Crypto market dynamics in stressful conditions

ABSTRACT

Understanding market liquidity and trading dynamics in one of the most innovative and volatile markets in the world, is crucial from the standpoint of both regulators and investors. In contrast to stocks, very little is known about the functioning of cryptos around extreme returns (ERs). Using high-frequency order-book and trade data for the 8 most widespread cryptos on 16 trading platforms over three years, we examine the contemporaneous and lagged influence of trading activity and liquidity on the occurrence of extreme returns (ERs) in a logistic regression framework adapted to rare events. Despite its huge volatility, we show that the trading and liquidity dynamics on the crypto market around ERs is not orthogonal to what traditional markets experience in stressful conditions. The number of trades is a particularly robust driver to explain the occurrence of ERs, followed by the relative spread. The same drivers are identified for traditional markets.

JEL CODES:

• G15

JEL CLASSIFICATION:

• Liquidity
• trading
• shocks
• bubble
• bitcoin
• cryptos

Acknowledgement

We thank the Editor, two anonymous reviewers, Thiago Winkler Alves, Neharika Sobti, Kotaro Miwa, Catherine D’Hondt, Yue Zhang, Ian W. Marsh, Rudy De Winne, Gunther Wuyts, as well as participants at the Australasian Finance and Banking Conference (Sydney, AUS), New Zealand Finance Meeting (Auckland, NZ), and ISCEF (Paris, FR). Christophe Desagre gratefully acknowledges financial support from the LFIN research center and the FNRS. Any errors are the full responsibility of the authors.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Supplemental data

Supplemental data for this article can be accessed online at https://doi.org/10.1080/00036846.2022.2108754.

Notes

¹ There is debate on the best way to value, price, or even classify cryptos. Still, cryptos do not behave like commodities, whether scarce or not. Even bitcoin, which was initially dubbed `digital gold’ because of its fixed-supply scarcity, is not a safe haven. On the contrary, cryptos are very sensitive to the level of risk aversion and seem to increasingly behave like small-cap, tech-oriented, publicly-quoted companies whose returns depend on the value of their intangible assets. In the case of cryptos, the intangible assets are their brand and in some cases their underlying technology. Interestingly, we find similarities in liquidity and trading dynamics between cryptos and stocks.

² The choice of this threshold is also in line with the literature on extreme risk. For example, standard measures of risk, such as the VaR and CVaR, are typically estimated at the 99% threshold. This is the default threshold recommended by the Basel Accord which requires large international banks to hold regulatory capital for the trading book based on a 99% VaR over a 10-day holding period. This threshold is also used in recent papers such as Ji et al. (2020).

³ Market share is determined by the relative proportion of trades on each platform.

⁴ Source: Forbes (2014). Bitcoin’s Mt. Gox Goes Offline, Loses $409 M – Recovery Steps and Taking Your Tax Losses. https://bit.ly/3dqVsgx.

⁵ We report the same figure for other cryptos, i.e. BCHUSD, EOSUSD, ETHUSD, LTCUSD, XLMUSD, XMRUSD, and XRPUSD, in the online Appendix (Figure 1).

⁶ As a robustness check, we also compute the annualized daily volatility based on 288 5-minute intervals. The Pearson correlation coefficient between both measures is equal to 94%.:

⁷ We report the same figure for other cryptos, i.e. BCHUSD, EOSUSD, ETHUSD, LTCUSD, XLMUSD, XMRUSD, and XRPUSD, in the online Appendix (Figure 2).

⁸ Interested readers should refer to Heinze and Schemper (2002) and Heinze (2006).

⁹ We conducted these analyses with and without fixed effects for the LOGIT specifications and the results are not sensitive to their inclusion or removal.

¹⁰ Figure 3 in the online Appendix demonstrates these dynamics graphically by depicting the distribution of volatility estimates across the three time periods.

¹¹ Dueker (2005) proposes a QUAL VAR methodology to forecast a binary variable with a VAR. However, El-Shagi and Von Schweinitz (2016, 293) argue that this methodology is `inadvisable when the chain of causality matters’.

¹² We use the $\mathrm{\Delta }$ to indicate the first-difference of a variable.

¹³ More detailed results are available upon request. The conclusions were identical with the standard Augmented Dicky-Fuller (ADF), Phillips-Perron (PP), and Kwiatkowski – Phillips–Schmidt – Shin (KPSS) tests.

¹⁴ We have also replicated this analysis by sub-periods, i.e. before, during, and after the bubble and in low, medium, and high volatility regimes. The results are highly similar and the figures are available upon request.

¹⁵ Causality in the sense of Granger does not necessarily mean that there is a strict chain of causality existing between two series. Granger causality is associated to the notion of time precedence, i.e. it tests whether one series precedes another series based on a number of lags and a time sequence. The nature of this precedence can be the direct consequence of a causal relationship, the result of pure luck, or a mixture of both effects. Granger causality tests cannot assess to which of these categories the highlighted relationship belongs.

¹⁶ Detailed results are available upon request. To complement this analysis, Figure 4 presents the results of the impulse response functions in the online Appendix. We use the same four variables, i.e. number of trades, relative spread, imbalance, and absolute log returns. A shock in the absolute log return tend to increase the spread, which is in accordance with our results. We also observe that a shock in the spread increases the absolute log return, suggesting a bivariate relationship between the variables. Finally, a shock in the number of trades also positively impacts the absolute log return. Again, the reaction of the dynamic system in response to external changes is sound as there is no evidence of market dysfunction.

¹⁷ We thank an anonymous reviewer for suggesting to incorporate this analysis in the paper.

¹⁸ The empirical results are available in the online Appendix in Tables A1 to A3.

¹⁹ The results of the contemporaneous model as well as the separate outcomes for downward and upward ERs are not presented here but available upon request. These results are also highly consistent with the ones presented in the previous Sections.