PaperlayerView exact revision on arXiv

Source-linked AI summary

Priority Gas Auction Cadence and Searcher Competition: Evidence from Flashblocks on Base

Pranay Anchuri, Edward W. Felten, Akaki Mamageishvili

arXiv:2610.03922v1cs.CEcs.GTq-fin.TR

TL;DR

The paper asks how dividing Base’s priority gas auction into shorter sequential auctions changes searcher bidding and competition. Using on-chain data and a difference-in-differences design, it finds lower searcher fees and less top-of-block fee concentration, while identifying latency-related ordering and limits to causal interpretation of reverts.

  • Problem

    The effect of faster priority gas auctions on searcher bids and winners had not been measured, and a before-and-after comparison cannot separate the intervention from concurrent market shifts.

    Source: §1 · p. 2, §4.1 · p. 6

  • Method

    The paper uses an address-day difference-in-differences design with activity-matched controls and reconstructs flashblock structure from on-chain priority-fee sequences.

    Source: §1 · p. 2, §4.3 · p. 8

  • Results

    Searcher effective priority fees fall by 0.187 gwei relative to a matched control, and top-of-block fee concentration drops from 0.98 to 0.28.

    Source: §7 · p. 15

  • Takeaways & Limitations

    Auction cadence is a governance-tunable design parameter, and the estimates inform how rollups set it.

    Source: §1 · p. 2

    The revert-rate response is not cleanly identified because the searcher classification conditions on revert rate and its pre-period differential trend is nonzero.

    Source: §5.2 · p. 10

Abstract

from arXiv · show

Flashblocks divide a block's priority gas auction into shorter sequential auctions that commit transaction order before the block is complete. We ask how this auction cadence affects bidding and competition among automated arbitrageurs, or searchers. In July 2025 Base replaced a single 2 s auction with ten 200 ms auctions. We use this change to estimate the searcher response from on-chain data alone. On an address-day panel of 3,032 searchers and 8,053 activity-matched controls, a difference-in-differences design estimates a 0.187 gwei fall in the effective priority fee (59% of the searcher pre-period mean). The share of searcher priority-fee value paid in the first tenth of block gas falls from 0.98 to 0.28. This fee compression is the identified effect. The revert-rate response is not (causally) identified and is reported descriptively. At the level of a single opportunity, the winning fee splits into an approximately invariant floor and a competitive premium. The premium scales about linearly with auction duration at high contention and less steeply at low contention. A first-price auction model with uncertain arrival and partial payments by losing searchers accounts for this pattern. We match attempts to the pool they contest. A higher-fee transaction reverts behind a lower-fee winner that landed earlier in 1.9 to 2.7% of contested opportunities by value, an upper bound. Latency therefore decides part of the ordering. A micro-auction model projects diminishing returns from shortening the window below 200 ms, as bids approach a nonzero floor. Rollups such as Arbitrum One now make auction cadence a governance tunable parameter, so these estimates inform how it is set.

1 Introduction

Base’s July 2025 shift from one 2 s priority gas auction to ten sequential 200 ms sub-auctions creates a setting to measure how auction cadence affects searcher bidding and competition. Using on-chain data, the paper estimates a large searcher-fee reduction, examines latency’s role in ordering, and distinguishes the identified fee effect from descriptively reported reverts.

1 Introduction

  • The July 2025 intervention replaces a single 2 s auction with ten sequential 200 ms sub-auctions while leaving the canonical block time at 2 s.Flashblocks commit transaction order as each preconfirmation sub-block is produced.Source: §1 · pp. 1–2
  • Searchers are the primary auction bidders and pay most priority fees, making their bidding costs relevant to ordinary users and auction-cadence design.The introduction notes that Arbitrum One made auction cadence a governance-tunable parameter.Source: §1 · p. 2
  • A difference-in-differences design compares active searchers with activity-matched, non-competitive addresses using an address-day panel and on-chain data alone.The analysis reconstructs sub-block structure from within-block priority-fee sequences because flashblocks have no on-chain marker.Source: §1 · p. 2
  • Searcher effective priority fees fall by 0.187 gwei (59% of their pre-period level) relative to a matched control, while top-of-block fee concentration drops from 0.98 to 0.28.The revert-rate response is not cleanly identified and is reported descriptively.Source: §1 · p. 3
  • At the opportunity level, a higher-bidding transaction loses because of late arrival in 1.9–2.7% of contested opportunities by value, an upper bound.The winning fee comprises an approximately invariant floor and a contestable premium; the premium scales about linearly with auction duration at high contention and less steeply at low contention.Source: §1 · p. 3

2 Background

The background describes Base’s flashblock intervention and the on-chain data used to study it. Aggregate monthly statistics show fee and revert-rate shifts, but those pooled changes do not isolate the effect of flashblocks.

2 Background

  • Base introduced ten 200 ms preconfirmation sub-blocks within each 2 s block in July 2025, and the canonical block time remained 2 s.June is the pre period, August the post period, and July is excluded as the transition period.Source: §2 · p. 3
  • Robinhood chain uses 100 ms blocks but applies first-come, first-served transaction ordering.Source: §2 · p. 3
  • Mean effective priority fee per gas falls from 0.038 to 0.025 gwei between June and August 2025, while the revert rate rises from 0.078 to 0.114.These pooled shifts mix searchers with other addresses and confound flashblocks with the July-to-August market shift.Source: Table 1 · p. 4, §2 · p. 4
  • Transaction- and block-level data for June and August 2025 come from a Base archive node, with system and deposit transactions excluded from per-transaction rates.The later analysis windows are processed identically.Source: §2 · p. 4

3 Model and predictions

The models link auction duration to opportunity value and searcher bidding. They predict a duration-sensitive contestable premium, while partial loser payments and uncertain arrival shape equilibrium bids and auction revenue.

3 Model and predictions

  • At high contention, the winning fee’s premium over an approximately invariant floor scales with block time, with measured exponents from about 0.5 to 1.2 across competitor counts.The floor includes a jump component that survives as the window approaches zero, while the diffusive component declines with window length.Source: §3 · pp. 4–5
  • In the first-price opportunity auction, the winner pays its full bid and each loser pays a fraction of its bid.This partial-payment format lies between winner-pays and all-pay first-price auctions; players can arrive in the current or next round, and late arrivals lose regardless of bid.Source: §3 · p. 5
  • With uncertain arrival, longer windows admit larger price dislocations and more timely arrivals, and the model predicts lower revenue as block time falls.The auction’s revenue increases with the current-round arrival probability when that probability exceeds the threshold stated in Proposition 2.Source: §3 · p. 5
  • The equilibrium point-wise bid need not increase with current-round arrival probability because that probability enters both the numerator and denominator of the bidding function.The paper notes examples in which bidding does not increase as arrival probability increases.Source: §3 · p. 5

4 Methods

The methods define transaction outcomes and identify the searcher response with an address-day difference-in-differences design. They classify a competitive core, reconstruct flashblock structure from fee-order resets, and measure fee concentration in the first gas decile.

4.1 Outcomes and identification

  • The design compares searchers with comparable non-competitive addresses in an address-day panel, using a two-way fixed-effects interaction to estimate the treatment effect.Address effects absorb fixed group differences and day effects absorb common shocks; identification requires the groups’ untreated changes to move together.Source: §4.1 · p. 6

4.2 Groups and classification

  • The control group is matched on activity, not fees, and its much lower fees make equal-percentage trends difficult to check.Searchers and controls are selected from addresses active in both months; requiring both competitive markers isolates a core at the cost of coverage.Source: §4.1 · p. 7, §4.2 · p. 7

4.3 Flashblock reconstruction

  • Flashblock reconstruction treats a cross-sender upward fee jump as an ordering reset, which can mark either a flashblock boundary or a live preemption.Same-sender jumps are discarded; because resets can also arise mid-window, the slice count approximates rather than bounds the number of flashblock boundaries.Source: §4.3 · p. 8

4.4 Fee-concentration index

  • The fee-concentration index measures the share of within-block priority-fee value paid in the first gas decile.A uniform per-gas rate gives the index its neutral value of 0.1, while concentrating the highest rates at the block top drives it toward 1.Source: §4.4 · pp. 8–9

5 Results

Flashblocks fragment block ordering and compress searcher fees. Opportunity-level evidence links the winning-fee premium to auction duration and contention, while model projections suggest smaller gains from shortening windows below 200 ms.

5.1 Block structure

  • Blocks average 19.5 inferred slices post-launch versus one pre-launch, with fee concentration C falling from 0.89 to 0.27; later structure persists, but does not identify long-run fee effects.The inferred slice count measures ordering fragmentation, not an exact sub-block count.Source: §5.1 · p. 9

5.2 Causal estimates

  • Searcher effective priority fee falls 0.187 gwei (59% of the 0.317-gwei pre-period mean), and declared bid falls 0.228 gwei more than controls.The fee event-study gap is flat before treatment (p = 0.79) and steps down across the post-period.Source: §5.2 · pp. 9–10
  • Searcher fee concentration C falls from 0.98 to 0.28, versus 0.07 to 0.09 for controls; the June pre-trend is flat but its test has low power.C measures the share of each group's within-block priority-fee value in the block's top gas decile.Source: §5.2 · p. 10
  • Revert-rate, overbid, and transaction-count estimates are descriptive, not causal, because they have nonzero pre-trends; revert-rate screening also selects on that outcome.The revert-rate pre-trend has p < 0.001.Source: §5.2 · p. 10, Table 3 · p. 11

5.3 Opportunity-level competition and late arrival

  • A higher-fee attempt reverts one reconstructed slice after its linked winner in 1.9–2.7% of contested captures by winner fee value, an upper bound on latency losses.Reverts can also reflect slippage or stale opportunities, and pool attribution can misassign multi-pool searchers.Source: §5.3 · p. 12

5.4 Testing the predictions

  • The uncontested floor is similar (0.177 gwei in June; 0.164 in August), while the premium's block-time exponent rises from 0.51 at two searchers to 1.12 beyond five.The premium partly reflects opportunity size as well as competition; the fee is a lower bound on arbitrage value, not a direct measurement.Source: §5.4 · pp. 12–13
  • Reverting losers account for 11.4% of contested-capture fee value, consistent with the first-price model's partial loser payments, but the evidence is qualitative.The median loser pays 0.15 times the winner's fee value; the equilibrium bid is not fitted to the observed winning fee.Source: §5.4 · p. 13

5.5 Shorter flashblock windows

  • Under the measured-competition model, winning fees fall under 10% from 0.181 to 0.169 gwei between 200 and 50 ms; the simpler model projects a 33–68% drop.Neither model can be checked below 200 ms, and their projections differ in whether β(c) stays the same at shorter windows.Source: §5.5 · pp. 13–14, §5.4 · p. 13

6 Related work

The related literature spans transaction-fee design, MEV and priority gas auctions, preconfirmations and faster blocks, and block-time arbitrage. This study focuses on priority-fee responses to changing auction cadence and connects arbitrage value to searcher fees.

Transaction-fee mechanisms

  • Earlier transaction-fee work focuses on the base fee, which flashblocks leave unchanged; this paper instead studies how priority fees respond to auction-cadence changes.Roughgarden [15] analyzes EIP-1559 incentives, while Leonardos et al. study base-fee dynamics and Reijsbergen et al. [14] measure block-size fluctuations.Source: §6 · p. 14

MEV and priority gas auctions

  • Prior MEV and priority-gas-auction studies examine ordering value at whole-block scale, while this work measures competition across ten sequential sub-auctions.Related studies document arbitrage-bot auctions and consensus risks, quantify extractable value [13], and measure speculative spam MEV on Base and Arbitrum [17].Source: §6 · p. 14

Preconfirmations and faster blocks

  • Prior work studies timing games, auctioned time advantages, and batching as a way to remove latency races; flashblocks instead create ten sequential races within a block.The paper measures the net effect on searcher bids using on-chain data.Source: §6 · p. 14

Block-time arbitrage

  • Block-time arbitrage research formalizes and measures how arbitrage depends on block time; this paper uses a diffusion-plus-jump decomposition to relate arbitrage value to searcher fees.The cited work includes loss-versus-rebalancing and its extensions [11] [10] [12] [4].Source: §6 · pp. 14–15

7 Conclusion

On Base, ten sequential 200 ms flashblock auctions compress searcher bidding relative to matched controls, while opportunity-level ordering also reflects latency. The paper projects diminishing returns below 200 ms, but that projection is a model extrapolation and the revert-rate response is not identified.

Conclusion

  • 0.187 gwei: flashblocks lower the searcher effective priority fee relative to a matched control, 59% of its pre-period level, and reduce top-of-block fee concentration from 0.98 to 0.28.The causal estimate is a level shift at the July intervention; the block-level regime persists eleven months after launch, but the causal effect is not re-estimated for later months.Source: §7 · p. 15
  • 1.9 to 2.7% of contested opportunities by value: higher-bidding transactions lose to a one-flashblock deficit, an upper bound attributed partly to latency.The unweighted upper bound is 5.4 to 6.6%.Source: §7 · p. 15
  • The measured fee compression is first-order at 200 ms and diminishes below it as bids approach a nonzero floor.The shorter-window projection comes from models, not an observed sub-200 ms regime.Source: §7 · p. 15
  • The revert-rate response is not identified under the study’s screen and is reported descriptively, not causally.The sub-200 ms projection is also a model extrapolation rather than a measurement.Source: §7 · p. 15

Contents of the Appendices

The appendices develop auction proofs and block-time arbitrage scaling, describe extracted data, and examine live queue sequencing and post-launch comparisons.

A Proofs

  • The proofs derive bidders’ allocation probabilities, equilibrium bids, and interim payments from uncertain arrival and winning probabilities.The proof passages also set up expected auction revenue.Source: A Proofs · pp. 16–17
  • Expected revenue increases with arrival probability for p > n/(n+2), while an example shows equilibrium bids need not remain monotone for sufficiently large p.In the example, the derivative’s sign is the sign of (1−c)v^2+6cv−3c, which is negative for sufficiently small v when c > 0.Source: A Proofs · pp. 18–19

B Block-time dependence of arbitrage value

  • The diffusive arbitrage-value term scales linearly in block time when fees are small relative to per-block moves, but as B^3/2 when fees dominate.Common arbitraged pools are described as near the linear case because price moves exceed fees on volatile pools, though this is not checked pool by pool.Source: B Block-time dependence of arbitrage… · p. 20
  • The diffusion term vanishes as B approaches zero, while discontinuous price moves contribute an invariant jump floor to arbitrage value.The paper takes the jump component from a jump-diffusion decomposition.Source: B Block-time dependence of arbitrage… · p. 20

C Extracted fields

  • The data pull extracts transaction records in block-and-index order, stores fee and value fields in wei, and uses a companion block-level pull for base fees.Table 4 presents the transaction fields in stored order.Source: C Extracted fields · p. 20, Table 4 · p. 21

D Live priority-queue pull

  • Post-flashblock fee sequences show 6.1 live-pull returns per block, absent pre-flashblocks; the figure contrasts representative blocks from each regime.The return pattern is indirect evidence of mid-window preemption because sequencer receive-times are unavailable.Source: D Live priority-queue pull · p. 21, Figure 4 · p. 22

E Regime stability of flashblocks

  • The regime-stability section compares outcomes at different times after flashblock launch.The supplied passage states the comparison’s scope but gives no results.Source: E Regime stability of flashblocks · p. 22

E.1 Comparison against a mature post-flashblock month.

The paper compares the August 2025 post-launch regime with October 2025 and June 2026 to assess whether its block structure persists. It reports persistent fragmentation and low fee concentration, alongside a decline in live-pull behavior.

E.1 Comparison against a mature post-flashblock month

  • Mean inferred segments fall from 19.5 to 12.0 to 10.0, while returns fall from 33.2% to 19.8% to 0.4%, making live pulls predominantly an early post-launch feature.By June 2026, 91.8% of blocks carry exactly ten segments and 3.3% exceed ten.Source: §E.1 · p. 23
  • Strict-core searcher fees are descriptively stable across post months, but cross-period levels remain confounded by changes in ETH price, congestion, base fee, and automated activity.The reported effective priority fees are 0.112, 0.110, and 0.123 gwei in August, October, and June 2026; the within-period difference-in-differences comparison identifies the causal fall.Source: §E.1 · p. 23
  • Realized volatility ranges from 66% to 76% annualized across windows, and the authors report that a volatility gap is not driving the estimate.June and August had comparable volatility; day effects absorb the component common to both groups.Source: §E.1 · p. 23
  • Cross-sender upward-fee jumps remain 94.4%, 92.8%, and 93.6%, while top gas-decile share C is 0.27, 0.27, and 0.36, below June 2025’s 0.89.The single 2 s auction is absent in all three post months, and the qualitative regime is unchanged eleven months on.Source: §E.1 · p. 22, Table 6 · p. 24

F Estimation details and robustness

Robustness checks preserve the direction of the fee and concentration results, while revert-rate estimates remain descriptive. Fixed-effects weighting and population definitions explain why reported estimates and levels can differ across analyses.

F Estimation details and robustness

  • Fee and concentration estimates keep their signs under both activity-matched and retail controls, but fee declines are larger against the activity-matched control.Revert-rate estimates remain descriptive because their pre-trends are nonzero.Source: F Estimation details and robustness · p. 25, Table 7 · p. 25
  • The DiD fee and bid estimates are −0.187 and −0.228 gwei, rather than raw differences of −0.168 and −0.183, because fixed-effects weighting reweights address-day cells.The transaction-count estimate is +55.1 rather than the raw +155.7.Source: F Estimation details and robustness · p. 25
  • Searcher fee and concentration levels differ across the causal core, in-period searchers, both-months aggregates, and all-sender populations, so they are not directly comparable.For example, the core searcher fee is 0.317 gwei and its pre-period concentration index is 0.98; the all-sender index is 0.89.Source: F Estimation details and robustness · p. 25

G Opportunity-level analysis

Opportunity-level analysis groups competitors by the liquidity pool they contest, then examines fee competition and late-arriving attempts. Its evidence includes both a mechanical pre-period check and secondary latency diagnostics.

G Opportunity-level analysis

  • Competitors are grouped by underlying liquidity pool because contract-based keys can split genuine competitors or pool unrelated transactions together.Reverted transactions lack Swap events, so their pools are inferred from successful transactions, covering 96% of core reverted fee value in August.Source: G Opportunity-level analysis · pp. 25–26
  • Among August contested captures, higher-fee attempts revert within two reconstructed slices after their linked winner in 9.8–11.9% of cases.The range reflects alternative losing-transaction pool-attribution rules; those outbidding losers bid a median 0.02–0.03 gwei above the winner.Source: G Opportunity-level analysis · p. 26
  • June’s late-arrival rate is 0.0% by construction because each block has one fee-sorted slice, making this a mechanical check rather than independent validation of matching.Winners pay on full execution, while reverting losers pay only for gas consumed before reverting.Source: G Opportunity-level analysis · p. 26
  • Tables 8 and 9 report the competitor-count distribution and winning-fee means for opportunities defined as pools within a 2 s block.The number of distinct core searchers contesting each opportunity is the competition measure.Source: Table 8 · p. 26, G Opportunity-level analysis · p. 26

H Transaction-level hypotheses

The transaction-level tests find concentrated preemption premiums and more retry activity, while timing and concentration tests qualify other proposed mechanisms. The volatility analysis highlights limits to causal interpretation, and the shorter-window projections suggest diminishing returns below 200 ms.

H Transaction-level hypotheses

  • Queue-reset transactions pay 9.9× interior transactions’ effective fee (0.139 versus 0.014 gwei) and overbid at twice the interior rate.The block-level fee distribution does not become bimodal: its mean bimodality coefficient is 0.89 in both regimes.Source: H Transaction-level hypotheses · pp. 26–27
  • Reverts per block rise from 17.6 to 25.1 (+43%), while reverts per inferred contest fall from 17.6 to 1.3 (−93%), consistent with more retries across less-contested sub-auctions.Inter-reset gaps show no 200 ms synchronization, and within-searcher concentration falls from 0.93 to 0.26, matching the pooled decline from 0.89 to 0.27.Source: H Transaction-level hypotheses · p. 27

I Market volatility as a confounder

  • Revert-rate effects are not causally identified because their pre-period gap is unstable, while comparable June and August volatility and a July event-study step limit—but do not eliminate—the volatility concern.The near-zero control fee makes differential priority-fee responses difficult to check.Source: H Transaction-level hypotheses · p. 27, Figure 5 · p. 28, I Market volatility as a confounder · p. 28
  • October 2025 volatility reaches 76% annualized, driven by a cascade day at 217%, while searcher effective fees track realized volatility more closely than revert rates do.On the cascade, the effective fee is 4.7 standard deviations above its monthly mean while revert rate and transaction count barely move.Source: I Market volatility as a confounder · pp. 27–28

J Extrapolating bids below the 200 ms window

  • The sub-200 ms analysis calibrates a micro-auction model from observed 2 s and 200 ms bids, with Poisson arrivals across sub-windows and a competition premium that falls as windows multiply.The model assumes cadence-invariant competition intensity and value distribution, and independent one-shot sub-auctions.Source: J Extrapolating bids below the 200 ms… · pp. 29–30
  • At 50 ms, the projected declared-bid premium ranges from 0.128 gwei under the power shape to 0.061 gwei under the hazard shape, a 33–68% reduction from 200 ms.The shapes diverge below 200 ms; the measured mean competition intensity is 1.56, versus 6–92 required by the Poisson-thinning fits.Source: J Extrapolating bids below the 200 ms… · p. 30
  • Sub-200 ms figures are model extrapolations, not identified estimates: the first-order bid compression from flashblocks is floored at 200 ms, with further window shortening yielding diminishing returns.Source: J Extrapolating bids below the 200 ms… · p. 30

K Extrapolation from measured competition

Measured winning fees rise with the number of competing searchers, even after controlling for opportunity value. A model using measured competition and value primitives predicts that shorter windows below 200 ms yield little further fee reduction, subject to a transportability assumption.

K Extrapolation from measured competition

  • Value-controlled and raw winning-fee estimates rise from about 0.16 to 0.95–1.03 gwei as competitor count increases, so measured gas-used composition does not drive the slope.Gas used is a coarse proxy for opportunity value.Source: K Extrapolation from measured… · p. 31
  • The projection transports the same-searcher response across window lengths, but observed bids differ between 2 s and 200 ms, making the sub-200 ms result illustrative rather than exact.At 200 ms, expected rivals per window are only 0.14, so most windows are uncontested and bids sit near the value-capture floor.Source: K Extrapolation from measured… · p. 32
  • The model predicts a winning effective fee of 0.181 gwei at 200 ms and 0.169 gwei at 50 ms, approaching a nonzero 0.164-gwei floor.The decline is under 10%, and raw and value-controlled paths agree within 0.01 gwei.Source: K Extrapolation from measured… · p. 32, Table 12 · p. 33
Loading 2610.03922v1…