# Just-in-Time Resale in an Ahead-of-Time Auction for Faster Execution

Crypto.world

> Blockchains need a rule for ordering transactions, and many designs sell that ordering right through an auction. Arbitrum's Timeboost is unusual because it sells a 200 ms ordering advantage ahead of…

**arXiv**: 2603.20175

**Published**: 2026-03-20

**Authors**: Burak Öz, Akaki Mamageishvili, Christoph Schlegel, Ali Taslimi

## AI summary

## Overview

**Research area:** Algorithmic game theory and market design applied empirically to blockchain transaction ordering — specifically, ahead-of-time (AOT) auctions, secondary/ resale markets, and Maximal Extractable Value (MEV) in Layer-2 rollups.

**Technical level:** Intermediate. The paper combines a light theoretical framework (interdependent-values auctions, AMM arbitrage theory) with applied econometrics (markouts, correlation analysis), but the core arguments are accessible to anyone familiar with basic auction concepts.

**Scope (one sentence):** An empirical study of Arbitrum's Timeboost auction (August 2025 – March 2026) examining how well ahead-of-time bids predict realized CEX–DEX arbitrage profits, and how the rise of a just-in-time resale intermediary (Kairos) reshaped bidding, auction revenue, and surplus allocation.

## What This Paper Is About

Blockchains need a rule for ordering transactions, and many designs sell that ordering right through an auction. Arbitrum's Timeboost is unusual because it sells a 200 ms ordering advantage **ahead of time** — the winner must bid before knowing what opportunities will actually materialize during the one-minute fast-lane interval. The paper asks two questions: can bidders accurately price that forward-looking right, and what happens to the market when a resale intermediary starts buying fast-lane access wholesale and reselling it just-in-time at the transaction level? Using on-chain auction bids, time-boosted transactions, and payment traces, the authors document a shift from two searchers competing directly in the primary auction to both of them sourcing access through a common intermediary, with primary-auction revenue collapsing while total searcher profits stay broadly flat.

## Key Contributions

1. **First large-scale empirical evaluation of a deployed AOT ordering auction.** The authors assemble a dataset of 298,530 Timeboost auction rounds, 30,310,245 time-boosted transactions, and 39,521 Kairos payments, and measure how auction bids relate to realized arbitrage profits estimated via CEX markouts.
2. **Evidence that AOT bids are noisy short-horizon predictors but informative long-horizon ones.** Round-level bid–PnL correlations during the competition phase are statistically significant yet economically modest (e.g., Pearson 0.321 for Wintermute's top bid), but rise sharply when profits are aggregated over windows from 5 minutes to 1 hour (Spearman up to +0.804).
3. **Documentation of a discrete regime change in market structure.** The paper identifies three on-chain events (February 12, February 18, February 25, 2026) that divide the transition window into a Competition, Kairos, Reserve Price Adaptation, and Steady State period, and quantifies how win shares, bid gaps, and volatility sensitivity changed across them.
4. **A framework distinguishing three mechanisms for resale — unbundling, weakened competition, and information — with distinct empirical signatures**, and a finding that the observed data most strongly support the weakened-competition mechanism.

## Main Findings

- **Bids predict direction better than magnitude.** During the competition phase, per-round Spearman correlations between bids and positive-PnL markouts are higher than Pearson correlations for both searchers (Wintermute: Spearman 0.449 top bid / 0.458 paid bid versus Pearson 0.321 / 0.348; Selini: 0.337 / 0.305 versus 0.115 / 0.109), suggesting bidders could identify favorable intervals but not forecast their size.
- **Predictability improves with aggregation horizon.** Over a sample of n = 201,737 round-level observations, Spearman correlations between mean bids and total PnL rise from +0.681 (5 min) to +0.804 (1 hour) for the top bid; Pearson correlations rise from +0.195 to +0.401 over the same windows.
- **Shorter windows do not help.** Restricting markouts to the first 15 or 30 seconds of each round yields correlations that remain positive but weaker than the full-minute results (e.g., Wintermute Pearson +0.240 at 15 s and +0.282 at 30 s for the top bid, n = 116,679).
- **Bids are persistent, per-round profits are not.** Auction bids exhibit strong and slowly decaying autocorrelation, while per-round markout PnL shows little autocorrelation beyond the first two lags. The authors read this as bidders tracking persistent market conditions while facing substantial uncertainty about any individual one-minute interval.
- **Primary-auction revenue collapses after the Kairos transition.** In the focused transition-analysis window, paid bids fall from 62.7% of the top bid in the February 1–12 pre-transition segment to 14.8% thereafter, while estimated arbitrage PnL remain broadly similar across regimes.
- **Kairos becomes the dominant winner.** In the coordination phase (47,689 auctions), Kairos wins 54.5%, Selini 23.6%, and Wintermute 21.6%, together 99.7% of wins. In the Steady State period, Kairos wins 78.4% of rounds (17,119 of 21,837).
- **Bidding participation looks competitive, but pricing does not.** There were 117,146 bids across 14 distinct bidding addresses, averaging 2.46 bids per round, and the share of rounds with all three entities bidding is highest after the transition (85.3% in Steady State). Yet the median daily relative bid gap rises from 37.3% in the Competition period to 57.5% in the Kairos period.
- **Kairos bids a largely fixed amount.** During the Kairos period it wins 99.6% of rounds and predominantly submits a fixed bid of 0.004 ETH, with the paid price typically only marginally above the reserve. The top bid shows no statistically significant relationship with ETH volatility in this period.
- **A single unattributed address set prices for most rounds.** Address `0x3141592601bf441700f150bef82db6040f4fe373` submitted a fixed bid of 0.003970 ETH (0.75% below Kairos's 0.004 ETH), later 0.002921 ETH, won only 13 rounds (about 1% of those it participated in), but set the paid bid in 1,238 rounds — 96% of its active rounds. The authors find no public attribution for it.
- **Mechanism verdict.** Weakened competition fits best: the same searchers remain central, total value extraction changes only modestly, and primary auction revenue declines. Evidence for unbundling or information-based efficiency gains is described as more limited.

## Methodology in Plain English

The authors reconstruct the market entirely from public on-chain data, primarily via Dune Analytics queries that they publish.

**Auction side.** They pull every Timeboost round with at least one bid from Arbitrum's public endpoints, recording the top bid, the price actually paid (Timeboost is a second-price auction), and the one-minute interval of fast-lane access. They then assign each time-boosted transaction to an auction round by block time.

**Identifying arbitrage.** They use a heuristic: a transaction counts as CEX–DEX arbitrage if it targets Wintermute's or Selini's known contracts, emits only one swap-related event, and trades a known liquid pair among WETH, WBTC, USDC, or USDT. This yields 11,320,741 Wintermute and 3,399,555 Selini arbitrage transactions out of their time-boosted activity.

**Estimating profit.** Profit is estimated with "markouts": for a trade buying x units of token A for y units of token B on a DEX, the authors value the position at Binance mid-prices m seconds later and subtract fees. The baseline horizon is m = 5 seconds, chosen because estimated PnL rises between 1 and 5 seconds but is comparatively stable afterwards — total estimated PnL is only 1.2% higher at ten seconds and 5.0% higher at thirty seconds relative to the five-second baseline. Unless stated otherwise, they restrict analysis to trades with positive estimated PnL.

**Linking bids to value.** They compute Pearson and Spearman correlations between auction bids and estimated PnL at the round level, then repeat the exercise aggregating over 5-, 10-, 15-, 30-minute and 1-hour windows, and also over shorter intra-round windows (first 15 and 30 seconds). They compare these patterns with the autocorrelation structure of bids and PnL.

**The transition analysis.** Three on-chain events (February 12, February 18, and February 25, 2026) split a focused February 1 – March 12 window into four sub-periods. The authors then tabulate win shares, bid-participation combinations, bid gaps, and the relationship between bids and one-second ETH/USDT realized volatility across periods.

**Caveats they state themselves.** The arbitrage heuristic only catches simple swaps in a small set of liquid assets; profit estimates assume the whole inventory is unwound at the markout horizon; DEX coverage is limited to what Dune indexes; and — critically — Kairos's parallel lane for CEX–DEX searchers operates on a subscription model with off-chain payments, so payments by the two main searchers to Kairos are unobservable. They also assume 1 USDT = 1 USDC = 1 USD. Because the methodology is applied identically across regimes, they argue any bias should affect all periods similarly.

## Why This Matters

**For research.** This is the first empirical test of theoretical warnings that ahead-of-time allocation mechanisms are vulnerable to secondary-market intermediation. The paper shows the predicted failure mode can be measured on-chain, and it supplies a template — markout-based PnL, bid-gap dynamics, volatility sensitivity — that can be reused for other ordering mechanisms.

**Real-world applications:**

- **Rollup and L2 protocol design:** Teams choosing between priority gas auctions, first-come-first-served, and AOT auctions now have evidence on how AOT interacts with resale intermediaries and reserve-price settings.
- **Auction parameter setting:** The reserve price was raised from 0.001 ETH to 0.0075 ETH and then lowered back to 0.001 ETH within a week; the paper documents both the near-shutdown of bidding at the higher reserve and the market's new equilibrium afterwards.
- **Market monitoring and competition policy:** The bid-gap metric — the wedge between the top bid and the price actually paid — is a cheap, observable indicator of weakening competition in a second-price auction.
- **Searcher and intermediary strategy:** The findings explain why routing through a JIT reseller can dominate direct forward bidding when the reseller can allocate close to the moment of value realization.

**Industry relevance.** Timeboost is the only major deployed AOT arbitrage auction for transaction ordering, so this study bears directly on live infrastructure decisions at Arbitrum and on how other rollups think about ordering-market centralization. The observation that a handful of addresses — and in one stretch, a single unattributed address setting 96% of paid prices — can determine auction outcomes is a concrete centralization risk.

## Future Directions

- **Quantify the unobserved surplus split.** Because payments through Kairos's subscription-based parallel lane happen off-chain, the paper cannot determine whether the searchers' net payoffs rose or fell. Access to those payments, or an alternative identification strategy, would let researchers close the surplus-accounting gap.
- **Distinguish coordination from collusion more sharply.** The authors use "coordination" descriptively and explicitly do not claim an observed agreement. Structural estimation of a common-value auction with a resale channel could test whether the observed bid compression is consistent with equilibrium behavior under weak competition or requires something more.
- **Test the mechanisms against each other directly.** Unbundling predicts a broader, more differentiated set of active searchers and higher total value extraction; information predicts a larger spread between primary-auction revenue and resale revenue. The paper finds limited support for both, but a longer post-transition sample and better resale-payment visibility would sharpen the test.
- **Examine design remedies.** Whether a higher or dynamic reserve price, a longer auction window, or restrictions on resale can preserve primary-auction revenue without shutting down bidding is an open policy question the paper's reserve-price episode raises but does not resolve.

## Target Audience

Protocol researchers and mechanism designers working on transaction ordering and MEV; economists studying auctions with resale and secondary markets; analysts and data scientists at rollups, market makers, and searcher firms who need to understand how Timeboost bidding actually behaves; and competition-policy researchers interested in how intermediation can emerge in thin auction markets. Readers need comfort with auction terminology but not with formal proofs — the paper's contribution is empirical, and its theoretical section is deliberately qualitative. Note that the supplied content is truncated partway through the Steady State analysis, so the paper's later discussion and concluding remarks are not reflected here.

## Authors’ abstract

We study Arbitrum's Timeboost auction, an ahead-of-time mechanism that sells a 200ms ordering advantage in an otherwise first-come, first-served transaction ordering policy. The market naturally divides into two phases: a competition phase, in which the dominant searchers compete directly in the primary auction, and a coordination phase, in which they source fast-lane access through Kairos, a Just-in-Time resale intermediary. We use auction bids, time-boosted transactions,…

[The full abstract on arXiv](https://arxiv.org/abs/2603.20175)

[Read the original paper](https://arxiv.org/abs/2603.20175)

---

Canonical: https://crypto.world/paper/just-in-time-resale-in-an-ahead-of-time-auction-for-faster-execution-2603.20175
