Anatomy of a Stablecoin’s failure: the Terra-Luna case

We use hourly closing price data from the CryptoCompare database (CryptoCompare,, 2022) for 61 cryptocurrencies from 01 May 2022 00:00 to 16 May 2022 23:00 (UTC)⁶⁶6The list of cryptocurrencies used in this paper and their corresponding sector according to taxonomy by Messari,, 2022, is provided as supplementary material.. This widely used proprietary data source excludes exchanges containing spurious volumes (see Alexander and Dakos,, 2020 and Vidal-Tomás,, 2022), ensuring a high standard of data quality. Specifically, we use data from the Kraken digital currency exchange (the $4^{th}$ largest exchange for traded volume according to CoinMarketCap, 2022b, ).

Figure 1 shows rescaled hourly closing prices for LUNA, UST and BTC. Dotted lines underline trigger events that led to the collapse: (a) 05 May 2022 12:00, (b) 07 May 2022 22:00, (c) 09 May 2022 14:00, (d) 11 May 2022 10:00. Since (a) 05 May 2022 12:00, a strong and persistent selling pressure was evident on BTC and LUNA, with negative hourly log-returns⁷⁷7Hourly log-return $r$ for crypto asset $i$ at time $t$ is defined as $r_{i,t}=\log(p_{t})/\log(p_{t-1})$, where $p$ is the closing price. at 13:00, 14:00 and 15:00, as reported in Table 1. On (b) 07 May 2022 22:00, UST lost its peg to USD for the first time as a consequence of the “liquidity pool attack” on Curve-3pool (Ashmore,, 2022; Chainanalysis,, 2022). In response to this event, the LFG intervened, defending the UST peg and recovering the price to around $\$0.995$. However, on (c) 09 May 2022 14:00, UST lost its peg for the second (and last) time, giving rise to the main decrease in both LUNA and UST prices. Finally, on (d) 11 May 10:00, Kwon, 2022a presented the last attempt to defend the peg endorsing the community proposal 1164. After a beneficial impact that led the UST price to $0.8$$, this announcement was interpreted by the market as the signal for the expected death of the Terra ecosystem, giving rise to the final crash in the market.

In order to analyse dependency structures in cryptocurrency market during Terra collapse, for each cryptocurrency $i$, we compute hourly log-returns $r_{i,t}$. Moreover, to analyse herding behaviour, we also compute equally-weighted (EW) market returns ($r_{m,t}$). Table 2 reports descriptive statistics for LUNA, UST, BTC and the EW market, which shows the negative tendency of the market during this period.

To provide a complete overview of Terra project’s collapse, we analyse public trades (i.e. transaction data) for BTC, LUNA and UST. Compared to hourly prices, which are retrieved from CryptoCompare database (CryptoCompare,, 2022), this new dataset is directly sourced from Kraken digital currency exchange using CCXT (Ccxt,, 2022) Python package. Figure 2 reports hourly imbalances: positive values denote a selling pressure, while negative ones denote a buying pressure.⁸⁸8In order to compute the imbalance, we firstly distinguish between public trades on the buy and sell side. We then multiply the volume of each transaction and the price at which it is executed, obtaining the transaction’s costs. We aggregate resulting transactions’ costs by hour. We finally subtract buy hourly transactions costs from sell hourly transactions costs. We observe interesting features that could be related to some of the events introduced in Section 1. First, on 05 May 2022 (a), we identify a strong, positive imbalance for BTC, with a value comparable to those detected during the main collapse of LUNA and UST (i.e. from 09 to 11 May 2022 (c-d)). We do not observe comparable high positive imbalance values for LUNA and UST on the same date. Therefore, even though we cannot confirm that short selling positions were opened against BTC as described by different social media sources (Hall,, 2022; Ashmore,, 2022; Locke,, 2022), we remark the presence of a considerable selling pressure on this cryptoasset. In other words, if short selling positions were truly opened against BTC, 05 May 2022 should be the most plausible day in which this happened.⁹⁹9In this line, we also observe higher BTC selling pressure, from (b) to (c). Thus, on these days, market actors could have spread more panic in the market by selling BTC. Second, focusing on the UST behaviour, we underline the existence of a remarkable selling pressure since 09 May 2022 (i.e. after the second UST de-pegging). This finding is particularly relevant, since the selling pressure after the first UST de-pegging (b) was remarkably lower than the second one, which could be in line with social media news when contending that attackers caused panic by selling whale-size UST holdings (Hall,, 2022; Ashmore,, 2022; Locke,, 2022). This strategy would be similar to the one used to bring down Iron Finance, given that attackers waited to the second de-pegging, taking advantage of the uncertainty that already dominated the market (Finematics,, 2021). Last but not least, it is relevant to emphasise that we cannot confirm the existence of a coordinated attack executed by related market agents. Indeed, the potential short selling against BTC and the Terra attack could have occurred at the same time by chance, in the context of a global economic uncertainty. In the case of Iron Finance collapse, no coordinated attack using a third cryptoasset was detected (Finematics,, 2021).

Anchor (Anchor,, 2022) was a savings protocol designed to allow users from Terra ecosystem to (i) lend capital (UST) and (ii) borrow capital (UST) using bonded assets (bAssets, i.e. tokens representing ownership of an asset on a Proof-of-Stake blockchain) as collateral (see Figure 3). Anchor protocol was at the base of the fast expansion of Terra project since it offered to its depositors up to 20% APY. This scheme was similar to the one used by Iron Finance, which offered more than 100% APY (Iron Finance, 2021b, ).

Looking at Table 3 it is possible to see how, from 17 June 2021 to 06 May 2022, total UST deposits in Anchor increased by 3826%¹⁰¹⁰10Data from Anchor protocol dashboard is available since 17 June 2021. No data are available on 07 May 2022.. Consequently, before the Terra collapse, this protocol kept the 75% of all the UST circulating supply, leaving the remaining 25% as a means of exchange. According to Platias et al., (2020), Anchor’s mechanism was originally designed to allow the borrower to pay less interest as the utilization ratio (borrowed deposits divided by total deposits) decreased, resulting in lower interests for the depositor. However, despite the decrease in the utilisation ratio, the deposit interest rate was kept around 20% since 2021 (Anchor Protocol,, 2021; Jung,, 2022). In addition, the deposit interest rate was computed considering the yields of all bAssets used as collateral for borrowing the stablecoin, while its value did not increase significantly since January 2022. The unsustainability of the protocol is confirmed by capital infusions into Anchor’s reserves in February 2022 (Kelly,, 2022) and by recurrent requests from community members for the modification of the interest rates’ mechanism (Anchor Forum,, 2022). Compared to fiat currencies, UST was mainly used for speculative purposes, which implied that it was more prone to extreme market events. Looking at Table 3, we observe that, from 06 May 2022 to 11 May 2022, the protocol lost around 9.5 billions in UST, i.e. the 69% of the deposits. Thus, once it became evident that both the Terra system and the high-interest rate mechanism were at risk, depositors sold their UST igniting the bank run.

To get more insights into Terra project’s failure, we describe the evolution of dependency structures among cryptocurrencies during the crash. In order to do this, we use instruments provided by network science. Networks have been extensively used to study financial systems (Mantegna,, 1999; Aste et al.,, 2010; Briola and Aste,, 2022) modeling dependencies among assets through correlations. Different correlation measures capture different relationships among assets. We decide to use the Pearson correlation coefficient to model linear relations among cryptocurrencies. It is worth noting that, in a condition of stress of the underlying system, pure correlations might be affected by an excessive sensitiveness. In order to partially mitigate such an effect, we assign a structure of weights to observational events, giving more relevance to the last observations of a given window. Weighted correlations are found to be smoother and recovering faster from market’s turbulence than their unweighted counterparts, helping also to discriminate more effectively genuine from spurious correlations. Following the definition in Pozzi et al., (2012), we define the Pearson correlation coefficient weighted with exponential smoothing as follows:

where $w_{t}=w_{0}e^{\frac{t-\Delta t}{\theta}},\forall t\in\{1,2,\dots,\Delta t\}\land\theta>0$ represents an exponentially smoothed weight structure such that $\sum_{t=1}^{\Delta t}w_{t}=1$ and $\overline{y}_{k}^{w}=\sum_{t=1}^{\Delta t}w_{t}y_{t}^{k}$. $\Delta t$ corresponds to rolling windows made of 24 hours with steps of 1 hour each and $\theta$ is here set to $0.3$.

Such a definition can be used to build a range of networks representing dependency structures among cryptocurrencies (Briola and Aste,, 2022). Here we use the state-of-the-art methodology, namely, Triangulated Maximally Filtered Graph (TMFG) (Massara et al.,, 2017). Such a filtering network comes with several advantages compared to its alternatives. It is able to capture meaningful interactions among multiple assets and it is characterised by topological constraints which help to regularise for probabilistic modeling (Aste,, 2022). As a measure of network centrality, we compute the eigenvector centrality, which allows us to measure the influence of a crypto-asset in the system. Intuitively, a cryptocurrency has an higher eigenvector centrality as long as it is connected to other relevant cryptocurrencies, which are also characterised by an high eigenvector centrality. In order to highlight relevant events and reduce the impact of secondary ones, also in this case, we compute an exponentially smoothed rolling average of the signal with a smoothing factor equals to $0.3$.¹¹¹¹11Results are consistent to different values of $\theta$.

In this study, we also use the approach proposed by Chang et al., (2000) to study the possible emergence of herd behaviour during Terra collapse. This method is based on the notion of herding towards market consensus in which “herds are characterised by individuals who suppress their own beliefs and base their investment decisions solely on the collective actions of the market, even when they disagree with its predictions” (Christie and Huang,, 1995). More specifically, Chang et al., (2000) analysed the existence of herding in a system of $N$ assets through the cross-sectional absolute deviation of returns (CSAD) as a measure of return dispersion,

and regressing the CSAD of returns with respect to market returns,

where $|r_{m,t}|$ is the absolute term and $r_{m,t}^{2}$ denotes the square of market returns. Chang et al., (2000) noted that rational asset-pricing models imply a linear relation between the dispersion in individual asset returns and the return on a market portfolio. As the market’s absolute return grows, cross-sectional return dispersion would also be expected to rise, in a linear fashion. Conversely, during periods of euphoria or fear, investors may react in a more uniform manner, giving rise to herding behavior. This phenomenon will decrease the dispersion among returns or at least increase at a less-than-proportional rate with the market return. Therefore, the emergence of herding behaviour will be consistent with a significantly negative coefficient $\beta_{2}$ in Eq.(3). However, in the absence of herding, we should only observe a positive and linear relation between the cross-sectional return dispersion and absolute market returns, i.e. a significant positive coefficient $\beta_{1}$, and a non-significant coefficient $\beta_{2}$. In this line, given that Eq.(3) represents the symmetric form that includes all the market return distribution, following Cui et al., (2019) and Chiang and Zheng, (2010) we also divided market returns to distinguish asymmetric herding when the market is up or down,

where $(1-D)$ and $D$ are dummy variables equal to 1 when $r_{m,t}\geq 0$ and $r_{m,t}<0$, respectively.¹²¹²12Following the corresponding literature (e.g. Chang et al.,, 2000; Chiang and Zheng,, 2010), in Eq.(3) and Eq.(4), we used Newey-West standard errors (Newey and West,, 1994).

Figure 4 reports exponentially smoothed average correlation coefficients for LUNA, UST, BTC and Kraken exchange¹³¹³13Kraken digital currency exchange represents the average correlation of all the 61 cryptocurrencies at our disposal.. We observe that, until 09 May 2022 12:00 (c), the market (Kraken) showed sporadic peaks of high correlations, such as the one observed after the first UST de-pegging (b). Starting from the second UST de-pegging, between 09 May 2022 12:00 (c) and 11 May 2022 10:00 (d), we identify a continuous increase in market correlations, until it stabilised for a short time around $0.8$. Considering that we use exponentially smoothed weighted correlations, which weight more the last hours of the window, we state that this constant increase in correlations shows that the market kept a high co-movement at each hour until 11 May 2022.

Therefore, from (c) to (d), we detect the specific period when most of the market reacted to Terra collapse, coinciding with the main decrease in UST and LUNA prices (see Figure 1). Afterwards, we observe a decrease in correlations in LUNA, UST, BTC and Kraken digital currency exchange. This new scenario shows that the first two cryptocurrencies were “excluded” from the system given their “death”, while the system seems to recover a lower and more stable degree of correlation.

After obtaining an initial picture through the study of market correlations, we compute the TMFG to obtain a better description of the system’s collective dynamics. Figure 5 reports the evolution of TMFG’s eigenvector centrality, through 24 hour rolling windows, for LUNA, UST and BTC. From (a) 05 May 2022 12:00 to (b) 07 May 2022 22:00, we observe that BTC was influencing the rest of the system, given its higher eigenvector centrality. This finding potentially supports the hypothesis of its involvement during the initial down-market since 05 May 2022 12:00. On (b) 07 May 2022 22:00, we identify the effect of the first Terra de-pegging on the cryptocurrency market’s dependency structure, since LUNA became part of the core of the network, influencing the cryptocurrency market for a short time (see Figure 6). Since that day, LUNA was excluded from the core of the network, gradually lowering its eigenvector centrality. This finding can be observed in the network reported in Figure 7, during the second UST de-pegging on 09 May 2022 12:00 (c). Finally, after (d) 11 May 2022 10:00, LUNA was completely excluded from the system, being relegated in the lowest eigenvector percentile.¹⁴¹⁴14Regarding UST, we do not observe relevant changes in terms of eigenvector centrality given its nature as stablecoin. Thus, even though connections with other cryptocurrencies could exist, most of them were non-significant.

Table (4) shows the results of the analysis of the herding behaviour in the cryptocurrency market during Terra collapse. Interestingly, despite the fact that the cryptocurrency market is more interconnected over time (Aslanidis et al.,, 2021) and herding is traditionally found on periods of uncertainty (da Gama Silva et al.,, 2019), we do not observe herding during the Terra collapse. In particular, we report significant positive coefficients for $|r_{m,t}|$, $(1-D)|r_{m,t}|$, and $D|r_{m,t}|$, while we underline non-significant coefficients for $r_{m,t}^{2}$, $(1-D)r_{m,t}^{2}$ and $Dr_{m,t}^{2}$.¹⁵¹⁵15For robustness purposes, we applied Eq.(4) through rolling windows of 7 days, i.e. 168 hours/observations. However, the absence of herding remains. Moreover, if we remove stablecoins from our sample, whose returns are mainly 0 (e.g. USDT or DAI), we keep obtaining the same result. This result supports the existence of an heterogeneous performance in the cryptocurrency market during this period. In Table 5 we report average hourly returns from (b) 07 May 2022 22:00 to (d) 11 May 2022 10:00, where we observe that some cryptocurrencies, like BTC and ETH, were characterised by a “better” (less negative) performance, compared to the rest of the market, whose average was equal to $-0.0031$. Consequently, even though we observe higher Pearson correlation coefficients during the collapse [see Figure 4], UST-LUNA failure did not dramatically affected the market due to the absence of herding and the low reaction of BTC [see Figure 6]. We could conjecture that investors considered this event as a non-structural shock, which did not posed real risk to jeopardise the future of the cryptocurrency system.

Considering that the failure of large stablecoins built upon inappropriate decentralized finance frameworks (e.g. lack of collateral and dependencies with liquidity providers) could give rise to potential financial stability risks, we believe that our results are relevant for both investors and policymakers. Applications of stablecoins within the crypto space remarkably increased since 2020, while poorly supported by adequate regulation. Initially, they were used as: (i) bridge between fiat currencies and crypto-assets; (ii) “parking space” for crypto volatility (Adachi et al.,, 2022). Nowadays, they are also used to provide most of the liquidity in DeFi applications such as decentralised exchanges and lending protocols (e.g. UST and Anchor protocol). Consequently, stablecoins play a critical role within the crypto space, where they are involved in almost 75% of the entire trading (Adachi et al.,, 2021). Largest stablecoins represent a primary source of risk according to the European Central Bank (ECB), which states that a “bank run” on Tether could disrupt trading and price discovery in crypto-asset markets, which could turn disorderly (ECB,, 2022). Therefore, even though the collapse of UST was not able to rise relevant herding effects, the failure of larger stablecoins could generate systemic effects on the whole crypto universe. In addition, considering the growing interlinks between crypto-assets and the traditional financial system, the failure of stablecoins could also have implications external to the crypto space. These scenarios remark the need for a Global Stablecoin Standard and continuously updated regulations able to guarantee (i) a process for fulfilling holders’ redemption claims, (ii) proper transparency of reserve asset composition and (iii) development of appropriate risk-management frameworks. Relevant regulatory standard-setting bodies (Financial Stability Board and Bank of International Settlements; see, e.g. Bank for International Settlements,, 2022) and governments (Japan, Hong Kong, United Kingdom, European Union and the United States; see, e.g. Asmakov,, 2022, European Council,, 2022) are already proposing robust rules for these digital assets. It would be beneficial that contributions from inside the DeFi community would also be proposed.

Most of the existing literature on stablecoins concerns their stability (Grobys et al.,, 2021), their role in portfolio diversification (Wang et al.,, 2020; Baur and Hoang,, 2021), and crypto asset price formation (Barucci et al.,, 2022; Kristoufek,, 2022). Due to the nascent interest in this research field and the rapid growth of decentralised finance, gaps in the literature still exist. Recent events around Terra project highlights the need to analyse specific topics that have been neglected until now: (i) stablecoins as a means of exchange, (ii) stablecoins’ role within the crypto space, (iii) regulation on stablecoins and related protocols and (iv) potential risk to financial stability. First, stablecoins could not be appropriate as a mean of exchange due to their vulnerability to “bank runs” and high transaction costs compared to non-blockchain alternatives (Mizrach,, 2022). Second, the role of stablecoins as “bridge” in the crypto space could be considered a double-edged sword since they could jeopardise the system’s stability once becoming systemic. Third, the dependence of UST on Anchor protocol remarks the need for better scrutiny of third party protocols related to any stablecoin. Finally, it would be necessary to consider scenarios in which specific fragilities within the stablecoins’ ecosystem may give rise to systemic financial stability risks.