betaFeedback

WorksComputer Science

Herding or Racing? Decomposing Cross-Account Order-Flow Persistence in On-Chain Markets

Abstract

Most of the autocorrelation in trade signs on decentralized exchanges comes from pairs of trades sent by different wallets. For Uniswap v3, this has been read as herding dominating order splitting – the reverse of what equity markets show. We show that this herding share mixes two quantities: the probability pp that consecutive orders come from the same account, which is small when thousands of wallets are trading, and the ratio κκ of cross-account to same-account sign correlation. For any pp, κκ equals one under the public-signal benchmark of Tóth et al. and zero under independent order splitting. On Uniswap v3 USDC-WETH (one week each in 2023 and 2024), the herding share is 0.65-0.86, but κκ is only 0.05-0.10. On a fully on-chain EVM limit order book, the herding share is similar (0.61-0.87), but κκ is 0.72-0.76; among Uniswap swaps above USD 100k, it is 0.34-0.52. Where κκ is high, cross-account co-movement is mostly a race for the price on a centralized exchange. It is concentrated in orders a few blocks apart, and a cross-fitted gradient-boosted filter that sees only Binance data removes about three quarters or more of it. This race flow accounts for most of the losses of liquidity providers. We recommend reporting κκ alongside the herding share and modeling on-chain order flow as splitting traders plus arbitrageurs who respond to the centralized-exchange price.

Introduction

Market-order signs are persistent. In equity markets, this persistence is mostly attributed to order splitting, as investors slice large metaorders into many small trades [16, 17]. The decisive evidence comes from splitting the sign autocorrelation into pairs of orders placed by the same agent and pairs placed by different agents [22]. On the London Stock Exchange, about three quarters of the lag-one correlation comes from the same member firm, and account-level data from Tokyo confirm the splitting mechanism quantitatively [21]; see [15] for a review. Tóth et al. also caution that a large cross-agent term is necessary but not sufficient for herding.

Decentralized exchanges (DEXs) allow the same decomposition at scale, since every on-chain order carries the address that sent it. Di Nosse, Gatta, Lillo and Jaimungal [10] study the 24 most active Uniswap v3 pools in 2023–2024 and find the opposite of the equity result: the cross-wallet (“herding”) component dominates the autocorrelation of swap signs. The result holds up when sandwich attacks are removed and routed swaps are aggregated. They connect it to the herding effect, “the tendency of investors to replicate the trades of others”, and note that cryptocurrency prices show “many strong price trends that can be attributed to herding effects”. They list it among the main ways decentralized finance (DeFi) departs from the stylized facts of traditional markets.

We ask what this herding share actually measures, and make three contributions.

Measurement. The herding share hh is an exact function of two quantities: the probability pp that two consecutive orders come from the same account, and the ratio κ\kappa of the sign correlation between orders of different accounts to that between orders of the same account. Under either standard benchmark, κ\kappa does not depend on pp: it is one when traders respond to a common public signal and zero when accounts split orders independently. When thousands of wallets trade, pp is only a few per cent, so hh can be large even if κ\kappa is close to zero. On the pool that Di Nosse et al. use as their main example, hh is 0.65–0.86 while κ\kappa is 0.05–0.10: pair by pair, small swaps behave like equity order flow. On a fully on-chain limit order book built on the Ethereum Virtual Machine (EVM), hh is similar, but κ≈0.75\kappa\approx0.75.

Mechanism. Where κ\kappa is high—on the order book and among Uniswap swaps above $100k—cross-account co-movement is mostly a race for the price on a centralized exchange (CEX). A cross-fitted gradient-boosted classifier that sees only Binance data removes about three quarters or more of it, and on the order book delaying its inputs by half a second makes little difference. The co-movement is concentrated in orders a few blocks apart. On the order book it is strong within a block and lasts for about ten blocks; among large Uniswap swaps it is weaker or negative within a block and concentrated in the next few blocks.

Economics. Liquidity providers pay mostly for the race. On the order book, orders in a cross-account race make up 54–74% of volume and carry 70–91% of makers’ losses. On Uniswap, after the pool fee, liquidity providers (LPs) lose 4–5 bp on the third of volume most aligned with Binance and earn about 5 bp on the least aligned third.

Related work. Beyond the order-flow literature above, our work connects to studies of maximal extractable value (MEV) and transaction ordering [9], CEX–DEX arbitrage and the losses it imposes on liquidity providers [14, 18, 19, 23], informed traders who pay for priority on DEXs [6], and latency arbitrage on stale quotes in traditional markets [3, 5, 12]. Recent identity-level studies of on-chain order books show that taker toxicity is a persistent wallet attribute [24], reconstruct address-level metaorders [4], and release full message data [2]. Client identity also predicts toxic flow in foreign exchange [7], and Binance and a few other venues lead price formation across bitcoin derivatives markets [1]. None of these decomposes sign persistence against a CEX signal. Methodologically, we borrow cross-fitting from double machine learning [8] to use a flexible learner [13, 20] as a measurement instrument.

Method

The herding ratio κ\kappa

Let ϵn∈{−1,+1}\epsilon_n \in\{-1,+1\} be the sign of the nn-th order (+1 buys the base asset), ana_n its account, and xn=ϵn−ϵˉx_n=\epsilon_n-\bar{\epsilon}. Following [10, 22], split the lag-ℓ\ell autocovariance C(ℓ)C(\ell), the mean of xnxn+ℓx_nx_{n+\ell}, into Csame(ℓ)C_{\mathrm{same}}(\ell) and Cother(ℓ)C_{\mathrm{other}}(\ell) according to whether an=an+ℓa_n=a_{n+\ell}. Let p(ℓ)p(\ell) be the share of same-account pairs, and let ρsame=Csame/p\rho_{\mathrm{same}}=C_{\mathrm{same}}/p and ρdiff=Cother/(1−p)\rho_{\mathrm{diff}}=C_{\mathrm{other}}/(1-p) be the mean products within each type of pair. The herding share is then

h(ℓ)=CotherC=(1−p)κp+(1−p)κ,κ(ℓ)=ρdiffρsame.(1)h(\ell)=\frac{C_{\mathrm{other}}}{C}=\frac{(1-p)\kappa}{p+(1-p)\kappa},\qquad\kappa(\ell)=\frac{\rho_{\mathrm{diff}}}{\rho_{\mathrm{same}}}. \tag*{(1)}

Two benchmarks anchor κ\kappa. In the public-signal model of [22], randomly selected traders respond to a common signal, every pair has the same correlation, and κ=1\kappa=1. With independently drawn account labels, p=∑iPi2p=\sum_i P_i^2, where PiP_i is account ii’s share of orders; this recovers their benchmark for the same-account share S=1−hS=1-h. If accounts split orders independently of one another, then ρdiff=0\rho_{\mathrm{diff}}=0 and κ=0\kappa=0.

The herding share inherits pp, which depends on how many accounts trade and how their orders cluster in time. Under both benchmarks, κ\kappa does not, and unlike the comparison of SS with ∑iPi2\sum_i P_i^2, it does not assume that account labels arrive independently. In data, κ\kappa can still vary with the mix of accounts and with the size threshold, and it can exceed one. Using account-specific means as in [10] instead of ϵˉ\bar{\epsilon} changes κ\kappa by at most 0.04 and hh by at most 0.02 in our samples.

A learned common-signal filter

To measure how much of the co-movement a public signal explains, we estimate s^n=E[ϵn∣Xn]\hat{s}_n=E[\epsilon_n\mid X_n] from features XnX_n observed before the order, and decompose the residual un=ϵn−s^nu_n=\epsilon_n-\hat{s}_n in the same way. The share of cross-account covariance that the filter removes is 1−ρdiffu/ρdiffx1-\rho^{u}_{\mathrm{diff}}/\rho^{x}_{\mathrm{diff}}. The learner is histogram-based gradient boosting [13, 20] with 300 iterations, learning rate 0.05, 31 leaves, an L2L_2 penalty of 1, and a minimum leaf size of 200 orders (order book) or 50 swaps (Uniswap). We cross-fit [8]: the sample is split by the parity of the clock hour, and each prediction comes from a model trained on the other half, so the filter cannot absorb in-sample noise into the residual statistics.

XnX_n contains only Binance data, read from a 100 ms grid:

  • returns over eight windows covering the 30 s before the order on the order book, or seven windows covering the 5 min before the slot on Uniswap;

  • signed trade volume over three windows (order book) or two (Uniswap);

  • realized volatility over 60 s (order book) or 5 min (Uniswap).

The latest input is the grid cell containing the block (order book) or slot (Uniswap) time, so it can include up to 100 ms of later Binance data. As a leakage control, we refit the order-book filter with all inputs delayed by 0.5 s, which leaves only data from at least 0.4 s before the block. A variant adds the gap between the venue’s price and the Binance-implied fair value, which is what arbitrageurs trade on but is not exogenous to the venue; on the order book it uses the gap at four lags and the spread. The fair value is the Binance price times a slowly varying basis: the 300 s rolling median of the venue-to-Binance midpoint ratio on the order book, and the median pool-to-Binance price ratio over the previous 400 swaps on Uniswap.

Data

On-chain order book. The venue is a central limit order book implemented as EVM smart contracts on a chain with sub-second blocks. Every order and fill is emitted as a contract event that carries the account identifier. We study its BTC perpetual contract over two samples: a main sample of 116 hours in late September and early October 2026, and an earlier sample of 72 hours in late August 2026. We use filled taker orders of at least $10. Smaller orders are 37% and 30% of orders but 0.2% and 0.6% of volume; including them gives κ=0.77\kappa=0.77 and h=0.76h=0.76 in the main sample. The main sample has 94,469 orders from 302 accounts and $49.5M of volume; the earlier sample has 142,659 orders from 457 accounts and $69.3M. Taking is concentrated in the main sample, where two accounts supply 75% of volume, but not in the earlier one (24%).

In the main sample, block times come from the first receipt of each block’s update on the venue’s public book feed, corrected for the median delivery delay; they differ from block times observed through a public node subscription by about 45 ms in median absolute terms. For the earlier sample, we interpolate the chain’s whole-second block timestamps and shift them by 0.65 s, their median offset from observed block times on days when both are available. The leader is Binance’s USD-margined BTCUSDT perpetual: best quotes on a 100 ms grid in the main sample, and last trade prices from Binance’s public archive in the earlier one (on a common day the two agree to 0.02 bp). Makers’ markouts are measured at 60 s, before trading fees and gas, against the venue’s midpoint (main sample) or the Binance price times a five-minute rolling median basis estimated from fill prices (earlier sample).

Uniswap v3. We use the USDC–WETH 0.05% pool, the main example of [10]. We extract all swap events for 3–9 June 2024 and 5–11 June 2023 from public Ethereum nodes, and take the transaction sender as the wallet, as they do. Swaps are signed +1 when they buy WETH, and we keep swaps of at least $10. That leaves 37,336 swaps from 17,982 wallets ($1.12bn) in 2024, and 31,843 swaps from 13,464 wallets ($1.93bn) in 2023. Sandwich and echo patterns detected with their algorithm are 0.35–0.62% of swaps and 1.9–6.5% of dollars; Table 1 keeps them, and we report their effect below. The leader is Binance’s ETHUSDT perpetual, aligned to the slot timestamp. LP markout is the execution price relative to the Binance-implied fair value 60 s later. The execution price recorded in the swap event includes the 5 bp pool fee, so this markout is the LPs’ result after the fee. All intervals are 95% hour-block bootstrap intervals with 2,000 resamples.

SampleNNAccountsppρsame\rho_{\mathrm{same}}ρdiff\rho_{\mathrm{diff}}κ\kappahhS/∑Pi2S/\sum P_i^2
Order book, main sample94,4693020.3310.8310.6340.76 [0.73, 0.82]0.61 [0.57, 0.64]2.2
Order book, earlier sample142,6594570.0980.9530.6890.72 [0.70, 0.74]0.87 [0.86, 0.88]4.5
Uniswap 2024, all swaps37,33617,9820.0250.8100.0370.05 [0.03, 0.06]0.65 [0.55, 0.73]199.8
swaps ≥$10\geq\$10k | 9,411 | 2,753 | 0.042 | 0.883 | 0.167 | 0.19 [0.15, 0.23] | 0.81 [0.77, 0.84] | 26.8
swaps ≥$100\geq\$100k | 3,816 | 526 | 0.104 | 0.761 | 0.257 | 0.34 [0.25, 0.43] | 0.75 [0.69, 0.79] | 9.6
Uniswap 2023, all swaps31,84313,4640.0160.7320.0720.10 [0.08, 0.12]0.86 [0.83, 0.89]63.3
swaps ≥$10\geq\$10k | 12,499 | 3,123 | 0.035 | 0.787 | 0.226 | 0.29 [0.24, 0.34] | 0.89 [0.87, 0.91] | 9.0
swaps ≥$100\geq\$100k | 5,215 | 748 | 0.059 | 0.667 | 0.346 | 0.52 [0.43, 0.62] | 0.89 [0.87, 0.91] | 4.4

Table 1. Decomposition of the lag-one autocorrelation of order signs. NN: orders of at least $10; pp: share of consecutive pairs from the same account; ρsame\rho_{\mathrm{same}}, ρdiff\rho_{\mathrm{diff}}: mean products of demeaned signs within each type of pair; κ=ρdiff/ρsame\kappa= \rho_{\mathrm{diff}}/\rho_{\mathrm{same}}; hh: herding share of C(1)C(1); S/∑Pi2S/\sum P_i^2: same-account share relative to the public-signal benchmark of [22] (about 12 on the London Stock Exchange). Brackets show 95% hour-block bootstrap intervals.

Results

Same herding share, different κ\kappa

Table 1 reports the decomposition. On Uniswap, the herding share is 0.65 in 2024 and 0.86 in 2023, confirming the finding of [10] for their example pool. Yet κ\kappa is only 0.05 and 0.10. Consecutive swaps from the same wallet agree with a correlation of 0.73–0.81, while consecutive swaps from different wallets agree with 0.04–0.07. The herding share is high because 97–98% of consecutive swaps come from different wallets. Measured against the public-signal benchmark, the same-wallet share is 63–200 times ∑iPi2\sum_i P_i^2, against about 12 times for the London Stock Exchange [22]. Removing sandwich and echo patterns lowers κ\kappa to 0.04 and 0.07 and raises the same-wallet correlation to 0.90 and 0.96. Pair by pair, small DEX flow is therefore dominated by splitting, as equity order flow is.

On the order book, the herding share is similar (0.61 and 0.87), but κ\kappa is 0.76 and 0.72. Orders from different accounts are nearly as correlated as orders from the same account, close to the public-signal benchmark. Uniswap swaps of $100k or more, which carry 76–82% of the pool’s volume, lie in between (κ=0.34\kappa= 0.34–0.52); sandwich and echo patterns change these values by at most 0.02. The size threshold matters in opposite directions. On Uniswap, κ\kappa rises with it (Table 1). On the order book it falls, to 0.73 for orders of at least $100 and 0.51 for the 22% of orders above $500.

Figure 1. Daily herding share against the herding ratio κ\kappa (lag one). Grey curves show Eq. (1) for three values of pp. The herding share confounds κ\kappa with pp. κ\kappa separates all Uniswap swaps from the order book; large Uniswap swaps lie in between.

Figure 1 shows the daily values. Each lies on the curve of Eq. (1) for its own pp, and the same herding share arises with values of κ\kappa that differ by more than an order of magnitude. On the last two days of the main sample, κ\kappa exceeds one (1.47 and 1.56), because the same-account correlation falls to 0.41 while the cross-account correlation stays at 0.61–0.65.

Where κ\kappa is high, accounts race for the CEX price

Timing. fig:2a,b shows the order book. The cross-account covariance of consecutive orders is 0.84 (main sample) and 0.80 (earlier sample) when both orders are in the same block, 0.81 and 0.75 in adjacent blocks, 0.72–0.77 two to ten blocks apart, and 0.24 and 0.22 beyond. Pairs at most ten blocks (a few seconds) apart contribute 90% and 96% of Cother(1)C_{\mathrm{other}}(1), and same-block pairs alone contribute 17% and 20%. Orders in the same block were submitted before either could observe the other’s fill, so their co-movement cannot be imitation of observed fills. Nor does the correlation rise once fills become observable.

Among large Uniswap swaps (Figure 2c,Figure 2d), the co-movement sits one to five blocks apart (0.42–0.56), and these pairs contribute 94% (2024) and 76% (2023) of Cother(1)C_{\mathrm{other}}(1). Within a block the covariance is negative in 2024 (−0.32) and weaker in 2023 (0.24). One interpretation is market design. In an automated market maker (AMM), every swap moves the pool price, so the first arbitrage swap of a block consumes much of the gap to the CEX price, and the race is decided block by block. On the order book, by contrast, several makers’ quotes can be stale at the same time and several accounts take them within one block. There, the first and second same-direction orders in a block capture 2.11 and 2.01 bp relative to the CEX-implied fair value. Testing this interpretation directly is left for future work.

Figure 2. Cross-account lag-one covariance of order signs by the number of blocks between the two orders, before (dark) and after (light) removing the prediction of the Binance-only filter. On the order book, the co-movement is strong within a block and up to ten blocks apart. Among large Uniswap swaps, it sits in the next few blocks, and within a block its sign is unstable.

A learned filter removes most of the co-movement. Table 2 reports the filter results. Using only Binance data, the filter explains 69% of the variance of order-book signs in the main sample. It removes 84% of the cross-account covariance: 90% within blocks and 95% across adjacent blocks. Delaying its inputs by 0.5 s barely changes this (68% and 83%), and adding the venue–Binance gap raises the two figures to 73% and 87%. The earlier, less concentrated sample gives 54% and 73%. A residual cross-account correlation of 0.33 (main) and 0.40 (earlier) remains, against 0.63 and 0.69 before filtering, so other common signals, such as other leading venues or the venue’s own state, may also matter.

Table 2: Learned common-signal filter. R2R^2: out-of-fold share of sign variance explained. Removed: share of the cross-account lag-one covariance removed by the filter, over all pairs, within the same block, and across adjacent blocks. Values above 100 mean the residual covariance is negative; “–” marks a negative raw covariance. Uniswap filters are fitted on swaps of at least $10k.

SampleFeaturesR2R^2 (%)AllSameNext
Order book, mainBinance69849095
delayed 0.5 sBinance68838494
with gapBinance, gap73879296
Order book, earlierBinance54737378
Uniswap 2024, ≥$100\ge\$100k | Binance | 60 | 109 | – | 112
with gapBinance, gap6689–96
Uniswap 2023, ≥$100\ge\$100k | Binance | 48 | 91 | 82 | 91
with gapBinance, gap61957798

Table 2.

For Uniswap swaps above $100k, the Binance-only filter explains 48–60% of signs and removes 91–109% of the cross-wallet covariance; with the pool–Binance gap, the figures are 61–66% and 89–95%. The residual cross-wallet correlation lies between −0.06 and 0.08.

In these samples, then, most of the cross-account co-movement where κ\kappa is high is a common response to the CEX price.

Race flow and adverse selection

Table 3 reports the costs to liquidity providers. Sorting orders into dollar terciles of their alignment αn=ϵns^n\alpha_n = \epsilon_n\hat{s}_n with the filter’s prediction, makers’ markouts on the order book fall from −0.51 to −2.31 bp in the main sample and from +0.08 to −2.48 bp in the earlier one. Orders that belong to a cross-account race are 54% and 74% of volume. They cost makers 1.87 and 1.70 bp, and account for 70% and 91% of makers’ losses; the remaining orders cost 0.94 and 0.46 bp.

Table 3: Adverse selection by flow type. Markouts at 60 s (bp, dollar-weighted) of the liquidity side by dollar tercile of alignment αn≡ϵns^n\alpha_n \equiv\epsilon_n\hat{s}_n with the Binance-only filter (T1 least aligned), and for orders in a cross-account race (another account’s same-direction order within three blocks on the order book, two on Uniswap) versus the rest. Order book: before trading fees and gas. Uniswap: swaps of at least $10k; LP markouts include the 5 bp fee.

SampleT1T2T3% of $ | Race | Rest
Book, main (makers)−0.51−1.51−2.3154−1.87−0.94
Book, earlier (makers)+0.08−1.73−2.4874−1.70−0.46
Uniswap 2024 (LPs)+4.57−2.56−4.2366−2.38+2.43
Uniswap 2023 (LPs)+5.29−2.33−5.0878−2.44+5.57

Table 3.

On Uniswap, LP markouts after the 5 bp fee fall from +4.6 to −4.2 bp (2024) and from +5.3 to −5.1 bp (2023) across alignment terciles. Before the fee, LPs roughly break even on the least aligned third (−0.4 and +0.3 bp) and lose 9–10 bp on the most aligned third. Swaps in a cross-wallet race (66–78% of volume) lose LPs about 2.4 bp even after the fee, while the rest earn 2.4–5.6 bp. The high-κ\kappa race component of order flow is thus the adverse-selection component, the source of what the AMM literature calls loss-versus-rebalancing [19].

Discussion

Reading DeFi stylized facts. The departure that [10] report is real, but it describes the population of accounts as much as their behavior. Pair by pair, small swaps on Uniswap are dominated by splitting, as in equity markets [21, 22]. What is new on-chain is a race among takers for the CEX price, in which orders from different accounts are nearly as correlated as orders from one account. Whenever venues or periods with different numbers of accounts are compared, we suggest reporting κ\kappa (or SS relative to ∑ipi2\sum_i p_i^2) next to the herding share.

Modeling. Di Nosse et al. motivate order-level simulators of DEX dynamics [10]. Such simulators, and agent-based models of liquidity provision against latency-sensitive arbitrageurs [11], should combine traders who split their orders with arbitrageurs who respond to an exogenous CEX signal. They should also reproduce the venue-specific timing of the race.

Machine learning. We use the learned filter as a measurement device. It isolates the predictable common component of order flow before the decomposition is applied, and cross-fitting keeps its own fit out of the residual statistics. Its output is also a pre-trade toxicity score whose terciles order liquidity providers’ losses monotonically. The broader point concerns markets populated by algorithmic agents: agents that react to the same public signal generate herding statistics without imitating one another, so herding measures should be conditioned on the signals those agents observe.

Limitations. We study one order-book venue and one contract over 188 hours, and one Uniswap pool over two weeks. An account is not a trader: one entity can spread its flow over several addresses, and routers or shared senders can merge many users. On the order book, merging accounts that share a sending address, which overmerges the users of shared gateways, raises κ\kappa from 0.76 to 0.85; merging the two largest takers gives 0.73. Order-book markouts exclude trading fees and gas, and the delayed-input check covers the main sample only. The filter sees only Binance; other leading venues would likely explain more of the co-movement.

Conclusion

The herding share of order-flow persistence confounds the number of accounts with how strongly they co-move. Separating the two shows that small DEX flow is dominated by splitting pair by pair, and that much of the genuine cross-account co-movement in on-chain markets is a race for the centralized-exchange price. A learned filter removes most of that co-movement, and liquidity providers pay mostly for it.

References

References

  1. [1]Jakob Albers, Mihai Cucuringu, Sam Howison, and Alexander Y. Shestopaloff. 2021. Fragmentation, Price Formation and Cross-Impact in Bitcoin Markets. Applied Mathematical Finance 28, 5 (2021), 395–448. doi:10.1080/1350486X.2022.2080083DOI
  2. [2]Jakob Albers, Mihai Cucuringu, Sam Howison, and Alexander Y. Shestopaloff. 2026. An Open Book: Level 4 Order Book Data from the Hyperliquid Exchange. SSRN working paper 6465720. doi:10.2139/ssrn.6465720DOI
  3. [3]Matteo Aquilina, Eric Budish, and Peter O’Neill. 2022. Quantifying the High-Frequency Trading “Arms Race”. The Quarterly Journal of Economics 137, 1 (2022), 493–564. doi:10.1093/qje/qjab032
  4. [4]Davide Barone and Fabrizio Lillo. 2026. Trading in the Sunshine or in the Shade: Market Impact and Adverse Selection on Hyperliquid. arXiv:2606.15715. https://arxiv.org/abs/2606.15715
  5. [5]Eric Budish, Peter Cramton, and John Shim. 2015. The High-Frequency Trading Arms Race: Frequent Batch Auctions as a Market Design Response. The Quarterly Journal of Economics 130, 4 (2015), 1547–1621. doi:10.1093/qje/qjv027
  6. [6]Agostino Capponi, Ruizhe Jia, and Shih-ao Yu. 2026. Price Discovery on Decentralized Exchanges. The Review of Financial Studies (2026). doi:10.1093/rfs/hhag002
  7. [7]Álvaro Cartea, Gerardo Duran-Martin, and Leandro Sánchez-Betancourt. 2026. Detecting toxic flow. Quantitative Finance 26, 4 (2026), 541–561. doi:10.1080/14697688.2026.2619539DOI
  8. [8]Victor Chernozhukov, Denis Chetverikov, Mert Demirer, Esther Duflo, Christian Hansen, Whitney Newey, and James Robins. 2018. Double/debiased machine learning for treatment and structural parameters. The Econometrics Journal 21, 1 (2018), C1–C68. doi:10.1111/ectj.12097DOI
  9. [9]Philip Daian, Steven Goldfeder, Tyler Kell, Yunqi Li, Xueyuan Zhao, Iddo Bentov, Lorenz Breidenbach, and Ari Juels. 2020. Flash Boys 2.0: Frontrunning in Decentralized Exchanges, Miner Extractable Value, and Consensus Instability. In 2020 IEEE Symposium on Security and Privacy (SP), 910–927. doi:10.1109/SP40000.2020.00040
  10. [10]Daniele Maria Di Nosse, Federico Gatta, Fabrizio Lillo, and Sebastian Jaimgungal. 2026. Deviations from tradition: stylized facts in the era of DeFi. Quantitative Finance 26, 8 (2026), 1231–1252. doi:10.1080/14697688.2026.2678231
  11. [11]Daniele Maria Di Nosse and Fabrizio Lillo. 2026. Mitigating Adverse Selection in Concentrated Liquidity AMMs with Dynamic Fees: An Agent-Based Model Approach. arXiv:2606.23070. https://arxiv.org/abs/2606.23070
  12. [12]Thierry Foucault, Roman Kozhan, and Wing Wah Tham. 2017. Toxic Arbitrage. The Review of Financial Studies 30, 4 (2017), 1053–1094. doi:10.1093/rfs/hhw103
  13. [13]Jerome H. Friedman. 2001. Greedy function approximation: A gradient boosting machine. The Annals of Statistics 29, 5 (2001), 1189–1232. doi:10.1214/aos/1013203451DOI
  14. [14]Lioba Heimbach, Vabuk Pahari, and Eric Schertenleib. 2024. Non-Atomic Arbitrage in Decentralized Finance. In 2024 IEEE Symposium on Security and Privacy (SP), 3866–3884. doi:10.1109/SP54263.2024.000256arxiv.org/abs/2401.01622
  15. [15]Fabrizio Lillo. 2023. Order Flow and Price Formation. In Machine Learning and Data Sciences for Financial Markets. Cambridge University Press, 107–129. doi:10.1017/9781009028943.008
  16. [16]Fabrizio Lillo and J. Doyne Farmer. 2004. The Long Memory of the Efficient Market. Studies in Nonlinear Dynamics & Econometrics 8, 3 (2004). doi:10.2202/1558-3708.1226DOI
  17. [17]Fabrizio Lillo, Szabolcs Mike, and J. Doyne Farmer. 2005. Theory for long memory in supply and demand. Physical Review E 71, 6 (2005), 066122. doi:10.1103/PhysRevE.71.066122DOI
  18. [18]Jason Milionis, Ciamac C. Moallemi, and Tim Roughgarden. 2025. Automated Market Making and Arbitrage Profits in the Presence of Fees. In Financial Cryptography and Data Security: 28th International Conference, FC 2024, Revised Selected Papers, Part I, Jeremy Clark and Elaine Shi (Eds.). Springer, 159–171. doi:10.1007/978-3-031-78676-1_9DOI
  19. [19]Jason Milionis, Ciamac C. Moallemi, Tim Roughgarden, and Anthony Lee Zhang. 2022. Automated Market Making and Loss-Versus-Rebalancing. arXiv:2208.06046. https://arxiv.org/abs/2208.06046
  20. [20]Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, Jake Vanderplas, Alexandre Passos, David Cournapeau, Matthieu Brucher, Mathieu Perrot, and Édouard Duchesnay. 2011. Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research 12, 85 (2011), 2825–2830. https://www.jmlr.org/papers/v12/pedregosa11a.html
  21. [21]Yuki Sato and Kiyoshi Kanazawa. 2023. Inferring Microscopic Financial Information from the Long Memory in Market-Order Flow: A Quantitative Test of the Lillo-Mike-Farmer Model. Physical Review Letters 131, 19 (2023), 197401. doi:10.1103/PhysRevLett.131.197401DOI
  22. [22]Bence Tóth, Imon Palit, Fabrizio Lillo, and J. Doyne Farmer. 2015. Why is equity order flow so persistent? Journal of Economic Dynamics and Control 51 (2015), 218–239. doi:10.1016/j.jedc.2014.10.007DOI
  23. [23]Fei Wu, Danning Sui, Thomas Thierry, and Malles h Pai. 2025. Measuring CEX-DEX Extracted Value and Searcher Profitability: The Darkest of the MEV Dark Forest. In 7th Conference on Advances in Financial Technologies (AFT 2025) (Leibniz International Proceedings in Informatics (LIPIcs), Vol. 354), Zeta Avarikioti and Nicolas Christin (Eds.). Schloss Dagstuhl – Leibniz-Zentrum für Informatik, 26:1–26:23. doi:10.4230/LIPIcs.AFT.2025.26
  24. [24]Daojing Zhai. 2026. Public Trader Identity: Adverse Selection and Return Pre-dictability. arXiv:2608.04373. https://arxiv.org/abs/2608.04373

Paper details

vlnc:W_dxntp5yw
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents