JournalofFinancialStability73(2024)101286 Contents lists available at ScienceDirect Journal of Financial Stability journal homepage: www.elsevier.com/locate/jfs Decentralization illusion in Decentralized Finance: Evidence from tokenized voting in MakerDAO polls Xiaotong Suna, Charalampos Stasinakisb,*, Georgios Sermpinisb aFaculty of Finance, City University of Macau, Avenida Padre Toma´s Pereira Taipa, Macau, China bUniversity of Glasgow Business School, University of Glasgow, Gilbert Scott Building, Glasgow G12 8QQ, United Kingdom A R T I C L E I N F O A B S T R A C T JEL classification: Decentralized Autonomous Organization (DAO) is very popular in Decentralized Finance (DeFi) applications as it G23 provides a decentralized governance solution through blockchain. We analyze the governance characteristics in G32 the Maker protocol, its stablecoin DAI and its governance token Maker (MKR). To achieve that, we establish O30 several measurements of centralized governance. Our empirical analysis investigates the effect of centralized O33 governance over a series of factors related to MKR and DAI, such as financial, network and Twitter sentiment Keywords: indicators. Our results show that governance centralization influences the Maker protocol and that the distri- decentralized finance bution of voting power matters. The main implication of this study is that centralized governance in MakerDAO blockchain tokens, governance very much exists, while DeFi investors face a trade-off between decentralization and performance of a DeFi protocol. This further contributes to the contemporary debate over whether DeFi can be truly decentralized. 1. Introduction which can be the origin of several problems. The most intractable issue is probably the agency problem, where the owners and managers of an Since the introduction of Bitcoin in 2008 (Nakamoto, 2008), block- organization have different interests. Managers can pursue their own chain has deeply changed financial markets. Various debates have profits at the expense of owners (Fama and Jensen, 1983). Therefore, the evolved around the potential democratization of financial services most challenging objective of governance is to align the interests of (Bollaert et al., 2021), blockchain competition and services improve- owners and managers. As discussed by Lee (2019), the decentralized ment (Choi et al., 2020; Zhang et al., 2022), as well as investment op- nature of blockchain brings forward the idea of a ‘token economy’, portunities in new tokenized assets (Howell et al., 2020; Anyfantaki where capital is better directed to those users actually contributing et al., 2021; Karim et al., 2022). It is widely accepted, though, that the content and services. Within the DeFi context, owners and managers are main disruption lies in the disintermediation of financial institutions theoretically identical, which creates an opportunity to investigate the from their centralized role. The absence of centralized third parties, e.g., premises of this debate once again. Another crucial intersection between central banks, in the blockchain universe and circumventing traditional traditional finance and DeFi is stablecoins and their links with the po- barriers to participation in financial markets are the major attributes of tential introduction of Central Bank Digital Currencies (CBDCs). Even if this market revolution. The role of a central authority is limited or ab- stablecoins are considered safer than other cryptocurrencies, central sent. Such decentralized frameworks theoretically allow all participants banks continue to scrutinize them. Are these tokens really needed to to be part of prominent decision-making and share risk (Abdikerimova ensure DeFi liquidity, and does introducing CBDCs ensure financial and Feng, 2022). Decentralization, therefore, is logically regarded as the stability from a stablecoin crash? core value proposition of blockchain (Harvey et al., 2021). Stemming from this background, evaluating the efficiency of DeFi is Decentralized Finance (DeFi) describes blockchain-based financial a crucial task. As Momtaz (2022) explains, one pathway to settling this applications which are designed to replicate most financial activities, e. debate is by examining the true decentralized nature of DeFi platforms. g., lending and borrowing, in traditional markets. Theoretically, Despite the fact that the market size of DeFi exceeds 80 billion dollars (as governance in DeFi is decentralized since all members are decision of April 2022), the debate on whether decentralization is realistic or an makers. In traditional finance, governance is inevitably centralized, illusion still stands (Aramonte et al., 2021; Carter and Jeng, 2021). * Corresponding author. E-mail addresses: xtsun@cityu.edu.mo (X. Sun), Charalampos.Stasinakis@glasgow.ac.uk (C. Stasinakis), Georgios.Sermpinis@glasgow.ac.uk (G. Sermpinis). https://doi.org/10.1016/j.jfs.2024.101286 Received 16 August 2022; Received in revised form 14 May 2023; Accepted 9 January 2024 Availableonline27May2024 1572-3089/©2024TheAuthors.PublishedbyElsevierB.V.ThisisanopenaccessarticleundertheCCBYlicense(http://creativecommons.org/licenses/by/4.0/). X. Sun et al. J o u r n a l o f F i n a n c i a l S t a b i l i ty73(2024)101286 Anker-Sørensen and Zetzsche (2021) argue that innovators prefer less the most traded stablecoins, with more than ten thousand daily trans- decentralized DeFi platforms for making profits; as a result, governance actions. Although the Maker protocol seems to be a big success of DAO, rights and modes of control in DeFi are highly likely to be centralized. the way it is governed in practice has not been rigidly examined. To the DeFi platforms showcase elements of centralization, usually in the form best of our knowledge, this is the first paper that focuses on providing of ‘governance tokens’ and power concentration to large coin-holders. empirical evidence of centralized governance in DeFi. This phenomenon can lead to collusion among core decision makers To achieve that, we collect information for the Maker protocol during the governance process. It is obvious, then, that governance be- governance, including all voters, their choice and voting power in Maker comes a critical dimension of the success of true decentralization in governance polls from 5th August 2019–22nd October 2021. Our DeFi. Decentralized Autonomous Organization (DAO) is one popular empirical analysis follows two stages. The first stage is to examine solution to decentralized governance and decision making. In a DAO, all governance polls by defining three novel measurements of centralized members are owners of the organization, and they have decision-making governance, namely voting participation, centralized voting power and power around its development. Usually, the suggested changes will be distribution of governance tokens. In the second stage, we investigate written in the form of an Improvement Proposal (IP), which is then voted the effect of centralized governance on the development of Maker pro- on through an established poll, where all members can make public their tocol. To achieve this, we expose MKR and DAI to several Maker-specific choices. DAO members state their choice through governance tokens. factors. These factors can be divided into several categories, including Usually, these governance tokens are also tradable cryptocurrencies. financial, network and Twitter sentiment indicators. We also investigate The votes are weighted by the number of governance token held by the ratios of collateral assets locked in Maker protocol. Such an empir- voters. In other words, governance in DAO is tokenized. Currently, DAO ical setup is consistent with similar studies in the field, such as Liu and is one of the most common governance mechanisms adopted by DeFi Tsyvinski (2020). Beside well-investigated factors, e.g., network factors, (World Economic Forum, 2021). we also consider transaction demand (e.g., trading volume), which is a So far, voluminous literature focuses on blockchain governance, and theoretical determinant of token price (Cong et al., 2020). Finally, since the debate revolves around the pros and cons of decentralization. users have to lock collateral assets before initiating loans from Maker, Decentralization will result in slower decision making, implying that the acceptable collaterals are a main issue discussed in Maker governance. If network becomes inefficient (Hsieh et al., 2017). Yermack (2017) argues collateral is insufficient, theoretically Maker protocol will be less safe that, in practice, blockchain governance is not completely decentralized. due to fewer users participating in it. In some extreme cases, the final decision is taken by only the core de- The empirical framework brings forward some very interesting velopers. For example, Bitcoin core developers once decided to lower findings. By examining governance polls in the Maker protocol, we transaction fees without discussing it with the related community observe signals of centralized Maker governance. Compared with the (Gervais et al., 2014). Recently, Jiang et al. (2022) discuss blockchain rapidly increasing number of users, voters are centralized in a small governance by evaluating the trade-off between stability and efficiency group, and the most dominant voters are heterogeneous in character- through the prism of sensitivity to transaction fees. The authors suggest istics. The unevenly distributed voting power, as a preliminary signal of that the decentralized and audible nature of the blockchain transaction governance centralization, leads to our measurements of governance is attractive, but transaction fee movements have led to fork splits and centralization in Maker protocol. By applying factor analysis, we find a endangered the system’s stability. Their findings show that when users complex nexus of effects of centralized governance around voting have balanced preferences between efficiency and stability, raising participation and distribution of voting power. Intuitively, more voters transaction fees reduces congestion in the platform. However, when it are a signal of larger voting participation, implying more decentralized comes to DeFi platforms and DeFi governance, the literature is quite governance. Voting participation can directly affect the financial factors silent. What would be the effect of powerful voters proposing and voting of DAI. For example, the trading volume of DAI decreases as more voters on polls that serve their own interests? Tsoukalas and Falk (2020) and vote in governance polls. This suggests that stablecoin can be affected by Carter and Jeng (2021) provide some insights on this question. Many participation in the polls and that decentralized governance could affect blockchain-based platforms apply a token-weighted voting mechanism, market performance of cryptocurrencies. Centralized distribution of the relying on the premise that tokenized voting incentivizes users towards governance token, i.e., MKR, can decrease the trading volume of MKR higher-quality voting and improves system performance. The four and DAI, implying that centralization may bring more serious problems. mentioned authors explain that this is not always correct, as this voting After expanding our work on other indicators, we find centralized mechanism discourages truthful votes and decreases the stability of the governance exerts complex influence on the adoption of Maker protocol. platform. Recently, Goldberg and Scha¨r (2023) utilize Decentraland, a This is a serious issue, as the more decentralized MakerDAO becomes, blockchain-based virtual world (i.e., metaverse), as a case study. By the fewer users start using DAI stablecoin. This paints a not very opti- analyzing voting behavior, they contend that voting power is not mistic picture not only for the long-term growth of Maker protocol but decentralized, which may cause rent extraction behavior and other also for other DAO-governed DeFi platforms. Finally, voting power related problems. Therefore, centralized governance in DeFi could exert distribution appears to play a significant role in the ratios of collateral negative effects. assets. Consequently, the centralized voting power of large voters may Though such papers provide both theoretical models and empirical change the proportion of main collateral assets (e.g., stablecoins) locked evidence of governance centralization in blockchain, the literature in the platform. Overall, all the above findings suggest that both the surprisingly remains silent when it comes to centralization in DeFi. degree of MakerDAO’s centralized governance and its performance pose Positioning the centralized governance debate in the DeFi universe at a relevant trade-off among DeFi investors. the forefront of the literature is the main motivation of this paper. We The remainder of our paper is organized as follows. Section 2 pro- focus on the Ethereum-based DeFi platform, Maker protocol, developed vides a summary of the governance process in MakerDAO. The dataset and managed by MakerDAO, as a case study. The rationale behind this and the measurements of centralized governance in the Maker protocol choice is simple. MakerDAO is one of the most influential DAOs. Since are defined in Section 3. The main empirical results are presented in 2017, when the DeFi universe expanded exponentially, Maker protocol Section 4, while robustness checks are provided in Section 5. Section 6 has emerged as the leading lending protocol, which conceptually rep- provides some concluding remarks. Finally, the appendix provides a licates the operation of a bank in cryptocurrency markets. In the Maker description of the utilized factors and relevant Granger tests, while protocol, Maker (MKR) is the governance token. In terms of its value, further technical details and robustness tests are given in the online one token equals one vote in the proposed polls. Apart from this toke- appendix. nized value, Maker protocol issues DAI, which is a stablecoin soft- pegged to the US Dollar (MakerDAO, 2020). Currently, DAI is one of 2 X. Sun et al. J o u r n a l o f F i n a n c i a l S t a b i l i ty73(2024)101286 2. Governance in the Maker protocol power is weighted by the amount of MKR that a voter owns and repre- sents, making the voting mechanism a token-weighted one (Tsoukalas 2.1. Decentralized Autonomous Organization (DAO), MakerDAO and and Falk, 2020). One MKR equals one vote, and the option with the most Maker protocol votes wins. In the Maker protocol, Maker IPs are structured and formalized for a voting event, and key issues and changes to the system Decentralized Autonomous Organization (DAO) is a novel mecha- are rigidly defined in Maker IPs. Usually, the Maker Foundation will nism of organizational governance and decision making. The DAO white draft the initial Maker IPs, and any community members can propose paper is first proposed by Jentzsch (2016). Technically, DAO can be competing IPs. Then, the final decisions will be made by MKR voters deployed on blockchain, and, currently, most DAOs rely on Ethereum, through the current Maker governance process. The above information which is a programmable blockchain. Ethereum’s yellow paper was is illustrated in Fig. 1. introduced by Wood (2014), and Ethereum users can write smart con- MKR holders can be voters and directly choose their options on the tracts in a Turing-complete programming language such as Solidity. By Maker Governance Portal. On the other hand, they can choose a Vote writing and executing smart contracts, users can actualize various in- Delegate to be their representative. As a result, delegates gain voting teractions and functions, e.g., transactions on Ethereum. The program- power from MKR holders, and these MKR holders can indirectly vote. mable character enables the implementation of DAO. The core of DAO The voting results are weighted by the amount of MKR voted for a governance is based on standard smart contract code instead of human proposal. Noticeably, Vote Delegates were not introduced in the very actors. In other words, DAO’s governance is tokenized. In practice, beginning. On July 30th, 2021, the guidance to Vote Delegates was live DAO-based protocols usually have their own governance token and in MakerDAO, while on 10th November 2021, 16 Delegates and governance token holders can vote on changes to the protocols. 65989.65 MKR were delegated. Currently, there are three types of MakerDAO was created in 2014, and it has become one of the most voting in Maker governance, i.e., Forum Signal Threads, Governance influential DAOs. The Maker protocol is a multi-collateral lending sys- Polls and Executive Votes. These are summarized in Table 1. tem, and the protocol is governed by MakerDAO teams, including in- Table 1 highlights the functions of the three types of votes. Forum dividuals and service providers (MakerDAO, 2020). Based on the Signal Threads are the least consequential, and the threads are a part of functions of Maker protocol, it is usually categorized as a Lending Pro- off-chain governance. Governance Polls and Executive Votes occur on- tocol (LP), resembling banks in cryptocurrency markets. Simply, users chain, and they can be accessed through the Maker Foundation’s can lend their tokens to LPs for economic incentives. On the other hand, Voting portal. Simply, Governance Polls determine processes outside the users can borrow tokens, and LPs usually require collateralization. The technical layer, while Executive Votes are about technical changes to the economic mechanism, mathematical models and the roles of LPs are well protocol. Executive Votes use the ’Continuous Approval Voting’ model discussed by Bartoletti et al. (2020). Maker protocol issues DAI and the to make the system more secure. The voting model means that new protocol is de facto a Multi-Collateral Dai (MCD) system. DAI is probably proposals need to surpass the voting weight of the last successful pro- the most notable stablecoin, which is soft-pegged to the US dollar. Sta- posal (MakerDAO, 2021). blecoins, such as DAI, are cryptoassets designed to cope with the vola- tility of traditional cryptocurrencies and provide a bridge with fiat 2.3. Governance centralization currencies (Wang et al., 2020). As MCD was launched in 2019, in Maker protocol every user can lock any supportive collateral such as ETH and a Centralization in the governance layer of blockchain has attracted corresponding amount of DAI will be generated as debt.1 the attention of the academic audience, with the discussion mainly In addition to DAI, we are also interested in the MKR token. In focusing on two problems, namely owner control and improvement practice, MKR plays two roles. On the one hand, MKR is the governance protocol (Sai et al., 2021). Gervais et al. (2014) argue that the author token of Maker protocol. MKR holders can vote on changes to the pro- Satoshi Nakamoto may accumulate significant Bitcoin since Nakamoto tocol. On the other hand, MKR contributes to the recapitalization of the participated in activities at the early stage of Bitcoin blockchain. Similar system. MKR is created or destroyed through the automated auction evidence of owner control exists in Ethereum as well (Bai et al., 2020). mechanism of Maker protocol. When the debt of protocol is outstanding, The large proportion of wealth controlled by the owners of blockchains MKR is created and sold for DAI. The protocol sells DAI for MKR and the may result in economic manipulation in blockchain. As for improvement surplus MKR is destroyed. At the inception of MakerDAO, one million protocol, this problem derives from the process of moderation in MKR were issued. The protocol sets maximum threshold and minimum blockchain. Usually, blockchain and DeFi adopt an improvement pro- threshold of MKR, and the total circulated MKR always fluctuates be- posal system in the decision-making process. By analyzing the authors tween the thresholds. and contributors of improvement proposals, Azouvi et al. (2018) find that core developers are the main contributors to the development of 2.2. Maker governance structure and voting process Bitcoin and Ethereum. In other words, these core developers have more power in the decision-making process. An innovative selling point of the Maker protocol is decentralized Apart from the two problems described above, DAO, as a new governance. In the Maker protocol, governance can be divided into two organizational form to automate governance, may bring both opportu- parts: on-chain governance and off-chain governance. In on-chain nities and challenges. Benefiting from blockchain technology, the governance, there are two types of votes, namely Governance Polls ownership is more transparent and voting can be more accurate (Yer- and Executive Votes. Any MKR holders can vote using the Maker Pro- mack, 2017). On the other hand, centralization seems to still be inevi- tocol’s on-chain governance system. Governance polls, which are about table in DAO. Maker Governance Polls do not attract much participation non-technical changes, measure the sentiment of MKR holders. Execu- and MKR held by the dominant voters may theoretically lead to collu- tive votes “execute” technical changes to the protocol. The voting results sion in some polls. However, in order to examine this, we need infor- are documented on blockchain. Off-chain governance is mainly about mation from many polls, and this is not an easy task. As shown in the informal discussion, e.g., discussion on the MakerDAO forum. Both MKR next section, the quest to obtain information from more polls is achieved holders and the larger community can express their opinions. Voting in this study. 1 As of April 2022, DAI is fifth in market capitalization among all stablecoins (over nine billion US dollars). For more details regarding DAI debt in Maker protocol, we refer the reader to Qin et al. (2021). 3 X. Sun et al. J o u r n a l o f F i n a n c i a l S t a b i l i ty73(2024)101286 Fig. 1. Governance in Maker Protocol. Note: Panel A shows the governance structure of Maker protocol. It is divided into two parts, on-chain governance and off- chain governance. Panel B illustrates the voting process of Maker governance polls. MKR holders can participate in polls as voters, or they can choose delegates as their representatives. blockchain via a transaction. Though the dates might be earlier than the Table 1 start dates when voters can choose options, the contents of polls are Types of votes in Maker protocol. already publicly accessible once the polls are sent to the blockchain. Poll Type of votes Functions 16 was the first governance poll that MKR holders could participate in. Forum signal (1) Determine consensus that something needs to be done in Some polls failed,2 so they are not documented in the portal. Hence, the threads response to a perceived issue, (2) determine consensus for a dataset consists of a total of 638 successful governance polls, and the concrete action to be taken in response to a perceived issue. voters’ public names can be found by searching for their addresses on Governance polls (1) Determine governance and DAO processes outside the Etherscan.io and Maker Governance Portal. To study the effects of technical layer of the Maker Protocol, (2) form consensus on important community goals and targets, (3) measure centralization in Maker governance, we consider two influential crypto sentiment on potential Executive Vote proposals, (4) ratify assets issued by Maker protocol, namely MKR and DAI.3 governance proposals originating from the MakerDAO forum signal threads, (5) determine which values certain system parameters should be set to before those values are then 3.2. Measurements of centralized voting power in Maker protocol confirmed in an executive vote, (6) ratify risk parameters for new collateral types as presented by Risk Teams. Executive votes (1) Add or remove collateral types, add or remove Vault types, This section introduces the novel measurements of centralized voting adjust global system parameters, adjust Vault-specific power in Maker protocol, namely voting participation, centralized parameters, (2) replace modular smart contracts. voting power and distribution of governance tokens. For each of the first Note: This table describes three types of votes in Maker protocol. Forum signal two measurements, we initially calculate the value at the poll level. threads are a part of off-chain governance, and all community members can Then, daily measurements can be generated. The distribution of gover- participate in the discussion on the Maker forum. Governance polls and execu- nance tokens can divulgate more information about centralized power of tive votes are on-chain. certain Maker protocol users, e.g., MakerDAO delegates and large MKR holders. 3. Data collection and identification 3.2.1. Voting participation 3.1. Data Collection To proxy voting participation, we use two measurements. One is the total votes of Maker governance polls on a given date. The other is the In Maker Governance Portal, the details of governance polls, e.g., the number of total voters on a given date. Here, a voter refers to an number of voters and results, are publicly available. To get the voters’ addresses, we query the voting history from MCD Voting Tracker. We investigate governance polls from Poll 16 (deployed on 5th August 2 Polls 28, 39, 47, 69, 78, 183, 282, 284, 286, and 500 failed. 2019) to Poll 663 (deployed on 22nd October 2021). The deploy dates of 3 The Maker portal is available at: https://vote.makerdao.com. The voting the polls are used to identify when the polls are added to Ethereum tracker is available at: https://beta.mcdgov.info. The DAI and MKR statistics can be found at: https://www.intotheblock.com/. 4

X. Sun et al. J o u r n a l o f F i n a n c i a l S t a b i l i ty73(2024)101286 Ethereum address.4 Intuitively, when these two measurements are decision speed of the largest voter. Here, the order of the largest voter higher, there are more voters and votes in governance polls. Assuming refers to the order in the whole history. This is to say, the voters with that there are n polls and m voters on a date d,we have: dominant voting power may change their choice later. Assuming that ∑n there are k records in the voting history of a governance poll i on a date Totalvotesd = Totalvotesi,d (1) d, we have: i=1 votingorderof thelargestvoter and Orderi,d = k (5) ∑m Assuming that there are n polls on a date d,we have Votersd = Votersi,d (2) ∑n i=1 Orderi,d 3.2.2. Centralized voting power Orderd = i=1 n (6) In order to capture centralized voting power, we utilize two mea- When order is smaller, the largest voter chooses an option earlier sures. The first is the Gini coefficient, which is traditionally used to than other voters. All voters can see the existing choices on Maker measure inequality (Dorfman, 1979). Assuming that there are l voters in Governance Portal. a governance poll, we have: ∑l ∑l ⃒ ⃒ votesi (cid:0) votesj ⃒ ⃒ 3.2. I 3 n . a D dd is i t t r i i o b n u ti t o o n t h o e f g m ov e e a r s n u a r n e c m e e t n o t k s e n st s e mming from voting behavior in Gini= i=1j=1 (3) Maker governance polls, we also consider the distribution of governance 2l2votes tokens in MakerDAO (i.e., MKR), which reveals more information about the characteristics of centralized decision makers. Nadler and Scha¨r Where votesi is the cast votes of voter i, and votes is the average votes in a (2020) show that token distribution is usually centralized in DeFi, and governance poll. we suspect such centralization also exists in MakerDAO. First, we After computing the Gini coefficient for each poll, we can calculate a calculate the balance of MKR controlled by MakerDAO delegates, which daily measurement by calculating the average. Assuming that there are n polls on a date d,we calculate the daily Gini coefficient, i.e., Ginid, via e g q at u e a s l . s 5 t A h s e s u s m um in g o f t h M at K t R h e o r w e n ar e e d l b d y e l d e e g l a e t g e a s t , e t s h e a n M d K r R e p b r a e l s a e n n c t e e d co b n y tr o d l e l l e e d - maximum likelihood estimation (see Taleb 2015). The number of voters by these delegates on a date d is: participating in governance polls can be very different. If we choose arithmetic means to measure the daily Gini coefficient, our measure- ∑l ment can be more biased and suffer from lower accuracy. Maximum delegated = delegatei,d (7) likelihood estimation, as an indirect method, can have a considerably i=1 lower error rate, especially when the sample sizes (in this case, the voters in a governance poll) vary. where delegatei,d is the MKR balance controlled by delegate i on date d. A higher delegate means more voting power is controlled by MakerDAO The second proxy for centralized voting power is the largest voter’s delegates, implying that the centralized governance is caused by these power in Maker governance polls. Here, largest voter refers to the voter influential MakerDAO users. that contributes most votes in a governance poll. The centralized voting Then, we compute the proportion of MKR controlled by large MKR power of largest voters can be approximated by the Largestvotingshare. holders. Here, we consider three categories of MKR holders, including In that way, we can also reflect on the relative voting power of the holders with a balance of more than 10,000 MKR, holders with a balance largest voter. For each poll, Order refers to the voting order of the largest between 10,000 and 100,000 MKR, and holders with a balance of more voter. When Order is smaller, the largest voter will choose their option than 100,000 MKR. On a date d, we assume that there are x,y,z MKR earlier. Assuming that there are n polls on a date d,we have: holders in the three categories, respectively. Then, three measurements ∑n can be calculated to reflect on the centralized distribution of MKR: Largestvotingsharei,d ∑ Largestvotingshared = i=1 n (4) >10kd = totalci x i r = c 1 u > la 1 te 0 d k M i,d KRd (8) Finally, the voting sequence can actually play a role, as this can be ∑ documented in several voting systems (see amongst others, Bo¨rgers 10k(cid:0) 100kd = total y i c = i 1 rc 1 u 0 l k a (cid:0) te 1 d 0 M 0k K i R ,d d (9) (2010), Brams 2008, and Brams and Fishburn 2002). Simply, voters ∑ h w a i v th e t d h i e f i f r e r g e o n a t l s s . t F r o at r e e g a i c e h s, p a o n ll d , w th e e d ir e fi v n o e ti n a g v a o r r i d ab er le p o r r e d f e e r r e to n c m e e w as i u ll r e v a th ry e >100kd = totalci z i r = c 1 u > la 1 te 0 d k M i,d KRd (10) where totalcirculatedMKRd is the amount of MKR circulating on Ethereum blockchain on date d. 4 An additional consideration about the number of voters is the potential When the three measurements above are higher, more voting power issue of wrapping complexity, as explored by Nadler and Scha¨r (2020). In is controlled by large MKR holders. Although governance polls are not simple terms, DeFi users can deposit their cryptocurrencies to smart contracts of deployed on a daily basis, the centralized distribution of MKR is a signal DeFi protocols. In other words, cryptocurrencies are pooled in these contracts but remained owned by DeFi users. When such smart contracts appear in the list of governance centralization, indicating that voting in MakerDAO is of voters in MakerDAO, it is crucial to trace the source of MKR tokens and dominated by large MKR holders. allocate them to their rightful owners. To address this challenge, we use Sco- pescan.ai and manually scrutinize all addresses identified as voters in Maker- 4. Empirical results DAO governance. Our data sample of MakerDAO voters do not include contracts that may suffer from problems related to this wrapping complexity. This section summarizes the empirical results of this study. The first The majority of voters are external owned accounts (EOAs), representing subsection presents the descriptive statistics of both polls and voters. Ethereum addresses controlled by individuals. While a few voters are associated with smart contracts created by EOAs, it is noteworthy that none of these are linked to DeFi or token wrappers. This ensures a more accurate and focused analysis of MakerDAO governance dynamics. 5 We query the voting power of MakerDAO delegates from dune.xyz. 5 X. Sun et al. J o u r n a l o f F i n a n c i a l S t a b i l i ty73(2024)101286 Table 2 Descriptive statistics of Maker governance polls. Total votes Total voters Breakdown votes Breakdown ratio Breakdown voters Votes of the largest voter Vote share of the largest voter Mean 36096.52 24.59 31529.94 88.78% 18.03 16941.61 52.66% Median 33097.15 23.00 28625.80 98.24% 16.00 17063.93 48.35% Maximum 131555.35 146 108694.07 100.00% 142 39403.85 98.51% Minimum 259.74 5 232.80 13.04% 1 203.27 20.28% Std 22383.18 12.76 19998.47 16.67% 11.63 8452.45 18.02% Note: This table reports the descriptive statistics of Maker governance polls. Breakdown votes refers to the votes of the winning option, and breakdown ratio is breakdown votes divided by total votes. Breakdown voters are the number of voters who choose the winning option. Fig. 2. The number of polls and voters in Maker governance (Aug 5, 2019 – Oct 22, 2021). Note: This figure presents the daily number of polls and voters in Maker governance. Panel A shows the daily number of Maker governance polls, while Panel B presents the number of voters daily in Maker governance polls. on the Maker protocol (MKR, DAI and locked collateral assets). Table 3 Descriptive statistics of voters of Maker governance polls. 4.1. First stage: Governance polls in the Maker protocol Involved polls Total votes First poll The highest votes Mean 12.55 18422.58 278.66 665.39 The collected information from the 638 governance polls is crucial Median 2.00 1.42 248.00 1.00 for understanding centralization in the Maker protocol. Table 2 presents Maximum 514 4170786.51 660 39403.85 the descriptive statistics of these polls. Minimum 1 0.00 16 0.00 Std 42.46 164269.26 194.65 3372.75 The votes are calculated in MKR tokens, and for each governance poll, breakdown ratio is the proportion of breakdown votes to total Note: This table reports the descriptive statistics of voters in Maker governance votes. In addition to votes and vote share of the largest voter, the order polls. For each voter, we calculate the number of polls that they participate in, their total votes and the highest number of votes in a single poll. Here, votes are of voting is considered. We also present the daily number of governance calculated in Maker (MKR), which is the governance token of the Maker Pro- polls and voters in Fig. 2. From the figures we can easily see that within a tocol. The first poll that a voter participated in is also presented. A lower number day, the number of deployed polls is usually less than 25. Usually, no means that the voter started to participate in Maker governance polls earlier. more than 700 voters will vote on the same day. For some polls, no more than ten voters will participate in decision making. The finding implies Then, the centralization in Maker governance polls is described by the that not all polls have large voting participation. Compared to the rapid calculations of the measurements of centralized governance defined in growth of Maker users, voters are a small group. Our analysis extracts a the previous sections. The second subsection summarizes the factor total of 1,250 unique voters in our dataset. For each voter, the number of analysis we perform to investigate the effects of centralized governance polls that they participate in can be surprisingly different. To showcase this, we present the following descriptive statistics in Table 3. 6 X. Sun et al. J o u r n a l o f F i n a n c i a l S t a b i l i ty73(2024)101286 Fig. 3. Total votes, votes of the largest voter and votes of the known voters in Maker polls (Poll 16 – Poll 663). Note: Panel A presents the total votes and votes of largest voters in Maker governance polls (Poll 16 – Poll 663). In most polls, the largest voter holds a significant proportion of voting power. Panel B shows the votes from the known voters in Maker governance polls (Poll 16 – Poll 663). The known voters include voting delegates and a16z and show strong voting power after Poll 600. Examining the total votes and the highest votes that a voter has in a Table 4 single poll shows that the voting power is not equally distributed across Gini coefficient in Maker governance polls. voters. This could be an early sign of voting centralization. However, to make this claim clearer, we need to delve deeper into the composition of Poll-level Daily the voters and their characteristics. To that end, we identify voters Mean 84.38% 18.57% whose identity is publicly available, the top ten voters that participate in Median 85.54% 0.00% most polls, the top ten voters with the largest total votes and the top ten Maximum 98.05% 94.04% Minimum 57.56% 0.00% voters that have the largest votes in a single poll. This information is Std 0.06 0.35 summarized in the online Appendix OA.1. We have some interesting findings towards identifying centralized Note: This table reports the Gini coefficient in Maker governance polls. In the governance in the Maker protocol. Apart from a16z,6 the known voters first column, we calculate the Gini coefficient for each poll. In the second col- umn, we first integrate a voter’s votes within a day, and then we compute the are delegates in Maker governance and their identity is publicly avail- daily Gini coefficient via a maximum likelihood estimation. able on the Maker governance portal, and more details of these known voters are given as well. The mechanism of voting delegates was intro- is unknown. Voters with the largest total votes are again heterogeneous duced in July 2021; therefore, most delegates started participating in in characteristics, while only two from the top ten are found to be del- governance polls in August 2021. Noticeably, the total votes and the egates (Field Technologies, Inc. and a shadow delegate). When ac- highest votes in a single poll are different among these known voters. counting for the voters with the largest single votes, we find again a Field Technologies, Inc. is the known voter with the largest number of different composition. We identify delegates such as Field Technologies, total votes (as of November 1st, 2021). In terms of voters that are Inc., Flip Flop Flap Delegate LLC, a shadow delegate and a16z being participating in polls, none of them has a public name, i.e., their identity dominant, while the remaining voters appear with unknown identities. Taking a wholistic look at these findings, we notice that some voters may both participate in many polls and have large total votes, namely voters 6 It is easy to establish by searching for other voters’ addresses on Etherscan.io with the addresses 0x4f…3f30 and 0x6a…ab40. In other cases, some voters might not participate in many polls, but when they do, they have that a16z represents the venture capital firm Andreessen Horowitz, which is the most influential venture capital in DeFi markets. significantly large votes in certain polls. For example, a16z only votes for 7 X. Sun et al. J o u r n a l o f F i n a n c i a l S t a b i l i ty73(2024)101286 Fig. 4. Gini coefficient. Note: This figure shows the Gini coefficient in Maker governance. Panel A reports the Gini coefficient at the poll level (Poll 16 – Poll 663). Panel B reports the Gini coefficient daily in Maker governance polls (Aug 5, 2019 – Oct 22, 2021). three polls, but their single votes are more than 30,000. These charac- Although the above could happen through descriptive information teristics of the dominant voters suggest that on-chain developments on extracted from our unique dataset, we take further steps to quantita- the protocol are driven by dominant voters and that decentralization tively establish centralized governance on the Maker protocol. First, we does not seem to hold. Voting power appears to be distributed unevenly measure the centralized voting power in Maker governance at a poll across different known or unknown small groups of voters, especially level and across days by utilizing the Gini coefficient estimations. The when total votes and large votes in a single poll are considered. results are summarized in the table and figure that follow. In order to further show this, we focus now on the notion of At a poll level, we find that the Gini coefficient is always more than centralized voting power in the Maker polls. We compare the largest 50% and exhibits a maximum of 98.05%. Given that the Gini coefficient vote for each poll with the total votes, and we find that the largest voter estimation is higher than 0.60 for most of the polls, highly centralized can account for a significant share of the total votes in most polls. voting power in the Maker governance is established. We also calculate Practically, the largest voter is the pivotal figure in implementing pro- and illustrate the daily Gini coefficient. The expected daily average Gini tocol changes, as they tend to account for around one third of the voting coefficient should be around zero, if no centralized voting occurs. share. In terms of the known voters (namely delegates and a16z), the However, we observe that there are days that the value is higher than trend is similar. These known voters are identified after the delegate 0.75, again implying strong centralization of voting power on particular regulation change in Maker protocol (after Poll 600) and their dominant days within our period under study. We further highlight the evidence of power is evident. However, it is hard to say if they were able to play an vote centralization by estimating the Lorenz curve of the cumulative important role in previous polls. All this information is illustrated in the total votes for particular polls. The results support the above findings following figure. and are presented in the online Appendix OA.1. To support the above, we also illustrate the total votes over the We also illustrate the voting power of large MKR holders and Mak- breakdown votes and their respective voters, the breakdown ratio, the erDAO delegates. For MKR holders whose MKR balance is between voting share of the largest voter and average voting share of the largest 10,000 and 100,000 (hereafter, major holders), their voting power is voter daily. The results show that winning polls are driven by the most around 25%. For MKR holders with more than 100,000 MKR (hereafter, votes, the largest voters contribute significant votes to the winning op- supermajority holders), their voting power accounts for a significant tions and the largest voters consistently concentrate at least 30% of the proportion, though their voting power has decreased since December average daily voting share. These voting patterns are presented in the 2020. The amount of MKR controlled by MakerDAO delegates should online Appendix OA.1. The key message remains that centralized voting not be ignored, given that the spikes of their MKR balances are close to power exists. 150,000 MKR. 8 X. Sun et al. J o u r n a l o f F i n a n c i a l S t a b i l i ty73(2024)101286 Fig. 5. Voting power of large MKR holders and MakerDAO delegates. Note: This figure illustrates the proportion of MKR controlled by large MKR holders and MKR balance controlled by MakerDAO delegates (Aug 5, 2019 – Oct 22, 2021). In Panel A, we calculate the proportion of MKR controlled by holders whose MKR balance is between 10,000 and 100,000. In Panel B, we focus on MKR holders whose MKR balance is more than 100,000. Panel C shows the MKR balance controlled by MakerDAO delegates. Table 5 Measurements of governance centralization in Maker. Voters TotalVotes LargestShare Order 10k-100k >100k >10k Delegate Mean 123.53 181335.26 0.54 0.41 0.33 0.28 0.61 105743.35 Median 68.00 114304.83 0.51 0.39 0.33 0.32 0.63 101145.62 Maximum 756.00 1251962.15 0.96 0.91 0.51 0.44 0.67 146462.93 Minimum 7.00 259.74 0.27 0.00 0.22 0.12 0.49 1151.71 Std 141.27 209387.53 0.16 0.20 0.07 0.10 0.05 23047.70 N of obs. 127 127 127 127 810 810 810 320 Note: This table presents the descriptive statistics of measurement of governance centralization in Maker. The first four columns report measurements of governance centralization in Maker governance polls. We first calculate these measurements for each poll and then convert it to daily level measurements. For example, we first calculate the number of voters for every poll, which we then add to the get the number of daily voters. The last four measurements reflect the influence of large MKR holders and delegates. Finally, the other measurements of governance centralization are 4.2. Second stage: Factor analysis established based on the definitions given in Section 3. Their descriptive statistics are provided in the following table. In this section, we first apply a series of univariate regressions, with To simplify the factor analysis, we implement Principal Component MKR and DAI used as dependent variables and the measurements of Analysis (PCA). Simply, higher explained variance ratios mean more centralized governance as independent variables. We consider financial, important measurements. In the following sections, we estimate re- network and Twitter sentiment factors. In other words, we estimate the gressions using the four measurements, including Gini,Voters, 100k(cid:0) following regressions: 100k, >100k. Gini and Voters represent voting centralization, while 100k(cid:0) 100k and >100k describe holding centralization in MakerDAO.7 factori,t =β 0 +β 1 centralt +ε t (11) Where: • i ={MKR,DAI} re 7 q u T e h s e t. B re e g si r d e e s s t i h o e n fo r u e r s u c l h ts o se fo n r m o e t a h s e u r r e m m e e a n s t u s, r e w m e e a n l t s s o e c x a a n m b in e e t p h r e o v in id fl e u d e n u c p e o o n f • centralt ={Voterst ,Ginit ,10k(cid:0) 100kt ,>100kt ,Delegatet } the voting power of MakerDAO delegates in Section 4, and a measurement ‘delegate’ is constructed. Compared to other centralization measurements, Given i, factors can be defined as a set: ‘delegate’ has fewer observations. We exclude ‘delegate’ from PCA analysis to factori,t ={financiali,j,t ,networki,k,t ,Twittersentimenti,l,t } evaluate the importance of other centralization measurements more accurately. 9 X. Sun et al. J o u r n a l o f F i n a n c i a l S t a b i l i ty73(2024)101286 Table 6 Table 7 Principal Component Analysis (PCA) of measurements. Financial factors (MKR, DAI). Panel A: Measurements related to governance polls PANEL A: MKR Explained Variance Ratio Explained Variance N of obs. Voters Gini 10k-100k >100k Delegate Gini 51.41% 0.15 127 Return -0.02 0.02 0.01 -0.01 -0.02 Voters 19.68% 0.06 127 (-0.64) (1.43) (1.19) (-1.06) (-0.72) TotalVotes 14.25% 0.04 127 ΔMktC -0.01 0.00 0.00 0.00 0.00 LargestShare 12.56% 0.04 127 (-1.26) (-0.19) (0.53) (-0.44) (-0.21) Order 0.02% 0.01 127 Volume 0.00 0.00 0.09*** -0.07*** -0.04 Panel B: Measurements related to MKR balance (0.45) (0.22) (6.56) (-6.40) (-1.14) Explained Variance Ratio Explained Variance N of obs. Volume_dex 0.02 0.01 0.12*** -0.10*** 0.01 10–100k 83.34% 0.19 810 (0.45) (0.29) (7.83) (-8.60) (0.27) >100k 16.67% 0.04 810 Volume_dex_usd 0.03 0.01 0.12*** -0.12*** 0.18*** >10k 0.00% 0.00 810 (0.68) (0.28) (9.38) (-13.52) (3.99) Volume_l 0.12 -0.04 0.11*** -0.09*** -0.03 Note: This table reports the results for Principal Component Analysis (PCA) of (0.82) (-0.76) (6.68) (-6.88) (-0.74) centralization measurements. For each measurement of governance centraliza- Volume_l_usd 0.10 -0.02 0.09*** -0.10*** 0.11*** tion, we present the ratio of explained variance and the variance explained by (1.13) (-0.70) (8.32) (-11.95) (3.00) this measurement, respectively. A higher ‘explained variance ratio’ implies that PANEL B: DAI the measurement can capture more information included by all measurements. Voters Gini 10k-100k >100k Delegate To simplify our empirical results, we only present results related to the mea- ΔReturn 0.01 0.00 0.01 -0.01 0.00 surements with highest ‘explained variance ratios’. (0.33) (0.17) (0.87) (-1.03) (0.36) ΔMktC 0.00 0.00 0.03*** -0.04*** 0.00 (0.00) (-0.08) (3.40) (-4.94) (-0.05) Where j=1,…,7, k =1,…, 4, and l =1,2. Volume 0.02* 0.01 0.05*** -0.05*** 0.03 The detailed description of the above set of factors8 is presented in ΔVolume_dex ( 0 1 .0 .6 2 8 ) ( 0 0 .0 .8 1 2 ) ( 0 6 .1 .7 2 0 * ) * * ( -0 -9 .1 .5 0 5 * ) * * ( 0 1 .0 .2 1 9 ) detail in Appendix A.1. For all statistically significant results, we run the (0.58) (0.64) (7.83) (-8.60) (0.27) Granger test to address potential reverse causality (see, Appendix A.2). ΔVolume_dex_usd 0.02 0.01 0.12*** -0.12*** 0.18*** We note that in all cases, the factors pass the Granger causality tests. (0.58) (0.63) (9.38) (-13.52) (3.99) Volume_l 0.02* 0.01 0.04*** -0.05*** 0.03 (1.71) (0.79) (6.28) (-9.07) (1.29) 4.2.1. Financial factors Volume_l_usd 0.02* 0.01 0.04*** -0.05*** 0.03 Maker governance polls are directly related to non-technical (1.70) (0.80) (6.24) (-9.01) (1.29) changes, e.g., adding a new collateral, to Maker protocol. These Note: This table reports the univariate regression coefficients and standard t- changes will add more financial functions or revise the parameters of statistics in parentheses for the financial factors of MKR (Panel A) and DAI transactions in the protocol. Therefore, it is crucial to examine whether (Panel B). *, ** and *** denote significance levels at the 10%, 5% and 1% levels, MKR and DAI factors, such as daily return, market cap, and trading respectively. The definitions of the factors are given in Table A.1. volume, are going to be affected by centralized governance, in the form of the metrics discussed in Section 3. Market capitalization, trading Table 8 volume, and daily return are commonly considered key characteristics Network factors (MKR, DAI). of cryptocurrencies (Liu et al., 2022). When studying the performance of PANEL A: MKR DeFi, market capitalization and trading volume also emerge as crucial indicators (Makridis et al., 2023). Measurements of governance Voters Gini 10k-100k >100k Delegate centralization can contribute significantly to research on the risks and ΔTotalWithBlc 0.00 0.00 0.00 0.00 0.00 returns associated with cryptocurrencies. Furthermore, we pay attention (-0.24) (0.47) (0.02) (0.03) (0.05) to substantial wealth transfers involving MKR and DAI on the Ethereum ΔNew -0.02 0.02 0.00 0.00 0.01 (-0.65) (0.96) (-0.04) (0.11) (0.23) blockchain, constructing two dependent variables, namely volume_l and ΔActive -0.01 0.01 0.00 0.00 0.00 volume_l_usd. These variables could be associated with manipulation (-0.38) (1.00) (-0.05) (0.13) (0.07) activities in crypto markets, such as pump-and-dump schemes (Dhawan ΔActiveRatio -0.02 0.03* 0.00 0.00 0.00 and Putnin¸ˇs, 2022). By investigating the relationship between gover- (-0.52) (1.75) (0.02) (0.05) (0.19) PANEL B: DAI nance centralization and large transactions, we can uncover additional Voters Gini 10k-100k >100k Delegate drivers behind significant transfers of crypto assets. Further studies can ΔTotalWithBlc -0.02 0.00 0.00 0.00 0.00 utilize our findings to examine whether governance centralization (-0.53) (-0.20) (0.14) (0.10) (-0.08) contributes to manipulation activities. ΔNew -0.08** 0.00 0.00 0.00 -0.01 In addition to the total transaction volume, we also examine trans- (-1.98) (-0.04) (-0.14) (0.28) (-0.16) ΔActive -0.05 -0.03 0.00 0.00 -0.01 action volumes on Decentralized Exchanges (DEXes) separately. Mak- (-1.11) (-1.02) (-0.09) (0.34) (-0.17) ridis et al. (2023) argue that CEX and DEX markets are segmented. ΔActiveRatio -0.03 0.04** 0.00 0.00 0.00 Aspris et al. (2021) contend that DEXes significantly differ from CEXes, (-0.84) (2.22) (-0.14) (0.14) (0.01) particularly in terms of the cryptocurrencies they list, which can lead Note: This table reports the univariate regression coefficients and standard t- crypto investors to have varying preferences when choosing crypto- statistics in parentheses for the network factors of MKR (Panel A) and DAI (Panel currency exchanges. By examining the trading volume of MKR and DAI B). *, ** and *** denote significance at the 10%, 5% and 1% levels, respectively. on DEXes, we aim to investigate whether governance centralization in The definitions of the factors are given in Table A.2. MakerDAO can exert influence on MKR/DAI traders’ preferences on cryptocurrency exchanges. 8 To avoid the problems of spurious regressions, we first examine if factors The findings of univariate regression for these factors are summa- are stationary. For the non-stationary variables, we choose the first differences rized in the following table. The two panels of the table bring forward of the variables instead. More details are given in Online Appendix 2. Beside the some interesting findings for the effects of centralized governance factors presented in Section 4, we further explore some other factors in online measures for MKR and DAI. We observe that 10k(cid:0) 100k and delegate appendix 8. 10

X. Sun et al. J o u r n a l o f F i n a n c i a l S t a b i l i ty73(2024)101286 Table 9 between DAI stability and MakerDAO governance. Twitter sentiment factors (MKR, DAI). All the above shows that centralized governance is significantly PANEL A: MKR evident for the financial factors relevant to MKR and DAI. In particular, holding centralization measurements (i.e., 10k-100k and >100k) appear Voters Gini 10k-100k >100k Delegate to exert influence on the trading volumes of MKR and DAI, while voting ΔNeutral 0.00 0.01 0.00 0.00 -0.01 centralization measurements (i.e., Voter and Gini) may not play a pri- (-0.26) (1.09) (0.05) (0.02) (-0.44) mary role. This distinction can be likened to the concept of ’buy-and- ΔNegative 0.00 0.00 0.00 0.00 0.00 (-0.03) (0.08) (0.04) (0.03) (-0.16) hold’ in corporate governance, where certain owners are hesitant to PANEL B: DAI actively intervene in governance matters despite observing performance Voters Gini 10k–100k >100k Delegate issues in a firm (Connelly et al., 2010). Given that not all MKR holders ΔNeutral 0.00 0.00 0.00 0.00 0.00 frequently participate in voting, it becomes evident that holding (0.70) (-0.44) (0.05) (0.02) (-0.06) ΔNegative 0.01 0.01 0.00 0.00 -0.01 centralization and voting centralization capture different dimensions of (0.16) (0.37) (-0.09) (0.05) (-0.24) governance centralization, consequently leading to dissimilar effects on the system. Note: This table reports the univariate regression coefficients and standard t- Generally, the results bring forward a trade-off between decentral- statistics in parentheses for the Twitter sentiment factors of MKR (Panel A) and ized governance and volume of these two tokens, and we also contend DAI (Panel B). *, ** and *** denote significance at the 10%, 5% and 1% levels, respectively. The definitions of the factors are given in Table A.3. that the influences of dominant decision makers can be complicated. Although the financial characteristics of coins are quite important, the have a significant positive effect on MKR volume, while >100k has the literature has shown that the researcher needs to expand on other in- dicators integral to the technical structure of coins, tokens, and protocols opposite. This conceptually means that the centralized voting power of in order to capture their potential fundamental value (Kraaijeveld and major MKR holders (i.e., those with 10,000 - 100,000 MKR holdings) De Smedt, 2020; Nadler and Guo, 2020; Liu and Tsyvinski, 2020; and delegates will boost MKR trading activities. However, if superma- Nakagawa and Sakemoto, 2022). For that reason, we next turn to other jority MKR holders (i.e., those with more than 100,000 MKR holdings) MKR and DAI related factors that are non-financial in nature to further accumulate higher MKR balances, the total volume, the volume on establish that centralized governance exists in the Maker protocol. DEXes, and the volume of large transactions will decrease. This finding could support the claimed value proposition of MakerDAO, i.e., decen- 4.2.2. Network and Twitter sentiment factors tralized governance. The above could parallel the findings of Meirowitz and Pi (2022), who analyze the shareholder’s dilemma through the lens In this subsection, we focus on factors capturing network charac- teristics and social media sentiment. In cryptocurrencies, network of voting and trading. Our results further imply that voting and trading adoption is a critical metric that can significantly influence long-term are not substitutes in some cases, and major stakeholders and super- development (Somin et al., 2018; Sockin and Xiong, 2023). Moreover, majority stakeholders can have different influences. In the context of DeFi is often seen as a more inclusive alternative to traditional finance DAO, major holders and supermajority holders can even have contrary (Cong et al., 2023). Governance centralization, however, appears to effects on the trading volume of governance tokens, which implies the mitigate this effect. Investigating the most important network statistics, complex relationship between governance and trading dynamics in such as total addresses, new addresses, active addresses, and their DeFi. active-to-total ratio should be investigated. This is examined by the For the case of DAI, we continue to observe significant effects of univariate regressions presented in Table 8. centralized governance. A higher number of voters has a notable impact It is clear that centralized voting affects MKR network factors. For on increasing the trading volume of DAI, including total volume and the both MKR and DAI, Gini exhibits significant positive effects on the volume of large transactions. Therefore, a higher voting participation active-to-total ratio. Therefore, centralized voting seems to be advan- rate proves advantageous for the DAI stablecoin. The DAI market cap is tageous. However, a higher number of voters is associated with a expected to be a metric of the performance of Maker protocol as well. We observe that 10k-100k is positively related to ΔMktC, while >100k has decrease in the growth of new addresses participating in DAI trans- actions. This implies that network adoption of DAI is negatively affected the opposite effect. Regarding the trading volume of DAI, 10k-100k and >100k show inverse influences as well. Therefore, though both major when MakerDAO governance is more decentralized. Furthermore, our analysis did not reveal how holding centralization is related to network holders and supermajority holders have centralized voting power factors. In summary, the network adoption of DAI and MKR appears to (compared with small MKR holders), their influences can be different. In benefit from voting centralization in MakerDAO governance, while a corporate governance, holding centralization often serves as a proxy for higher level of voting participation may not necessarily have a positive conflicting interests (Demsetz and Villalonga, 2001). Corporate owners impact on network factors. The relationship between governance with varying levels of ownership can exert different effects on firm centralization and network dynamics is complex. performance (Dalton et al., 2003), a parallel that resonates with our In terms of social media sentiment, we focus on Twitter. Currently, findings. Twitter is the main social media platform where DeFi investors express However, we do not observe significant results related to how their opinions, and Twitter sentiment analysis has proven to be a valu- governance centralization affects DAI price or volatility. A decreasing able tool for predicting cryptocurrency price movements (Kraaijeveld price for DAI, a stablecoin, can be a signal of depegging, implying that and Smedt, 2020; Naeem et al., 2020). The crypto community actively decentralized governance can cause problems for it. That in essence engages in discussions and exchanges on this platform, while the Maker defeats the main purpose of stablecoins, namely price stability. For protocol also maintains official Twitter accounts. Grover et al. (2019) example, Tether (USDT), a stablecoin mostly backed by cash and com- conducted a study on Twitter users’ discussions about blockchain and mon cash equivalents, suffered from depegging in late 2022. This created uncertainty on whether its reserves’ cushion would be sufficient found that these discussions predominantly focus on the benefits of to meet its obligations during a flood of USDT redemption requests.9 blockchain rather than its drawbacks. We anticipate that Twitter dis- cussions related to MKR and DAI will be influenced by MakerDAO’s Based on our analysis, we do not identify a significant relationship governance. By concentrating on Twitter sentiment factors provided by IntoTheBlock.com, we aim to explore whether governance centraliza- tion exerts an influence on the negative and neutral sentiment expressed 9 https://coingeek.com/tether-panics-as-loan-scrutiny-mounts-throws-alame by Twitter users. As shown in Table 9, our regression results did not da-under-bus/ reveal any statistically significant relationships. Based on our analysis, it 11 X. Sun et al. J o u r n a l o f F i n a n c i a l S t a b i l i ty73(2024)101286 Fig. 6. Proportions of three different collateral assets. Note: This figure illustrates the proportions of three different collateral assets locked in Maker protocol (Nov 18, 2019 – Oct 22, 2021), including ETH and stablecoins. The datasets are queried from dune.xyz. accounted for more than 25% of value of total collateral assets, and the Table 10 spikes of stablecoin ratio go above 50%. Collateral ratios. To explore how collateral assets are driven by centralized gover- Voters Gini 10k-100k >100k Delegate nance, we estimate the following regressions: ΔETH_ratio 0 (0 .0 .6 2 0 ) - ( 0 -1 .0 .1 2 5 ) 0 (2 .0 .0 2 4 * ) * - ( 0 -1 .0 .7 1 3 * ) - ( 0 -2 .0 .8 3 8 * ) * * Collateralt =β 0 +β 1 centralt +ε t (12) ΔStablecoin_ratio -0.01 0.02 -0.02** 0.01* 0.04*** (-0.29) (0.85) (-1.96) (1.82) (3.03) Where: Note: This table reports the univariate regression coefficients and standard t- statistics in parentheses for collateral ratios in Maker protocol. *, ** and *** • centralt ={Voterst ,Ginit ,10k(cid:0) 100kt ,>100kt ,Delegatet } denote significance at the 10%, 5% and 1% levels, respectively. The definitions of the factors are given in Table A.4 in Appendix 1. Given i, factors can be defined as a set: Collateralt ={ΔETHratiot ,ΔStablecoinratiot } appears that governance centralization within MakerDAO may not be The detailed description of the above set of factors is presented in generating substantial discussion on social media platforms and may not detail in Table A.4 in Appendix A.1. For all statistically significant re- be garnering significant attention from many investors. sults, we run the Granger test and the results do not suffer from reverse causality. 11 4.2.3. Collateral ratios The picture from Table 10 is clear. The distribution of MKR is related Finally, we investigate how centralized governance affects collateral to collateral ratios in Maker protocol. Increased voting power of >100k assets locked in Maker protocol. To initiate loans from Maker protocol, and Delegate can decrease the growth of ETH ratio, while 10k(cid:0) 100k has Maker users must lock collateral. Therefore, collateral assets accepted by the opposite effect. For stablecoins, 10k(cid:0) 100k and Delegate show the Maker protocol are an important issue in Maker governance. Moreover, contrary influences as well. Again, centralized voting power can affect if the locked collateral assets become risky, Maker protocol may be Maker protocol from the aspect of collateral assets, given that Maker affected. In March 2023, DAI was severely influenced by the depegging of USD Coin (USDC),10 since USDC is one of the most important governance decides collateral onboarding and offboarding. Further- more, MKR holders have dissimilar preferences of collaterals, which collateral assets in Maker protocol. may explain why major holders and supermajority holders influence the In this subsection, we examine how centralized governance relates to protocol differently. Overall, our two-stage analysis provides substantial components of collateral assets. We first consider two categories of empirical evidence of centralized governance in MakerDAO, as several collateral assets, including Ether (ETH) and stablecoins. The reason for significant univariate relationships are established across different including ETH classes of factors.12 The next section provides further robustness checks is intuitive. ETH, as the native cryptocurrency of Ethereum block- towards that end. chain, is one of the earliest accepted collateral assets in Maker protocol. Stablecoins play a crucial role in Maker protocol. For each category of collateral assets, we compute the proportion of its value (in USD) to the total value of locked collateral assets in Maker protocol. Fig. 6 shows that ETH was the dominant collateral before September 2020. Since September 2020, ETH ratio has been much lower, while stablecoins become important collaterals. After September 2020, stablecoins usually 11 More details about results for Granger causality test are given in Table A.9 in Appendix A.2. 10 https://cointelegraph.com/news/maker-dao-files-emergency-proposal-a 12 Given the extent of the factors and univariate regressions examined, we also ddressing-3–1b-usdc-exposure present a summary of the relationships in Online Appendix OA.3. 12 X. Sun et al. J o u r n a l o f F i n a n c i a l S t a b i l i ty73(2024)101286 Table 11 2-SLS IV regressions (financial factors – MKR). Panel A: Estimate 10k-100k using an instrument (1) (2) (3) (4) (5) (6) Volume Volume_dex Volume_l Volume_l_usd Off-chain 0.32*** 0.31** (6.02) (5.81) 10–100k 0.22 0.45*** 0.16 0.24* (1.01) (2.47) (0.63) (1.64) Durbin-Wu-Hausman test 0.20 4.36 0.00 0.00 p-value 0.66 0.04 0.95 0.95 Adj. R-sq -0.01 -0.51 0.02 0.16 N 126 127 127 127 Note: This table reports results of the 2-SLS IV regressions. Panel A, Columns (1) and (4) report the results of the following first stage regression: 10k(cid:0) 100kt =β 0 + β 1off (cid:0) chaint + ε t, where off(cid:0) chain is an instrumental variable. Columns (2)–(3) and columns (5)–(6) report the results of second stage: factort = β 0 + β 1 10k(cid:0)̂ 100kt +ε t. In Columns (1) and (4), partial F-statistics are reported in parentheses. In Columns (2)–(3) and (5)–(6), t-statistics are reported in parentheses. *, ** and *** denote significance at the 10%, 5% and 1% levels, respectively. Table 12 Regression discontinuity (summary). Panel A: MKR Panel B: DAI Measurements Financial Measurements Financial Network Factors Factors Factors Voters Voters Volume↑ ΔNew↓ Volume_l↑ Volume_l_usd↑ Gini Gini ΔActiveRatio↑ 10k-100k Volume↑ 10k-100k ΔMktC↑ Volume_dex↑ Volume↑ Volume_dex_usd↑ Volume_l↑ Volume_l↑ Volume_l_usd↑ Volume_l_usd↑ >100k Volume↓ >100k ΔMktC↓ Volume_dex↓ Volume↓ Volume_dex_usd↓ Volume_l↓ Volume_l↓ Volume_l_usd↓ Volume_l_usd↓ Delegate Volume_dex_usd↑ Delegate ΔVolume_dex↑ Volume_l_usd↑ ΔVolume_dex_usd↑ Note: This table summarizes the results for regression (13), where the KuCoin hack is constructed as a dummy. For example, the frist column and second row report Volume↑, which means that the increase of 10k-100k leads to a significant increase in Volume. The detailed regression results are presented in Online Appendix OA.5. 5. Robustness checks governance incorporates the ‘one-person, one-vote’ principle, unlike on-chain governance where voting rights depend on ownership of tokens 5.1. Addressing endogeneity: off-chain governance as an instrumental or exploitation of hashing power. In MakerDAO governance, forum variable (MKR) signal threads are functionally a warm-up for on-chain voting, so off-chain voting in these threads affects Maker protocol via the on-chain The empirical results presented in Section 4 could face criticism due voting that follows. This is supported by theoretical and empirical evi- to potential endogeneity concerns. To alleviate this issue, we use the dence (e.g., Dursun and Üstündag˘ 2021, Reijers et al. 2021 and Han instrumental variable approach and estimate two-stage least squares et al. 2023). For each thread, we document the first post date and the (2SLS) regressions. We construct an instrumental variable (IV) using number of unique voters, and the daily number of off-chain voters are datasets for forum signal threads, which are a part of the off-chain utilized as an IV for on-chain governance in MakerDAO. governance in the Maker protocol (Brennecke et al., 2022). Anyone To further validate off-chain voters as an IV, we conduct an analysis can participate in the discussion and voting in the threads. That means to examine the causal relationship between protocol performance and that, unlike the on-chain governance investigated previously, off-chain off-chain governance. Our expectation is that the specific factors of the governance does not require MKR in one’s account. Even people who IV, used as dependent variables, are not affected by the Maker protocol do not use blockchain can share their opinions and click an option in the in Section 4. To address this concern, we add the lagged terms of thread. For some signal threads, the informal discussion will eventually protocol-specific factors to the first stage regressions and subsequently turn to Maker IP, where participants can vote. re-estimate the regressions. The detailed results are presented in Online The results of signal threads are related to on-chain governance. Appendix OA.4 (see Tables OA.11 – OA.13). These results suggest that Zhao et al. (2022) describe how off-chain governance acts as the foun- off-chain governance is a valid IV. The following table presents first and dation of on-chain governance, and they also show that pre-voting dis- second stage regressions for the financial factors of MKR, and the results cussions on strategic decisions are beneficial for a DAO in certain cases. show consistency with findings in Section 4.13 Xu et al. (2023) show that off-chain governance has growing signifi- cance in DeFi governance by analyzing the content of off-chain discus- sion on forums. Reijers et al. (2021) also explain that off-chain 13 For the sake of space, the remaining 2SLS results for MKR and all the equivalent 2SLS results for DAI are provided in Online Appendix OA.4. 13 X. Sun et al. J o u r n a l o f F i n a n c i a l S t a b i l i ty73(2024)101286 Table 13 Before using an instrument, we first test if our instrument suffers Categories of Maker governance polls. from weak instruments concerns. The results for the first stage re- Number of polls gressions show that our instrument can be used in 2SLS regressions. Then, we are curious whether measurements of centralized governance Risk Parameter 252 are endogenous to factors for MKR and DAI. To test the endogeneity, we Ratification Poll 27 Inclusion Poll 70 apply Durbin-Wu-Hausman test. Simply, the test examines whether Collateral Onboarding 50 predictor variables in the univariate regressions are endogenous. Since Collateral Offboarding 2 the null hypothesis is that endogeneity does not exist, usually, we do not Greenlight 146 observe endogeneity between our measurements and most factors for Real World Asset 28 Misc Governance 18 MKR and DAI, meaning that the corresponding OLS regressions in Sec- Misc Funding 3 tion 4 are reliable. For results where endogeneity is observed (e.g., MakerDAO Open Market Committee 11 Column (4), Table 11), we compare results of 2SLS regressions and re- MIP 106 sults in Section 4, and the findings are consistent. Therefore, the mea- Budget 25 surements of centralized governance are generally not found to be Oracle 38 System Surplus 6 endogenous to MKR and DAI factors. DAI Direct Deposit Module 1 Multi-chain Bridge 1 5.2. Regression discontinuity Technical 17 Auction 20 Delegates 0 In September 2020, there was a significant security breach at Peg Stability Module 11 KuCoin, a centralized exchange, resulting in the theft of approximately Core Unit Onboarding 17 $280 million worth of cryptocurrencies (Hui and Zhao, 2021). This Dai Savings Rate 28 Black Thursday 4 event is widely regarded as an exogenous shock to the DeFi market, as Multi-Collateral DAI Launch 5 highlighted by previous research, such as Makridis et al. (2023). Prioritization Sentiment 2 Following their approach, we construct a dummy variable ’shock’ to Note: This table reports the number of governance polls (Poll 16 – Poll 663) in examine the impact of the exogenous shock on our results in Section 4. different categories. One poll can have multiple labels. The value of ’shock’ equals to 1 during the period spanning from Poll 287 (deployed on September 14, 2020) to Poll 412 (deployed on January 11, 2021), when the largest voting share was lower than the Table 14 average. For the remaining sample, the value of ‘shock’ equals to 0. We Descriptive statistics of ‘Voters’ and ‘Gini’ based on ‘risk parameter’ polls. estimate the following regression14: Voters TotalVotes LargestShare Order Gini factori,t =β 0 +β 1 centralt +β 2 shockt +ε t (13) Mean 51.58 81514.40 0.56 0.41 0.33 Median 40.00 54565.51 0.53 0.40 0.00 Where: Maximum 206 365383.75 0.96 0.93 0.94 Minimum 7 259.74 0.26 0.00 0.00 Std 39.10 75042.97 0.17 0.22 0.41 • i ={MKR,DAI} N of obs. 107 107 107 107 107 • centralt ={Voterst ,Ginit ,10k(cid:0) 100kt ,>100kt ,Delegatet } Note: This table presents the descriptive statistics of measurement of governance centralization in Maker, and the measurements are calculated using ‘risk Given i, factors can be defined as a set: parameter’ polls. In the first four columns, we first calculate these measurements factori,t ={financiali,j,t ,networki,k,t ,Twittersentimenti,l,t } for each poll and then convert them to daily level measurements. For example, we first calculate the number of voters for every ‘risk parameter’ poll, then we Where j=1,…,7, k =1,…, 4, and l =1,…,2. add them to get the number of daily voters. Daily Gini is calculated using the We expect that the influences of governance centralization will not maximum likelihood method described in Section 3. be affected by the KuCoin shock. In other words, the coefficients of centralization measurements in regression (13) should be consistent Table 15 with the results in Section 4. The table below summarizes the results, Measurements of governance centralization based on ‘risk parameter’ polls which aligns with our findings in Section 4.15 (summary). Panel A: MKR 5.3. Certain types of governance polls Measurements Financial Factors Network Factors Twitter Sentiment Factors Governance polls can be categorized. Their labels are publicly Voters Volume↑ observable on the Maker governance forum. Generally, the label reflects Volume_dex_usd↑ on what a governance poll is focused on. For example, ‘collateral Volume_l_usd↑ onboarding’ polls are about new collateral assets that can be used to Gini Volume_dex_usd↓ ΔNew↑ ΔNeutral↑ Volume_l_usd↓ ΔActive↑ ΔNegative↑ initiate loans from Maker protocol, while ‘MIP’ polls discuss improve- Panel B: DAI ment proposals of Maker protocol (see Table 13). We focus on the label Measurements Financial Network Factors Twitter Sentiment of ‘risk parameter’, which has the most governance polls. Using the Factors Factors subset, we calculate the centralization measurements again. Table 13 Voters ΔVolume_dex↑ ΔTotalWithBlc↓ ΔVolume_dex_usd↑ summarizes the descriptive statistics. Generally, ‘risk parameter’ polls Gini ΔReturn↑ ΔMktC↓ Note: This table reports the relationship between measurements of centralized 14 We also estimate the regression:factori,t = β 0 + β 1centralt + β 2shockt + voting power and the factors of MKR and DAI. The measurements are calculated β 3centralt × shockt + ε t. These results can be provided upon request and using the datasets for governance polls with the label ‘risk parameter’. The resemble the ones presented in the manuscript. detailed regression results are presented in Online Appendix OA.6. 15 More detailed results for regression discontinuity analysis are presented in Online Appendix OA.5. 14

X. Sun et al. J o u r n a l o f F i n a n c i a l S t a b i l i ty73(2024)101286 have fewer voters and fewer total votes, while the daily Gini coefficient investigation whether voting centralization and holding centralization and the largest voters’ share is higher. The preliminary results imply that show inconsistent influences across various DAOs. governance is even more centralized in these polls. The optimal governance structure for DeFi remains an open question Then, using the subset, we re-estimate the univariate regressions in the current landscape. When governance is excessively centralized, it mentioned in Section 4. Given that ‘risk parameter’ polls can decide key leaves DeFi systems vulnerable to manipulation by a select few agents. variables of Maker protocol, e.g., interest rates and debt ceilings, it is not This paper implies that centralized governance in DAO to some extent very surprising that governance centralization in these polls can affect may contribute to trading activities of the underlying DeFi protocol. financial factors (for both MKR and DAI). Overall, the results prove that With our findings, we make a compelling case in favor of the argument more decentralized governance (e.g., a higher number of voters) has that decentralization in DeFi platforms is an illusion and that the trade- positive effects (e.g., an increase in trading volume), while higher Gini off between market performance and decentralization exists. The trade- can negatively affect MKR and DAI, such as lower trading volume of off is similar to the ones observed in the corporate world, where unex- MKR and slower growth of market capitalization of DAI. pected results of governance processes may be caused by different Overall, this subsection shows that decentralized governance is ad- preferences of decision makers (Garlappi et al., 2017; Donaldson et al., vantageous. Further studies can combine the contents of governance 2020). polls in different categories with the performance of Maker protocol, Although our findings appear conceptually and empirically robust, revealing which issues are more crucial. We can also investigate voters’ they should be interpreted with their limitations in mind. First, the voting patterns in different types of polls, in an effort to reveal their identity of the dominant voters is unknown. Anonymity is another private benefits. character of blockchain and DeFi, and we may not know the identity of voters until they are willing to announce it. Therefore, sybil attacks are a 6. Conclusion potential problem in our analysis. A single entity can control multiple identities (e.g., blockchain addresses), which can undermine the secu- Decentralization is a crucial innovation of blockchain, and the rapid rity of a blockchain-based system by gaining an unfair and overly growth of DeFi relies on decentralization. Complete decentralization is influential position (Douceur, 2002). Given the difficulty of clustering theoretically impossible (Abadi and Brunnermeier, 2022), and empirical multiple addresses controlled by an entity, the analysis in this paper is evidence of centralization is detected in different layers of blockchain based on Ethereum addresses. Further research can dive into DAO (Sai et al., 2021). In this paper, we focus on governance in DeFi and governance more deeply by implementing more advanced techniques particularly on the Maker protocol, which is governed by MakerDAO. for clustering multiple addresses controlled by a single entity. Second, Decentralized governance is a crucial domain for DeFi and Maker pro- when studying DAI flows to different on-chain financial systems, we do tocol is an ideal case since its voting history is considered transparent not track the subsequent transactions of transferred DAI. Nadler and and precise (Beck et al., 2018). By examining Maker governance polls, Sch¨ar (2020) propose mapping algorithms that can expose the sources of we find that voters are centralized in a small group and voting power is cryptocurrencies stored in an on-chain financial system. This can be a unequally distributed among these voters. In most voting activities, the solution towards revealing the real owners of cryptocurrencies traded on largest voters could account for a significant proportion of votes. Pre- DeFi applications. Further studies can investigate the effects of gover- viously, Gervais et al. (2014) and Azouvi et al. (2018) argue that a few nance centralization in DAO by applying similar methods. Third, we key developers have unilateral decision-making power in blockchain focus on casted votes in MakerDAO governance without delving into the governance. This problem might derive from the requirement of pro- potential voting power of MKR holders. After tracking the MKR balance gramming skills. Our results expand the discussion to the of all MKR holders on Ethereum blockchain, we find that a significant token-weighted voting system in DeFi. Particularly in Maker, any MKR portion of MKR holders choose to be passive holders rather than actively holder can easily participate in governance by clicking an option on the engaging in governance.16 This leads to two questions for further website, which would indicate that governance would be more decen- research: Why do the MKR holders not actively engage in governance? tralized. Interestingly, our results show that governance in Maker pro- What influences do large stakeholders exert on DeFi and DAO? Drawing tocol is highly centralized. parallels from corporate finance literature (e.g., Meirowitz and Pi, To show that, we first construct two categories of centralization 2022), which acknowledges that shareholders may either engage in measurements, namely voting centralization (including Voters and Gini) governance or trade shares as part of their investment portfolios based and holding centralization (including 10k-100k, >100k, and Delegate). on private information, we posit that a similar phenomenon might exist Voting centralization exerts complex influences on Maker protocol. For in DeFi, given that governance tokens are tradable cryptocurrencies. example, a higher number of voters can lead to an increase of trading Moreover, the distribution of MKR remains highly centralized over time, volume of DAI, however, it can also negatively affect the growth of as indicated by the gini coefficient of approximately 96.85% during the network adoption of DAI. But a higher Gini, usually regarded as a signal sample period.17 This emphasizes the need for understanding the impact of centralized governance, can bring forward more active investors for and effects of ownership concentration in DeFi and DAO. Addressing both MKR and DAI. The findings imply that governance centralization in these questions would contribute to a more comprehensive under- DAO resembles a double-edged sword. standing of DAO governance. Finally, we do not know whether the au- Holding centralization also exerts influence on Maker protocol. This thors of Maker IPs are dominant voters. If a dominant voter proposes observation is related to research on ownership structure, where large changes to the Maker protocol, the aim of such proposals might be tied stakeholders often exert influence through private engagements with to their own vested interests. With their large voting power, this could management (Jensen and Warner, 1988; Connelly et al., 2010; Fichtner lead to further centralization of power and potential collusion during the et al., 2017). In DAOs, owners (i.e., governance token holders) and development of Maker protocol. Currently, writing Maker IPs requires managers (i.e., participants in DAO governance) are theoretically both programming skills and understanding of technical structure of identical, so collusion with management is not necessarily a concern. DeFi. Assuming that not many voters have such competence, key But our findings show that in DAOs, ownership structure, particularly the governance power held by large stakeholders, remains crucial. The influences of large stakeholders are reflected in on-chain voting. 16 The historical distribution of MKR token can be queried utilizing the Furthermore, our research presents an intriguing finding: Major stake- datasets provided by Steakhouse.Financial: https://dune.com/steakhouse/mk holders and supermajority stakeholders have opposing influences on r-token-decentralization-and-distribution-metrics certain cryptocurrencies and the collateral status of DeFi protocol. 17 We calculate the historical gini coefficient based on the distribution of MKR Future research can explore why that is the case, while it is worthy of token on Ethereum blockchain. The dataset can be provided upon request. 15 X. Sun et al. J o u r n a l o f F i n a n c i a l S t a b i l i ty73(2024)101286 developers may be the only people that can guide voters by proposing addresses publicly available so that users can detect suspicious activities specific Maker IPs, implying that the centralized power of core de- of developers. velopers exists in DeFi. This could also be supported by studies sug- gesting that delegating tasks to a group of experts can lead to better Funding sources aggregation of information (Fehrler and Janas, 2021). As things stand, though, Maker users rely on developers to provide detailed proposals, This research did not receive any specific grant from funding the aims of codes and explanations of all possible outcomes in an un- agencies in the public, commercial, or not-for-profit sectors. derstandable way. Another possible solution is to make IP authors’ Appendix A.1 Description of utilized factors The following tables summarize the factors used in our univariate regressions. Table A.1 Financial factors for MKR and DAI Description Return Daily return MktC Price (in USD) of the tokens times the circulating supply Volume Total amount (in tokens) of tokens transferred on Ethereum blockchain within a day Volume_usd Total amount (in USD) of tokens transferred on Ethereum blockchain within a day Volume_dex Sum of the amount (in tokens) traded on Decentralized Exchanges (DEXes) Volume_dex_usd Sum of the amount (in USD) traded on Decentralized Exchanges (DEXes) Volume_l Aggregated daily volume, measured in tokens from on-chain transactions of more than $100,000 Volume_l_usd Aggregated daily volume, measured in USD from on-chain transactions of more than $100,000 Note: The factors are provided by intotheblock.com. Table A.2 Network factors for MKR and DAI Description TotalWithBlc The number of addresses that actually have a balance New The number of new addresses created daily Active The number of addresses that made a transaction Active ratio The percentage of addresses with a balance of tokens that made a transaction during a given period (Active Addresses / Addresses with a Balance). Note: The factors are provided by intotheblock.com. Table A.3 Twitter sentiment factors for MKR and DAI Description Positive The number of tweets that are related to a given token have a positive connotation. Neutral The number of tweets that are related to a given token have a neutral connotation. Negative The number of Tweets that are related to a given token have a negative connotation. Note: The Twitter sentiment factors utilize machine learning algorithm to determine if the texts used in the Tweets related to a given token have a positive, neutral or negative connotation. The factors are computed and provided by intotheblock.com. Table A.4 Collateral ratios Description ETH_ratio The value (in USD) of ETH locked as collateral divided by the total value (in USD) of locked collateral in Maker protocol Stablecoin_ratio The value (in USD) of stablecoins locked as collateral divided by the total value (in USD) of locked collateral in Maker protocol Note: We focus on three types of collateral assets, including ETH, stablecoins and Wrapped Bitcoin (WBTC). The variables are queried on dune.xyz. A.2 Results for Granger test The following tables summarize the results for the Granger test. For the measurements based on governance polls, the number of observations is not enough for implementing the Granger test. Therefore, we fill the null values using linear interpolation. 16 X. Sun et al. J o u r n a l o f F i n a n c i a l S t a b i l i ty73(2024)101286 Table A.5 Granger test results for MKR (linear interpolation) PANEL A: Network Null Hypothesis Obs. df F-stat. Prob. Null Hypothesis F-stat. Prob. Gini does not Granger Cause ΔActive ratio 807 2 0.09 0.96 ΔActive ratio does not Granger Cause Gini 4.16 0.13 Note: This table reports the results for Granger tests based on Vector Autoregression (VAR) models. Column ‘df’ shows the optimal lag order. Using the optimal lag order, we run Granger tests for the hypotheses stemming from our empirical findings. For each test, both F-statistics and probability are presented. Table A.6 Granger test results for MKR (measurements related to MKR distribution) PANEL A: Financial factors Null Hypothesis Obs. df F-stat. Prob. Null Hypothesis F-stat. Prob. 10–100k does not Granger Cause Volume 804 6 1.70 0.12 Volume does not Granger Cause 10–100k 1.19 0.31 10–100k does not Granger Cause Volume_dex 806 4 3.43 0.01 Volume_dex does not Granger Cause 10–100k 0.15 0.96 10–100k does not Granger Cause Volume_dex_usd 801 9 1.08 0.37 Volume_dex_usd does not Granger Cause 10–100k 0.25 0.99 10–100k does not Granger Cause Volume_l 804 6 2.02 0.06 Volume_l does not Granger Cause 10–100k 1.47 0.19 10–100k does not Granger Cause Volume_l_usd 799 11 0.65 0.79 Volume_l_usd does not Granger Cause 10–100k 0.44 0.94 >100k does not Granger Cause Volume 804 6 1.22 0.29 Volume does not Granger Cause >100k 0.82 0.55 >100k does not Granger Cause Volume_dex 806 4 3.45 0.01 Volume_dex does not Granger Cause >100k 0.60 0.66 >100k does not Granger Cause Volume_dex_usd 801 9 1.18 0.30 Volume_dex_usd does not Granger Cause >100k 0.36 0.95 >100k does not Granger Cause Volume_l 804 6 1.74 0.11 Volume_l does not Granger Cause >100k 0.89 0.50 >100k does not Granger Cause Volume_ l_usd 799 11 0.83 0.61 Volume_l_usd does not Granger Cause >100k 0.10 1.00 Delegate does not Granger Cause Volume_dex_usd 319 1 5.08 0.02 Volume_dex_usd does not Granger Cause Delegate 0.05 0.82 Delegate does not Granger Cause Volume_ l_usd 319 1 3.28 0.07 Volume_ l_usd does not Granger Cause Delegate 0.59 0.44 Note: This table reports the results for Granger tests based on Vector Autoregression (VAR) models. Column ‘df’ shows the optimal lag order. Using the optimal lag order, we run Granger tests for the hypotheses stemming from our empirical findings. For each test, both F-statistics and probability are presented. Table A.7 Granger test results for DAI (linear interpolation) PANEL A: Financial factors Null Hypothesis Obs. df F-stat. Prob. Null Hypothesis F-stat. Prob. Voters does not Granger Cause Volume 703 7 0.20 0.99 Volume does not Granger Cause Voters 1.64 0.12 Voters does not Granger Cause Volume_l 704 6 0.24 0.97 Volume_l does not Granger Cause Voters 1.63 0.14 PANEL B: Network Null Hypothesis Obs. df F-stat. Prob. Null Hypothesis F-stat. Prob. Voters does not Granger Cause ΔNew 704 6 1.07 0.38 ΔNew does not Granger Cause Voters 0.98 0.44 Gini does not Granger Cause ΔActiveRatio 695 15 0.23 1.00 ΔActiveRatio does not Granger Cause Gini 0.36 0.99 Note: This table reports the results for Granger tests based on Vector Autoregression (VAR) models. Column ‘df’ shows the optimal lag order. Using the optimal lag order, we run Granger tests for the hypotheses stemming from our empirical findings. For each test, both F-statistics and probability are presented. Table A.8 Granger test results for DAI (measurements related to MKR distribution) PANEL A: Financial factors Null Hypothesis Obs. df F-stat. Prob. Null Hypothesis F-stat. Prob. 10–100k does not Granger Cause ΔMktC 706 4 3.97 0.00 ΔMktC does not Granger Cause 10–100k 0.22 0.93 10–100k does not Granger Cause Volume 706 4 3.96 0.00 Volume does not Granger Cause 10–100k 0.22 0.93 10–100k does not Granger Cause Volume_l 706 4 3.74 0.01 Volume_l does not Granger Cause 10–100k 0.22 0.93 10–100k does not Granger Cause Volume_l_usd 706 4 3.71 0.01 Volume_l_usd does not Granger Cause 10–100k 0.22 0.93 >100k does not Granger Cause ΔMktC 706 4 4.60 0.00 ΔMktC does not Granger Cause >100k 0.65 0.63 >100k does not Granger Cause Volume 709 1 65.58 0.00 Volume does not Granger Cause >100k 0.00 1.00 >100k does not Granger Cause Volume_l 709 1 60.72 0.00 Volume_l does not Granger Cause >100k 0.00 0.99 >100k does not Granger Cause Volume_ l_usd 709 1 59.98 0.00 Volume_l_usd does not Granger Cause >100k 0.00 0.99 Delegate does not Granger Cause ΔVolume_dex 315 5 0.57 0.72 ΔVolume_dex does not Granger Cause Delegate 0.53 0.75 Delegate does not Granger Cause ΔVolume_dex_usd 315 5 0.57 0.72 ΔVolume_dex_usd does not Granger Cause Delegate 0.53 0.75 Note: This table reports the results for Granger tests based on Vector Autoregression (VAR) models. Column ‘df’ shows the optimal lag order. Using the optimal lag order, we run Granger tests for the hypotheses stemming from our empirical findings. For each test, both F-statistics and probability are presented. Table A.9 Granger test results for collateral ratios Null Hypothesis Obs. df F-stat. Prob. Null Hypothesis F-stat. Prob. 10–100k does not Granger Cause ΔETH_ratio 700 4 2.80 0.03 ΔETH_ratio does not Granger Cause 10–100k 0.74 0.57 10–100k does not Granger Cause ΔStablecoin_ratio 700 4 1.95 0.10 ΔStablecoin_ratio does not Granger Cause 10–100k 1.53 0.19 >100k does not Granger Cause ΔETH_ratio 698 6 3.52 0.00 ΔETH_ratio does not Granger Cause >100k 1.09 0.36 >100k does not Granger Cause ΔStablecoin_ratio 698 6 2.79 0.01 ΔStablecoin_ratio does not Granger Cause >100k 1.90 0.08 Delegate does not Granger Cause ΔETH_ratio 319 1 4.94 0.03 ΔETH_ratio does not Granger Cause Delegate 0.65 0.42 Delegate does not Granger Cause ΔStablecoin_ratio 319 1 4.31 0.04 ΔStablecoin_ratio does not Granger Cause Delegate 1.12 0.29 17 X. Sun et al. J o u r n a l o f F i n a n c i a l S t a b i l i ty73(2024)101286 Note: This table reports the results for Granger tests based on Vector Autoregression (VAR) models. Column ‘df’ shows the optimal lag order. Using the optimal lag order, we run Granger tests for the hypotheses stemming from our empirical findings. For each test, both F-statistics and probability are presented. Appendix B. Supporting information Supplementary data associated with this article can be found in the online version at doi:10.1016/j.jfs.2024.101286. References Grover, P., Kar, A.K., Janssen, M., Ilavarasan, P.V., 2019. Perceived usefulness, ease of use and user acceptance of blockchain technology for digital transactions – insights from user-generated content on twitter. Enterp. Inf. Syst. 13 (6), 771–800. Abad 2 w i 5 , 2 4 J 5 0 . 4 , 7 0 B ) 7 . r 〉 u N . n a n ti e o r n m a e l i B er u , r M ea . u , 2 o 0 f 2 E 2 c . o B n l o o m ck i c c h R ai e n s e E a c r o c n h o . m 〈h ic t s t p ( s N :/ B / E w R w W w o .n rk b i e n r g .o P rg a / p p e a r p N er o s . / Han, c J o . m , L / e so e, l 3 J / ., P & ap L e i r , s T .c . f ( m 2 ? 0 a 2 b 3 s ) t . r D ac A t O _id G = o 4 v 3 er 4 n 6 a 5 n 8 c 1 e, ( S A S c R c N es . s A e v d a : i 2 la 5 b l S e e a p t t : e h m t b tp e s r : / 2 / 0 p 2 a 3 p ) e . rs.ssrn. Harvey, C.R., Ramachandran, A., Santoro, J., 2021. DeFi and the Future of Finance. John Abdikerimova, S., Feng, R., 2022. Peer-to-peer multi-risk insurance and mutual aid. Eur. Wiley & Sons. J. Oper. Res. 299 (2), 735–749. Howell, S., Niessner, M., Yermack, D., 2020. Initial Coin Offerings: Financing Growth Anker-Sørensen, L. and Zetzsche, D.A. (2021). From Centralized to Decentralized Finance: with Cryptocurrency Token Sales. Rev. Financ. Stud. 33 (9), 3925–3974. The Issue of ’Fake-defi’, SSRN. Available at: https://papers.ssrn.com/sol3/papers. cfm?abstract_id=3978815/ (Accessed: 29 March 29 2023). Hsieh, Y., (JP) Vergne, J., Wang, S., 2017. The Internal and External Governance of Blockchain-based Organizations. Bitcoin And Beyond. Taylor and Francis Group, Anyfantaki, S., Arvanitis, S., Topaloglou, N., 2021. Diversification benefits in the pp. 48–68. cryptocurrency market under mild explosivity. Eur. J. Oper. Res. 295 (1), 378–393. Hui, A., & Zhao, W. (2021). Over $280M Drained in KuCoin Crypto Exchange Hack. Aramonte, S., Huang, W., Schrimpf, A., 2021. DeFi Risks and the Decentralisation Illusion (BIS Qartely Review, December 2021). Bank Int. Settl. 〈https://www.bis.or Available at: https://www.coindesk.com/markets/2020/09/26/over-280m-drained- g/publ/qtrpdf/r_qt2112b.htm〉. in-kucoincrypto-exchange-hack/ (Accessed: 24 August 2023). Jensen, M.C., Warner, J.B., 1988. The distribution of power among corporate managers, Aspris, A., Foley, S., Svec, J., Wang, L., 2021. Decentralized Exchanges: The “Wild West” shareholders, and directors. J. Financ. Econ. 20, 3–24. of Cryptocurrency Trading. Int. Rev. Financ. Anal. 77, 101845. Jentzsch, C. (2016). Decentralized Autonomous Organization to Automate Governance Azouvi, S., Maller, M., Meiklejohn, S., 2018. Egalitarian Society or Benevolent [White paper]. https://lawofthelevel.lexblogplatformthree.com/wp-content/ Dictatorship: The State of Cryptocurrency Governance. Financ. Cryptogr. Data Secur. uploads/sites/187/2017/07/WhitePaper-1.pdf. 127–143. Bai, Q., Zhang, C., Xu, Y., Chen, X., & Wang, X. (2020). Evolution of Ethereum: A Temporal Jiang p , l a S t . f , o L rm i, Y s . t , a W bil a i n ty g , a S n . d , Z e h ffi a c o i , e L n . c , y 2 . 0 E 2 u 2 r . . B J l . o O ck p c e h r. a i R n e c s. o m 29 p 6 e t ( i 3 ti ) o , n 1 : 0 t 8 h 4 e – t 1 r 0 a 9 d 7 eo . ff between Graph Perspective. arXiv.org. Available at: https://arxiv.org/abs/2001.05251/ Karim, S., Lucey, B., Naeem, M., Uddin, G., 2022. Examining the Interrelatedness of (Accessed: 29 August 2021). Bartoletti, M., Chiang, J., & Lluch-Lafuente, A. (2020). SoK: Lending Pools in Decentralized NFTs, DeFi Tokens and Cryptocurrencies. Financ. Res. Lett., 102696 Kraaijeveld, O., De Smedt, J., 2020. The Predictive Power of Public Twitter Sentiment for Finance. arXiv.org. Available at: https://arxiv.org/abs/2012.13230/ (Accessed: 29 Forecasting Cryptocurrency Prices. J. Int. Financ. Mark., Inst. Money 65, 101188. August 29 2021). Lee, J.Y., 2019. A Decentralized Token Economy: How Blockchain and Cryptocurrency Beck, R., Müller-Bloch, C., King, J., 2018. Governance in the blockchain economy: a Can Revolutionize Business. Bus. Horiz. 62 (6), 773–784. framework and research agenda. J. Assoc. Inf. Syst. 19 (10), 1020–1034. Liu, Y., Tsyvinski, A., 2020. Risks and Returns of Cryptocurrency. Rev. Financ. Stud. 34 Bollaert, H., Lopez-de-Silanes, F., Schwienbacher, A., 2021. Fintech and access to (6), 2689–2727. finance. J. Corp. Financ. 68, 101941. Liu, Y., Tsyvinski, A., Wu, X., 2022. Common Risk Factors in Cryptocurrency. J. Financ. Brams, S., 2008. Mathematics and Democracy. Princeton University Press, Princeton. 77 (2), 1133–1177. Brams, S., Fishburn, P., 2002. Chapter 4 Voting Procedures. Handbook Of Social Choice MakerDAO. (2020). The Maker Protocol White Paper [White paper]. Available at: and Welfare. Elsevier, pp. 173–236. Brennecke, M., Guggenberger, T., Schellinger, B., & Urbach, N. (2022). The De-Central https://makerdao.com/en/whitepaper/ (Accessed: 11 February 2022). MakerDAO. (2021). A Guide to Participating in MakerDAO Governance. Available at: Bank in Decentralized Finance: A Case Study of Makerdao, Handle Proxy. Available at: https://makerdao.world/en/learn/governance/how-voting-works/ (Accessed: 11 http://hdl.handle.net/10125/80077 (Accessed: 07 September 2023). Carte A r v , a N il . a , b & le J a e t n : g h , t t L p . s ( : 2 // 0 p 2 a 1 p , e A rs u . g ss u r s n t .c 6 o ). m D / E so F l I 3 P / r p o a t p oc e o rs l . R cf i m sk ? s: a b T s h t e r a P c a t_ r i a d d = ox 3 8 o 6 f 6 D 6 e 9 fi 9 . SSRN. Mak F ri e d b i r s u , a C r . y A . 2 , 0 F 2 r 2 o¨w ). is, M., Sridhar, K., Bo¨hme, R., 2023. The Rise of Decentralized Cryptocurrency Exchanges: Evaluating the Role of Airdrops and Governance Tokens. (Accessed: 22 March 22 2023). J. Corp. Financ. 79, 102358. Choi, T.M., Guo, S., Liu, N., Shi, X., 2020. Optimal Pricing in On-demand-service- Meirowitz, A., Pi, S., 2022. Voting and trading: The shareholder’s dilemma. J. Financ. platform-operations with Hired Agents and Risk-sensitive Customers in the Econ. 146 (3), 1073–1096. Blockchain Era. Eur. J. Oper. Res. 284 (3), 1031–1042. Cong, L.W., Tang, K., Wang, Y., & Zhao, X. (2023, February 13). Inclusion and Momtaz, P.P. (2022). How Efficient is Decentralized Finance (DeFi)?. Available at: http://dx.doi.org/10.2139/ssrn.4063670/ (Accessed: 15 April 2022). Democratization through Web3 and Defi? Initial Evidence from the Ethereum Ecosystem. Nadler, M., & Sch¨ar, F. (2020, December 16). Decentralized Finance, Centralized NBER. https://www.nber.org/papers/w30949. Ownership? An Iterative Mapping Process to Measure Protocol Token Distribution. arXiv. Cong, L., Li, Y., Wang, N., 2020. Tokenomics: dynamic adoption and valuation. Rev. org. Available at: https://arxiv.org/abs/2012.09306/ (Accessed: 22 March 2023). Financ. Stud. 34 (3), 1105–1155. Nadler, P., Guo, Y., 2020. The Fair Value of a Token: How Do Markets Price Connelly, B.L., Hoskisson, R.E., Tihanyi, L., Certo, S.T., 2010. Ownership as a Form of Cryptocurrencies? Res. Int. Bus. Financ. 52, 101108. Corporate Governance. J. Manag. Stud. 47 (8), 1561–1589. Naeem, M.A., Mbarki, I., Suleman, M.T., Vo, X.V., Shahzad, S.J., 2020. Does Twitter Dalton, D.R., Daily, C.M., Certo, S.T., Roengpitya, R., 2003. Meta-analyses of financial Happiness Sentiment Predict Cryptocurrency? Int. Rev. Financ. 21 (4), 1529–1538. performance and equity: fusion or confusion? Acad. Manag. J. 46 (1), 13–26. Nakagawa, K., Sakemoto, R., 2022. Cryptocurrency network factors and gold. Financ. Demsetz, H., Villalonga, B., 2001. Ownership structure and corporate performance. Res. Lett. 46, 102375. J. Corp. Financ. 7 (3), 209–233. Dhawan, A., Putnin¸ˇs, T.J., 2022. A new wolf in town? pump-and-dump manipulation in Nakamoto, S. (2008). Bitcoin: A Peer-to-peer Electronic Cash System. Available at: https:// bitcoin.org/bitcoin.pdf/ (Accessed: 22 March 2023). cryptocurrency markets. Rev. Financ. 27 (3), 935–975. Qin, K., Zhou, L., Afonin, Y., Lazzaretti, L., & Gervais, A. (2021). CeFi vs. DeFi – Donaldson, J.R., Malenko, N., Piacentino, G., 2020. Deadlock on the board. Rev. Financ. Comparing Centralized to Decentralized Finance. arXiv.org. Available at: https://arxiv. Stud. 33 (10), 4445–4488. org/abs/2106.08157 (Accessed: 23 February 2022). Dorfman, R., 1979. A Formula for the Gini Coefficient. Rev. Econ. Stat. 61 (1), 146–149. Reijers, W., Wuisman, I., Mannan, M., 2021. Now the Code Runs Itself: On-chain and Off- D D o u u rs c u e n u , r , T J ., . R Ü . s , t 2 ü 0 n 0 d 2 a . g˘ , ‘ T B h . e B . S , y 2 b 0 i 2 l 1 A . t t A a c N k o ’. v P el e e F r r - a t m o- e P w ee o r r k S y f s o t r . P 2 o 5 l 1 ic – y 2 6 B 0 a . s ed On-chain chain Governance of Blockchain Technologies. Topoi 40 (4), 821–831. Sai, A., Buckley, J., Fitzgerald, B., Gear, A., 2021. Taxonomy of centralization in public Governance of Blockchain Networks. Inf. Process. Manag. 58 (4), 102556. blockchain systems: a systematic literature review. Inf. Process. Manag. 58 (4), Fama, E., Jensen, M., 1983. Separation of ownership and control. J. Law Econ. 26 (2), 102584. 301–325. Sockin, M., Xiong, W., 2024. A Model of Cryptocurrencies. Manag. Sci. (forthcoming). Fehrler, S., Janas, M., 2021. Delegation to a Group. Manag. Sci. 67 (6), 3714–3743. Somin, S., Gordon, G., Altshuler, Y., 2018. Network Analysis of ERC20 Tokens Trading Fichtner, J., Heemskerk, E.M., Garcia-Bernardo, J., 2017. Hidden power of the big three? on Ethereum Blockchain. Unifying Themes Complex Syst. IX 439–450. passive index funds, re-concentration of corporate ownership, and new. Financ. Risk. Taleb, N. (2015). How to (Not) Estimate Gini Coefficients for Fat Tailed Variables. arXiv.org. Bus. Polit. 19 (2), 298–326. Available at: https://arxiv.org/abs/1510.04841/ (Accessed: 15 April 2022). Garlappi, L., Giammarino, R., Lazrak, A., 2017. Ambiguity and the Corporation: Group Tsoukalas, G., Falk, B., 2020. Token-weighted crowdsourcing. Manag. Sci. 66 (9), Disagreement and Underinvestment. J. Financ. Econ. 125 (3), 417–433. 3843–3859. Gervais, A., Karame, G., Capkun, V., Capkun, S., 2014. Is Bitcoin a Decentralized Wang, G., Ma, X., Wu, H., 2020. Are stablecoins truly diversifiers, hedges, or safe havens Currency? IEEE Secur. Priv. 12 (3), 54–60. Goldberg, M., Sch¨ar, F., 2023. Metaverse Governance: An Empirical Analysis of Voting against traditional cryptocurrencies as their name suggests? Res. Int. Bus. Financ. 54, 101225. within Decentralized Autonomous Organizations. J. Bus. Res. 160, 113764. 18 X. Sun et al. J o u r n a l o f F i n a n c i a l S t a b i l i ty73(2024)101286 Wood, G. (2014). Ethereum: A Secure Decentralised Generalised Transaction Ledger. Xu, J., Perez, D., Feng, Y., & Livshits, B. (2023) Auto.gov: Learning-based On-chain Available at: https://ethereum.github.io/yellowpaper/paper.pdf/ (Accessed: 15 Governance for Decentralized Finance (DEFI), arXiv.org. Available at: http://arxiv.org/ April 2022). abs/2302.09551 (Accessed: 07 September 2023). World Economic Forum. (2021). Decentralized Finance (DeFi) Policy-Maker Toolkit Yermack, D., 2017. Corporate Governance and Blockchains. Rev. Financ. 21 (1), 7–31. [White paper]. Available at: https://www.weforum.org/whitepapers/decentralized- Zhang, Z., Ren, D., Lan, Y., Yang, S., 2022. Price competition and blockchain adoption in finance-defi-policy-maker-toolkit/ (Accessed: 15 April 2022). retailing markets. Eur. J. Oper. Res. 300 (2), 647–660. Zhao, X., Ai, P., Lai, F., Luo, X. (Robert), Benitez, J., 2022. Task management in decentralized autonomous organization. J. Oper. Manag. 68 (6–7), 649–674. 19