized volatility (FRV) paths that control the entire trajectory up to a fixed horizon $H$ with finite-sample, distribution-free marginal validity. The construction is model-agnostic: A monotone (isotonic) baseline across horizons and robust per-horizon scales are fitted to the training data; a scaled sup-norm score calibrated on a chronological holdout yields the uniform envelope. To address serial dependence, we employ block-maxima calibration; for regime sensitivity, we add group conditional (Mondrian) variants based on training-only state variables. On eight liquid assets, the method achieves conservative uniform coverage while adapting the width to tail risk: Bands are widest for crypto and oil, and tightest for broad equities and treasuries. A simple operational law emerges, namely that the mean one-sided width grows approximately with $\sqrt{H}$, turning the horizon design into a transparent safety-tightness trade-off. Practical guidance is given as follows: $\alpha \approx 0.05$ is a reliable default for thinner tails, while $\alpha \in [0.05,0.025]$ increases safety where bursts are frequent. Relative to parametric benchmarks such as heterogeneous autoregressive realized volatility (HAR-RV) and generalized autoregressive conditional heteroskedasticity (GARCH) models, our bands remain valid across regimes and stay width competitive in calmer markets. The algorithm is linear in $nH$ and agrees with deployment diagnostics. Overall, uniform FRV envelopes provide an operationally transparent, model-agnostic tool for pathwise volatility control, with tunable conservatism and simple extensions for dependence, covariate shift, and cross-split stability.