Non-Atomic Arbitrage in Decentralized Finance

Since the transition to PoS, Ethereum runs on two layers. The execution layer, largely taken over from the former PoW protocol, is tasked with validating and executing transactions. Conversely, the consensus layer, situated atop the beacon chain, is dedicated to reaching consensus among validators, i.e., the PoS equivalent of miners. To become a validator, one must lock, i.e., stake, 32 ETH in a specified smart contract.

Time is divided into slots, periods of twelve-second length, on the Beacon chain. In every slot, a single validator is responsible for proposing a block, i.e., being a proposer. Thus, there is a chance for a block to be proposed and for the blockchain to be extended in every slot. For building a block, a validator receives a fixed consensus layer block reward ($\approx 0.04$ ETH). Additionally, the validator also receives a variable execution layer reward for building the block. This variable reward stems from fees being paid for transactions to be included in the block. These fees can be relayed to the proposer either directly through transaction fees or through direct payment (i.e., coinbase transfers). On average, these execution layer block rewards are around 0.12 ETH but vary significantly as we find.

PBS is a novel scheme that was designed for Ethereum PoS [5] and through which the majority (around 90% [3]) of blocks are proposed. With PBS the previous roles of the block proposer are separated into two tasks, namely block building and block proposing. In short, specialized block builders are tasked with building high-value blocks. Amongst these blocks built by the builders, the validator is expected to choose the block with the highest bid, i.e., the amount the proposer would receive, and proposes the block to the network.

Figure 1 depicts the PBS scheme and we detail the roles of each of the players in the following:

Searchers scan the mempool and look for profitable (bundles of) transactions, generally these transactions involve MEV opportunities, such as sandwich attacks, arbitrage, or liquidations. They pass their bundles to builder(s) privately to avoid them being copied by predatory trading bots scanning the mempool.

Builders combine bundles received from searchers they are connected to, with private and public transactions to build high-value blocks. Once a block is built they forward the block to one or many relays along with how much they are willing to give the proposer to choose their block – the bid. In case their block is later chosen by the proposer, the builder’s profit is the difference between the block’s value and the bid.

Relays are connected to builders. When they receive blocks from builders, they check that the block is valid, that the payment to the proposer is as indicated in the bid, and that the block complies with any censoring filters implemented by the relay. They identify the block with the highest bid. Then, they send the block header, along with the size of the payment to the proposer.

Proposers request blocks from all relays they are connected to. They are expected to choose the block with the highest bid if the value of this bid exceeds the value of the block they built themselves. Once they have chosen a block, they sign the block header received from the relay which will prompt the relay to broadcast the block.

Note that in blocks built through PBS, the block builder is indicated as the fee recipient and the builder transfers the promised amount to the proposer in the block’s last transaction. We further highlight that one entity can play all the roles at once. For instance, if a builder also runs a searcher we would refer to this as an integrated seacher.

DEXes are one of the most popular DeFi applications. They allow users to exchange cryptocurrencies on the blockchain without giving up custody of their assets. While traditional CEXes generally utilize a limit order book to match orders, most DEXes are constant function market makers (CFMM). With a CFMM, users execute their order against a liquidity pool holding reserves of two or more cryptocurrencies. Trades against the pool execute according to a pre-defined trading function specified in the pool’s smart contract, i.e., the pool only accepts trades that keep its trading function constant. Importantly, every trade that executes in a CFMM pool moves the pool’s price deterministically, dependent only on the pool’s reserves ahead of the trade and the trade’s input size. Consequently, if an arbitrage opportunity arises, the assets locked in the liquidity pool determine how large the arbitrage trade can be. The larger the liquidity pools, the larger the required arbitrage trades are to synchronize the prices across different trading venues.

MEV refers to the profit that can be extracted through including, excluding, and reordering transactions in a block [6]. The most commonly observed and measured types of MEV on Ethereum are sandwich attacks, cyclic arbitrage, and liquidations. Succinctly, a sandwich attack has the attacker front- and back-running the victim’s transaction on a DEX for a profit. A cyclic arbitrage profits from price differences across DEXes. Liquiditations close under-collateralized loans on lending protocols. The liquidator can buy the collateral of a position at a discount, as a reward for repaying the debt. Oftentimes the liquidator then sells the collateral on a DEX[7].

Decentralized Exchange Trades. We collect data for all DEX trades on Uniswap V2, Uniswap V3, Curve, Balancer and Sushiswap. These five biggest DEXes in terms of total value locked (TVL) on the Ethereum blockchain [8]. In total, our data set comprises 77,019,583 swaps on DEXes. For each trade, we identify the pool in which the trade was executed as well as the amount of tokens exchanged. We further, record whether the trade was executed successfully, i.e., whether the execution of the transaction was successful.

Maximal Extractable Value. To differentiate non-atomic arbitrage trades from types of MEV and to compare their respective prevalence on DEXes, we collect MEV transaction labels from zeromev [9]. These labels specify whether a transaction was part of a sandwich attack (i.e., victim and attacker transaction), performs a cyclic arbitrage, or a liquidation.

Rewards Data. We collect Ethereum execution layer reward data. In particular, for each block we calculate the rewards received by the block’s fee recipient: (1) priority gas fees, and (2) tips, i.e., direct transfers to the fee recipient address. Note that the execution layer rewards are also referred to as the block value. We, further, record information regarding the PBS scheme rewards. In the PBS scheme specification, the block builder is indicated as the block’s fee recipient and in the block’s last transaction transfers the agreed-upon value to the block proposer. Thus, for blocks that follow the PBS specification we record the amount transferred by the fee recipient to the block proposer.

We collect data from eleven relays: Aestus, Agnostic, Blocknative, bloXroute (Ethical), bloXroute (Max Profit), bloXroute (Regulated), Eden, Flashbots, Manifold, Relayoor and UltraSound. The relays implement public APIs that give access to the blocks delivered by the relays to the proposers, as well as the bids received by the builders. Note that some relay APIs, namely Blocknative, bloXroute (Ethical), Eden, and Relayoor, were not reachable. For these, our data only extends to 31 April 2023, i.e., the last time we scraped the data. Further, bloXroute (Ethical), bloXroute (Max Profit), and bloXroute (Regulated) removed access for builder bids from their APIs. When querying the builder bids API for blocks that passed through the respective bloXroute relay, the response is consistently empty. Thus, for the three bloXroute relays the bid data only extends to 31 January 2023, i.e., the last time we scraped that data and received non-empty responses.

Importantly, the data by the relays is self-reported and thus requires some trust in the relays, as it cannot be fully verified. However, we combine our Ethereum blockchain reward data set with our PBS data set to verify that the block values reported by the relays correspond to those received by the validators. If this is not the case, we disregard any relay bid data for that block.

To obtain block times, i.e., the time at which a block was first seen in the network, and to identify private transactions we use Ethereum network data from the Mempool Guru project [10]. The Mempool Guru project runs geographically distributed Ethereum nodes. Each node records the timestamp at which it first saw any block or transaction. The projects data collection method is detailed in [11]. We take the earliest timestamp across the nodes operated by the Mempool Guru project as the block time and consider a transaction private if not one node saw the transaction before the block it is included in was seen in the network.

We obtain historical cryptocurrency price data from Binance.com [12] and CoinMarketCap [13]. In particular, we obtain candlestick data from Binance.com with an interval of a second as well as aggregate trade data for ETH and BTC. We further obtain daily candlestick data from CoinMarketCap for 60 cryptocurrencies.

In this section we develop a model for the profit of a non-atomic arbitrageur between two currencies $X$ and $Y$. As discussed in Section 2.3, most DEXes are a CFMM with the largest DEX being Uniswap [8]. Uniswap currently has two active versions, V2 [14] and V3 [15]. Here, we focus on the V2 version, which is also employed by the DEX Sushiswap [16], but the model also holds for V3 between two ticks. The Uniswap CFMM imposes, $x\cdot y=L^{2}$, where $x$ and $y$ stand for the number of $X$- and $Y$-tokens locked in the liquidity pool, and $L$ is referred to as the liquidity. The (marginal) price $P$ for $X$-tokens in terms of $Y$-tokens is given as $P=y/x$. The reserves can be expressed in terms of the liquidity and price by solving for $x$ and $y$, yielding $x=L/\sqrt{P}$, $y=L\sqrt{P}$.

The price on a DEX on Ethereum is $P_{\mathrm{on}}$ and the average off-chain price is $\tilde{P}_{\mathrm{off}}$, i.e., the average price received for an order.¹¹1The average price $\tilde{P}_{\mathrm{off}}$ can contain transactions on multiple off-chain exchanges. In general, it is the best price attainable for a transaction of a given size. We denote the price difference between the off-chain and on-chain as $\Delta P=\tilde{P}_{\mathrm{off}}-P_{\mathrm{on}}$. Without loss of generality, we assume the searcher discovers $\Delta P>0$, i.e., the price of $X$-tokens off-chain is higher such that the arbitrageur buys $X$-tokens on the DEX. During the trade, they pay $\Delta y$ of which $f\cdot\Delta y$ is paid in fees to the liquidity providers, where $f$ is the protocol fee. With this and the constant product $L^{2}=xy$, the amount of $X$-tokens the trader receives is given by

Apart from $f$, the arbitrageur is charged a fee by the off-chain exchanges which we denote as $g$.

As long as the off-chain price remains higher than the on-chain price, the arbitrageur can extract more value by buying more $\Delta x$ on the DEX. Thus, the arbitrageur will trade on the DEX until the price they receive for an infinitesimal unit of $X$ is $\tilde{P}_{\mathrm{off}}(1-g)$.²²2Here we assume that the off-chain price is unaffected by the arbitrage. In practice, the arbitrageur would receive a quote for a given trade size and can be used as $\tilde{P}_{\mathrm{off}}$. The pool’s marginal price at the end of the trade is then given by

Simultaneously, to this buy order, the arbitrageur sells the same amount of $X$-tokens, $\Delta x$ through the off-chain exchanges at an average price of $\tilde{P}_{\mathrm{off}}$.

Thus, the arbitrageur receives

where we used the previous two equations. Finally, we can compute the arbitrageurs profit by solving $\Delta P$ for $\Delta y$ and inserting it in the above equation for $\Delta y_{\mathrm{off}}$ which leads to

where we also expressed $x$ and $y$ in terms of $L$ and the starting on-chain price $P_{\mathrm{on}}$. The profit as a function of the price change $\Delta P$ is plotted in Figure 3. The profit becomes significant for larger price changes, showing that sufficient price difference will attract non-atomic arbitrage trades. Note that profitable arbitrage only exists for $P_{\mathrm{on}}<\tilde{P}_{\mathrm{off}}(1-f)(1-g)$, as otherwise the fees paid to the liquidity pool exceed the price difference that can be profited from.

We further note that while an arbitrageur still needs to pay the builder to have their trade included, the result shows how valuable the arbitrage trade is for the searcher and limits the bid they would be willing to submit.

To provide a better understanding of non-atomic arbitrage, we go through a case study of block 18,360,789 – a block with a significant price change in the lead-up to the block, i.e., the time between the previous block proposal and the block proposal itself. In Figure 4, we plot the time and the value of bids from the builders. Note that we hightlight bids from rsyncbuilder (shown in red) and beaverbuild (shown in blue), as these two builders were previously identified as having integrated searchers that perform these non-atomic arbitrage trades [4]. The bids from all remaining builders are shown in yellow. We further indicate the time of the bid corresponding to the block that was chosen by the proposer, i.e., the relays deliver the highest bid to the proposer when the proposer requests it. Finally, we also plot the relative price change of ETH-USDT and BTC-USDT on Binance.com during the same time. Recall that USDT is a stablecoin pegged to the $.

One immediately notices that the time in the lead-up to the block is characterized by extremely high price volatility of both ETH and BTC in comparison to USDT. The relative price change exceeds 0.6% and 1.3% in less than ten seconds respectively. This price change creates an arbitrage opportunity, as the prices on DEXes do not move in between blocks. Thus, as outlined in the previous section, the first transaction in a DEX pool can close this price difference and thereby profit from the arbitrage opportunity.

Coming back to Figure 4, we can observe that around five seconds after the price starts to change on Binance.com the bids start to increase. At this point, the price difference appears to be big enough for non-atomic arbitrage to be profitable. Further, we find that bids from the builders that are associated with non-atomic arbitrage transactions are higher than those from the rest and that the bids continue to increase as the ETH and BTC prices on Binance.com increase. The block chosen by the proposer was built by rsyncbuilder and its bid, i.e., the value received by the proposer was a staggering 10.32 ETH ($\approx$$23,000 at the time of this writing). In Table I, we take an in-depth look a the non-atomic arbitrage trades we identified with our heuristics that we will introduce in the following Section 5. In the block, there were 26 transactions with volume exceeding $6 million that we identified as performing non-atomic arbitrage trades. Remarkably, all but three of these transactions were by the rsyncsearcher3 – a searcher we identified to be linked to the rsyncbuilder that won the block (cf. Section 6.3). Additionally, we highlight that the rsyncsearchers paid a remarkable 11.15 ETH (i.e., more than the PBS bid) for their transactions. Thus, the rsyncbuilder built a high-value block largely attributed to fees paid by trades we labeled as non-atomic arbitrage by its own integrated searcher.

Note that most of trades buy ETH or BTC for a stablecoin. As we saw in Figure 4 the prices of these two cryptocurrencies rose on Binance.com, and they were thus trading for cheaper on DEXes at the beginning of the block. For example, if we look at the price at the beginning of the block, we see that the price of the ETH-USDT Uniswap V3 pool was 1563.57, and at the end of the block, it was 1574.15. The Binance.com price was 1564.61 at the beginning of the block’s slot, while it rose to 1574.63 by the time the block was proposed. Notice that the arbitrageurs drove the price almost exactly to the off-chain price, as we show to be optimal in our previous analysis (cf. Section 4.1). Additionally, some transactions exchange cryptocurrencies for cryptocurrencies (i.e., no stablecoins), such as the transaction at index 2 and 6, which exchange ETH for BTC. The relative price increase measured in USDT of BTC was higher than that of ETH. Thus, BTC relative to ETH was also available for cheaper on DEXes at the beginning of the block. Hence, the searcher also swaps ETH for BTC.

We start by providing a general overview of the non-atomic arbitrage on DEXes. On average, we identify 5216 trades as non-atomic arbitrages per day during our data collection window that spans a little over a year. This figure is in comparison to 5219 sandwich attacks (front- and back-running transaction counted as one), 1245 cyclic arbitrage (i.e., atomic on-chain arbitrage), and 14 liquidations. Recall, that sandwich attacks, cyclic arbitrage, and liquidations are considered the most common and well-studied types of MEV. Thus, non-atomic arbitrage, in terms of the number of occurrences is the second most prevalent type of MEV out of these four on the Ethereum blockchain. We provide a more detailed comparison between non-atomic arbitrage and these other types of MEV in Appendix C.

In Figure 5, we consider the cumulative volume of non-atomic arbitrage over time and highlight the eleven biggest searchers performing non-atomic arbitrage by volume. We provide a mapping from searcher name to address in our data set [17]. Further note that some searchers (beaversearcher1, beaverseacher2, rsyncseacher1, rsyncseacher2, rsyncseacher3, and mantaseacher) have names that suggest that they are linked to large builders in the PBS scheme (beaverbuild, rsyncbuilder and mantabuilder) [21]. This naming is due to our suspicion that the respective builders and searchers are controlled by the same entities and we will detail our reasoning in Section 6.3.

Returning to Figure 5, in total, the volume of trades we identify as non-atomic arbitrage are responsible for $132 billion on DEXes, e.g., Uniswap V2, Uniswap V3, Sushiswap, Curve, and Balancer. Startlingly, the total volume on these five DEXes in the same period of time is $460 billion [22]. Thus, the identified non-atomic arbitrage accounts for nearly 30% of the total volume of the five biggest DEXes on the Ethereum blockchain.

We reiterate here that there is no clear-cut way of identifying non-atomic arbitrage, as we lack data transparency regarding the off-chain. However, the fact that merely eleven searchers are responsible for 80% of the non-atomic arbitrage volume identified in our analysis and that we identify the majority of their transaction as non-atomic arbitrage trades for all but one searcher (cf. Appendix D), leads us to believe that our heuristics are well-suited to identify non-atomic arbitrage transactions. Additionally, these eleven searchers can further be clustered into eight entities, three searchers are associated with rsyncbuilder and two searchers are associated with beaverbuild (cf. Section 6.3 for a detailed description of the searcher/builder relationship identification). Thus, a very small number of entities is responsible for the vast majority of the non-atomic arbitrage volume on DEXes. In fact, the two biggest searchers, i.e., beaversearcher1 and jumpsearcher, account for 49.2% non-atomic arbitrage volume. We further point out that many of these large searchers (beaversearchers, jumpsearcher, rsyncsearchers, and mantasearcher) are rumored to be linked to HFT firms operating in traditional finance [4, 23], a further indicator that our heuristics capture non-atomic arbitrage as well as a sign that big players from traditional finance turning to DeFi for profits.

Finally, we highlight that two periods of time exhibit significant increases in the cumulative volume: (1) mid-November 2022, which is a period of time characterized by high volatility of cryptocurrencies during the FTX bankruptcy [24], and mid-March 2023, which corresponds to the USDC depeg and high cryptocurrency price volatility [25]. High cryptocurrency price volatility — leading to price differences between cryptocurrency markets — drives non-atomic arbitrage opportunities and these dramatic increases in non-atomic arbitrage volume during periods of high volatility are thus expected. We provide a more in-depth analysis of the correlation between price volatility and non-atomic arbitrage volume in Section 6.2.

In regards to Figure 5, we conclude that the non-atomic arbitrage volume on DEXes is immense and controlled by a few large entities. To gain a better understanding of the evolution over time, we take a more in-depth look at the share of non-atomic arbitrage volume controlled by the large searchers in Figure 6. We start by noting that on nearly 75% of days, merely two searchers are responsible for at least 50% of the arbitrage volume – underlying the high concentration in the non-atomic arbitrage market. Interestingly, beaverseacher1, the biggest non-atomic arbitrage searcher, is the only major searcher operating through our entire data collection period. Generally, beaversearcher1 accounts for at least 10% of the volume except for during the two time periods characterized by very high cryptocurrency price volatility in the previous, i.e., mid-November 2022 and mid-March 2023. Potentially, the non-atomic arbitrage market becomes more competitive during days characterized by exceptional events in the blockchain ecosystem, and beaverseacher1 loses some market share as a result. The second biggest searcher, jumpsearcher, on the other hand, was operating since the merge but stopped in late July 2023, while searcher1, the third biggest searcher, only started operations in mid-March 2023 and from then on was operating throughout our entire data collection window. Interestingly, the three searchers associated with rsyncbuilder all operate during non-overlapping time windows, it appears that one searcher is replacing the other – potentially upgrades to the searcher smart contract. Furthermore, there was a period of time between May and September 2023, when no searcher we could associate with rsyncbuilder was operating.

To get a better understanding of the differences in strategies utilized by the various searchers, we plot the distribution of trade sizes for each searcher as a violin plot in Figure 7. A violin plot includes a boxplot, which indicates the lower quantile, median, and upper quantile, as well as a kernel density plot to visualize the distribution of values. We observe clear differences in the trade size distribution between searchers (cf. Figure 7). For instance, beaversearcher1’s trade size distribution is quite smooth with the peak and median being around $10,000 and trade sizes ranging from less than $100 to more than $10 million. Interestingly, except beaverseacher2, all other major searchers have larger median trade sizes for their non-atomic arbitrage swaps. We further note that while their median trade sizes are slightly higher than that of beaverbuild, mantaseacher and builder1searcher exhibit similar trade size distributions to beaversearcher1. A couple of searchers (e.g., jumpsearcher, searcher1, searcher2, and searcher 3) appear to be specialized in high-volume non-atomic arbitrage trades. The median trade sizes of the former three all exceed $100,000. Finally, we observe that the three rsyncbuilders tend to have several peaks in their trade size distributions, i..e, they seem to be biased towards several distinct trade sizes. Further, we observe that with each iteration of the rsyncsearcher, the median trade size declines. While we cannot be sure what the cause of this is, it could be a sign that with time rsyncsearcher is turning towards smaller and likely less fought-after (as they are less valuable) non-atomic arbitrage transactions.

In Table III, we further analyze what proportion of non-atomic arbitrage volume and swaps by each of the largest eleven searchers are in pools between ETH, BTC, USDC, USDT, and DAI. These are the biggest cryptocurrencies on DEXes in terms of the number of swaps. One would assume that a significant proportion of volume is in such pools, given that these cryptocurrencies also have liquid markets on both DEXes and CEXes. We find that for all searchers but searcher3, who has exclusively had non-atomic arbitrage trades between these five tokens, the proportion of trades between these five tokens is smaller than the respective proportion of volume (cf. Table III). This discrepancy indicates that higher volume non-atomic arbitrage trades are executed between these five tokens, which is expected given that comparatively the DEX pools for these five tokens have higher liquidity. Therefore, it requires greater trade sizes to move the price as outlined in Section 4.1. Except for searcher3, we observe largely similar behavior in terms of the proportion of volume/trades executed between these five highly liquid tokens or otherwise for all searchers. Interestingly, we notice again that with each iteration of the rsyncsearchers, the proportion of trades/volume executed between the top five tokens drops. Thus, it again appears as if with time the rsyncbuilder is turning to potentially less fought-after non-atomic arbitrage transactions.

In the following, we investigate the drivers of non-atomic arbitrage trades on DEXes. Recall, that non-atomic arbitrage leverages price differences between two markets (likely an on-chain DEX and an off-chain CEX), and these are expected to be heightened in times of high cryptocurrency price volatility. In more detail, price volatility on CEXes can drive non-atomic arbitrage transactions, as it signals a divergence from the price at an on-chain DEX. Thus, we commence by investigating the relationship between the daily volume of trades we identified as non-atomic arbitrage trades and the daily price volatility of BTC and ETH in Figure 8. Note that we measure the BTC and ETH price volatility since the change in cryptocurrency prices is closely correlated to changes in the two [26]. We measure the daily volatility by $\log_{10}\left({P_{\text{high}}}/{P_{\text{low}}}\right),$ where $P_{\text{high}}$ and $P_{\text{low}}$ are the daily high and low of the price respectively.

Looking at Figure 8, we observe that days with a high volume of non-atomic arbitrage trades tend to correspond to days with high BTC and ETH price volatility. Additionally, the highest volume days correspond to two periods identified earlier: the FTX collapse in November 2022 and the USDC depeg in March 2023. We further note that the correlation between the daily ETH price volatility and the volume of non-atomic arbitrage is 0.725 with a p-value of $1.30\cdot 10^{-69}$ – inferring extremely high statistical significance of the correlation. For the BTC price volatility, the correlation is even higher at 0.729 with a p-value of $2.86\cdot 10^{-68}$.

Note that the strong correlation between price volatility and non-atomic arbitrage volume does not necessarily imply causation. To further explore relationship, we calculate the correlation between volume (excluding non-atomic arbitrage) and non-atomic arbitrage volume. The correlation is 0.655, and thus, lower than the correlation between non-atomic arbitrage volume and price volatility of 0.725 and 0.729 for ETH and BTC respectively. The stronger correlation between non-atomic arbitrage volume and price volatility, underlines the relationship between the two. Additionally, we note that the correlation is lower despite our heuristics being conservative and likely undercounting the prevalence of non-atomic arbitrage, i.e., some remaining volume is probably non-atomic arbitrage and thereby correlated. Thus, price volatility of ETH and BTC on CEXes likely drives non-atomic arbitrage on DEXes.

In Figure 9, we visualize the volume of non-atomic arbitrage trades by weekday and hour using UTC as the timezone. Observe that the highest volume in non-atomic arbitrage trades is seen Monday through Friday between 14:00 UTC and 21:00 UTC, which corresponds to US market openings. Especially US market opening at 14:30 UTC is associated with increased non-atomic arbitrage volume. An additional, less noticeable cluster of non-atomic arbitrage volume is observed Monday through Friday around 0:00 UTC, which corresponds to the opening of the Asian markets, and even less noticeable Monday through Friday at 8 am UTC – European market opening. This heightened volume during market hours and especially during the respective market openings is expected, as cryptocurrency prices are correlated with the US markets [27]. Thus, volatility in cryptocurrency prices is expected to increase during (US) market trading hours. Hence, we see a rise in the volume of non-atomic arbitrage trades during these times, as cryptocurrency prices are most volatile. We observe a final and interesting peak in non-atomic arbitrage volume on Wednesdays around 18:00 UTC. This corresponds to the time of the press conference following the Federal Open Market Committee Meeting at 19:00 UTC and the preceding release of notes at 18:30 UTC [28]. In this meeting decisions regarding the federal interest rate are made and they have led to significant market and cryptocurrency price movement [29].

To conclude, we identified the expected link between cryptocurrency price volatility and non-atomic arbitrage.

As mentioned previously, the searchers looking for and executing non-atomic arbitrage opportunities and the builders in the PBS scheme building the block need not be separate entities. In fact, previous work [4] has demonstrated that three builders (i.e., beaverbuild, rsyncbuilder, and mantabuilder) are likely to operate their own non-atomic arbitrage searchers but did not establish links between these builders and specific searchers. These builders are referred to as HFT builders. We move to link non-atomic arbitrage searchers and builders in this section.

To start, we analyze what proportion of the non-atomic arbitrage volume by each of the biggest eleven searchers is included in blocks built by each of the biggest nine builders (cf. Figure 10). Note that we identify the biggest searchers and builders by the non-atomic arbitrage volume they traded and included in their blocks respectively. We provide a mapping from searcher/builder name to their address/public key in our data set [17].

The darker the tone of blue in Figure 10, the larger the proportion of non-atomic arbitrage volume by the searcher in blocks by the corresponding builder. We, for instance, observe that more than half (i.e., 79% and 82% respectively) of the non-atomic arbitrage volume from beaverseacher1 and beaverseacher2 is in blocks by beaverbuild. Thus, it is likely that these two searchers are associated with beaverbuild and we name them accordingly. Additionally, we find evidence that beaverbuild was subsiding its integrated searchers for a month as we detail in Appendix A.2. This further shows that they are very likely to be the same entity.

Similarly, 83% of the non-atomic arbitrage volume of mantasearcher is found in blocks built by the mantabuilder, while 88% of volume from the rsyncsearcher2 and an astonishing 98% of volume from the rsyncsearcher3 is included in blocks by the rsyncbuilder. Note that there is no link between rsyncsearcher1 and rsyncbuilder in Figure 10 even though our naming suggests there to be a link. However, we are able to establish a link between rsyncsearcher1 and rsyncsearcher2, as the bytecode of their contracts is identical and they are thus likely to be the same entity as established in previous work [20]. We further comment that rsynbuilder was largely not operating at the same time as rsyncsearcher1 (cf. Figure 13(b)). Finally, we find a fourth, previously not identified, builder that is likely running an integrated non-atomic arbitrage builder: builder1. To be precise, 83% of the non-atomic arbitrage volume by builder1seacher is located in blocks built by builder1seacher.

Figure 10 allows us to observe ties between builders and searchers. However, at the same time, we also find that non-integrated searchers (e.g, jumpsearcher, searcher1, searcher2, and seacher3) have a harder time having their non-atomic arbitrage trades included in blocks by the four builders that have their own integrated builders, i.e., their non-atomic arbitrage trades are generally in blocks by flashbots, builder0x69, etc. Thus, it seems that either other searchers are less likely to submit their transactions to these four builders with integrated searchers or the four builders favor their integrated searchers.

We further take an in-depth look at the connections between HFT builders and their integrated searchers in Appendix A.1. Our analysis finds that for all builders the correlation between the share of daily blocks for which they won the PBS auction and the daily share of arbitrage volume by the respective integrated searcher is positive and the correlation is significant. Additionally, this correlation is slightly lower for beaverbuild and rsyncbuilder. This difference possibly stems from the fact that these builders also include significantly more sandwich attacks and cyclic arbitrage (i.e., common MEV types) than builder1 and mantabuilder. Thus, the value of their blocks is likely not solely driven by non-atomic arbitrage transactions.

To conclude our analysis, we discuss the impact of non-atomic arbitrage and the vertical integration of searchers specializing in this arbitrage. We start by noting that from 2023 onwards generally more than 50% of non-atomic arbitrage volume is found in HFT builder blocks (cf. Figure 11). Often this share rises to around 75%. Recall, that there are just four HFT builders, at no time were more than three of them in operation, and from May 2023 onwards only two were still operating (cf. Figure 14). Thus, we conclude that these builders have a strong grasp on the non-atomic arbitrage market.

As we established in the previous section, cryptocurrency price volatility can drive non-atomic arbitrage transactions, as it signals a divergence from the price at an on-chain DEX. Thus, we analyze the relationship between ETH and BTC price volatility and the share of blocks won by HFT builders (cf. Figure 12). We notice both an increase in the daily share of HFT builder blocks over time but also a correlation between the share of blocks won by HFT builders and the price volatility. To be precise, in October 2023 the correlation between the share of blocks built by HFT builders and the Ethereum price volatility is 0.783 with a p-value of $1.85\cdot 10^{-7}$, whereas it is 0.650 for the BTC price volatility with a p-value of $7.51\cdot 10^{-5}$. Thus, we establish that on high volatility days, these HFT builders specialized in non-atomic arbitrage transactions are more likely to win blocks. This finding is in line with a previous analysis by Gupta et al. [4] that showed that between April and May 2023 beaverbuild, mantabuilder, and rsyncbuilder were more likely to win the PBS auction as the price volatility of the ETH-USD market on Binance.us increases.

To further investigate the effects of cryptocurrency price volatility and thereby non-atomic arbitrage on the system, we take a closer look at the volatility in the lead-up to a block, i.e., for block $n$ the price volatility between the time block $n-1$ and block $n$ were first seen in the network. On average (if blocks are not missed) this is twelve seconds on Ethereum. Increased volatility during this time should lead to increasingly profitable and prevalent non-atomic arbitrage transactions.

Figure 13 plots the cdf of gas usage and block value of non-atomic arbitrage depending on the ETH and BTC price volatility. Observe that the higher the BTC and ETH price volatility in the lead-up to the block, the larger the share of the block’s gas usage by non-atomic arbitrage trades (cf. Figure 12(a)). Overall, in less than 1% of blocks, non-atomic arbitrage transactions make up more than 5% of the block’s total gas usage (red line). However, for blocks above the 99.9th percentile of BTC price volatility (yellow line), we see that in around 40% of blocks, non-atomic arbitrage transactions use more than 10% of block space, i.e., gas. Note that this is in addition to already larger blocks during times of high volatility (cf. Appendix B). Thus, the increased prevalence of non-atomic arbitrage likely leads to increased block competition during times of high volatility. Importantly, this also leads to an increased base fee (i.e., the minimum fee required for block inclusion) for consecutive blocks as the base fee automatically increases based on the previous block size. Further, this increase in the price for transaction inclusion as well as the overfullness of the block likely leads to increased congestion in the mempool. Thus, the block space used by non-atomic arbitrage during times of high volatility impacts all Ethereum users.

We also analyze the share of block value (i.e., priority fees and direct transfers) that come from non-atomic arbitrage transactions in Figure 12(b). Again we find that the share of value from non-atomic arbitrage transactions increases with the volatility. In fact, in around 50$\%$ of blocks where the ETH and BTC volatility is in the 99.9${}^{\text{th}}$ percentile, over 80$\%$ of the block value comes from non-atomic arbitrage transactions. Note again that not only the share of value from non-atomic arbitrage but also the block value increases in times of high volatility (cf. Appendix B). We further outline that the increased value from non-atomic arbitrage transactions in times of high volatility explains the higher likelihood that HFT builders to win blocks during times of high price volatility in ETH and BTC.