Can Generative AI Preserve Tail-Risk Networks? Evidence from Cryptocurrency Markets

Abstract Generated financial scenarios inform risk estimation and portfolio decisions, but agreement with standard distributional diagnostics does not show that a generator preserves the structure a decision depends on. We evaluate generated scenarios in three layers: conventional distributional and risk diagnostics; preservation of a conditional lower-tail network, judged against the sampling uncertainty of the empirical network; and realised outcomes of an allocation rule that uses the network’s node rankings. Using hourly data for ten cryptocurrencies over 2020–2025, we compare a stationary bootstrap, a GARCH-t copula model and a denoising diffusion probabilistic model (DDPM). The layers yield different conclusions about these generators. The discrepancy between the aggregated GARCH-t network and the observed network exceeds the bootstrap reference band at all 16 evaluation refits, yet overall distributional discrepancy stays within its band in 96 of the 157 GARCH-t panels that fall outside both the conventional-diagnostic and network bands. DDPM meets the joint conventional-diagnostic criterion in no panel but meets the single-panel network criterion in 120 of 160. Network differences change allocation weights and volatility exposure. After ex ante risk matching, differences in daily expected shortfall relative to the historical-network rule range from −0.09 to −0.02 percentage points, and every paired 95% interval contains zero. Controlled simulations show when ranking errors are costly: if tail hubs differ mainly in asymmetric downside dependence, a correlation-matching copula reverses the hub ranking and raises population expected shortfall by 3.1% relative to the true-network benchmark. Conventional diagnostics, network preservation and allocation outcomes give distinct assessments of these generators. JEL classification: C45 · C58 · G11 · G32