When Small Coins Roar: Rethinking Systemic Risk in Crypto Volatility Forecasting

In traditional finance, systemic risk is often linked to size — the bigger the institution, the bigger the threat. But in crypto? The rules are different. A recent paper from researchers at Jinan University rewrites the forecasting playbook by demonstrating that systemic influence in crypto markets is more about network positioning than market cap.

The authors introduce a state-adaptive volatility model that integrates multi-scale realized volatility measures (like semivariance and jump components) with time-varying quantile spillovers, producing a high-resolution view of inter-asset contagion — especially under stress.

Beyond Averages: Volatility is Asymmetric and Regime-Dependent

Using 5-minute data from Binance across six major cryptocurrencies (BTC, ETH, XRP, LTC, DASH, XLM), the authors analyze volatility through the lens of quantiles. Instead of assuming a one-size-fits-all transmission mechanism, their framework adapts to whether the market is in a low, normal, or high volatility state.

Here’s how it works:

1. Identify market regimes using volatility quantiles (e.g., top 5% = high volatility).
2. Find the most influential risk transmitter to the target asset (like BTC) in each regime.
3. Feed this source’s lagged volatility into a HAR model, but only if the market is in the corresponding regime.

This dynamic selection — encoded in what they call the state-adaptive spillover variable — allows the model to reflect evolving relationships, like XRP becoming a key transmitter during volatile periods, even if it’s a smaller-cap asset.

Why Log-HAR Alone Isn’t Enough

The standard HAR-RV model (Corsi, 2009) decomposes volatility into daily, weekly, and monthly components, capturing long memory. A common upgrade is the Log-HAR, which applies a log transform to smooth extreme variations. Still, it assumes fixed spillover channels.

By contrast, the proposed SA-Log-HAR model integrates the state-adaptive variable into the Log-HAR structure, enabling it to condition on volatility regime and source. The authors test 12 variants — including Lasso-regularized versions to control overfitting — across multiple horizons and features:

Key Insight: Size Doesn’t Equal Spillover Power

Perhaps the most striking finding is a structural asymmetry: ETH, despite being the second-largest crypto, often acts as a net receiver of volatility, while XRP and LTC frequently transmit shocks to BTC. This pattern holds across multiple volatility features and quantile levels.

Example: Net Spillover to BTC by Quantile (RS Feature)

XRP’s role as a high-quantile shock transmitter is especially pronounced, hinting at its central role in systemic risk during stress periods.

Forecasting Results: SA Models Dominate

The authors use out-of-sample testing over both 300- and 500-day periods and multiple evaluation metrics, including MSE, MAE, RMSE, QLIKE, and the Campbell-Thompson R². Results show that:

• SA-Log-HAR-RS (semivariance version) consistently ranks top for 1- and 5-step forecasts.
• Lasso-SA-Log-HAR-RS achieves the highest R² of 0.776, outperforming all HAR and GARCH variants.
• GARCH models underperform, sometimes producing negative out-of-sample R² — a signal that traditional volatility models are misaligned with crypto’s complex dynamics.

Practical Takeaways

• For traders: Tail risks are not just fatter — they’re directionally amplified. Use spillover-aware models for position sizing.
• For risk managers: Don’t trust market cap alone. Influence in the crypto risk network is regime- and quantile-dependent.
• For researchers: The fusion of quantile spillovers and adaptive selection opens new paths for regime-aware forecasting.

Closing Thought

This paper pushes volatility modeling forward not just by improving forecast accuracy, but by reshaping our intuition of where risk comes from in crypto. In this market, the loudest signal may come from the smallest coin — but only when the market’s screaming.

Cognaptus: Automate the Present, Incubate the Future.