Assessing the Credit Risk of Crypto-Assets Using Daily Range Volatility Models

Research report

TLDR

Range-based volatility forecasts are especially useful in short term volatility forecasting contexts and reasonably effective at longer time horizons. This paper utilizes range based approaches to forecast the likelihood of a token “dying” over a certain time horizon.

Range based volatility measures are especially effective for forecasting purposes as opposed to return based approaches in short-term volatility forecasting use cases and also long term default/credit risk models.
Infinite time horizon credit risk models for insolvency risk, broadly known as zero-price probability models (ZPPs) are well-suited to crypto applications, particularly for non-blue chip tokens.

Concrete utilizes a myriad of infinite time horizon default risk models, primarily a modified Cramer-Lundberg model, which falls under the class of broader ruin theory approaches. A broader ZPP approach seems to be better suited for our use case, given the infrequency of death/insolvency events.
This model remains highly academic and requires considerable engineering and data science resources for execution/implementation. Cramer-Lundberg models remain the Pareto optimal approach assessing infinite time horizon insolvency risks.

The proposed, and four other, models are tested at a range of time horizons, both over and under 30 days: a simple random walk, heterogeneous autoregressive, autoregressive fractionalized integrated moving average, and conditional autoregressive range models.
The above 4 models are themselves modified slightly. For example, the CARR model is modified with certain GARCH models. The models aim to forecast the probability of a coin “dying” according to various definitions presented in the paper. This is analogous to ruin theory models (page 13).
The Cauchit model is identified as the best model, particularly when using the professional rule that defines a dead coin as one whose value drops below 1 cent. The ZPP computed using an MS-GARCH(1,1) model remains the best model when employing the simple random walk (pages 14-16).
A table is provided that includes AUC, Brier scores, models included in the MCS, and numerical convergence failures for three different criteria to classify a coin as dead or alive.