Research
AIDA: An AI Outage Registry with Evidence from On-Chain MEV
Joins a 5,868-incident AI outage registry to Ethereum and Base MEV activity, letting researchers test whether coding-assistant failures shift arbitrage, liquidation and sandwiching.

- arXiv
- 2611.01656
- Published
- 2026-11-01
- Authors
- Wenbin Wu
AI summary
Overview
Research area: Empirical market microstructure and crypto-asset research, at the intersection of MEV (maximal extractable value) on public blockchains and the economics of AI infrastructure reliability. The paper is filed under q-fin.TR (Trading and Market Microstructure) and is slated for the International Review of Financial Analysis.
Technical level: Intermediate. The core argument is accessible, but the paper relies on transaction-level log decoding, Poisson pseudo-maximum-likelihood (PPML) panel models, and matched-episode event studies that assume some familiarity with econometrics and on-chain data.
Scope in one sentence: The paper builds a multi-vendor registry of AI service outages (AIDA) and pairs it with verified on-chain MEV activity on Ethereum and Base to test whether AI coding tools shape who builds trading bots versus who wins execution.
What This Paper Is About
Generative AI coding assistants have made writing trading software cheaper, but no systematic dataset records when those AI services go down — so nobody has been able to measure whether trading-bot development actually depends on them. The author builds AIDA, a registry of 5,868 deduplicated AI outage incidents across 30 providers, and links it to 5,041,034 verified MEV transactions by 8,516 Ethereum addresses plus a 22-month Base Layer-2 panel. The goal is to separate two things that price-and-volume data usually conflate: the off-chain building of trading code and the on-chain execution of live trades, using AI outages as a natural experiment.
Key Contributions
-
A public AI outage registry. AIDA covers 30 providers from January 2025 to October 2026, built from 8,189 raw incident reports collapsed by complete-linkage deduplication into 5,868 distinct incidents, of which 1,543 are corroborated by multiple sources and 2,977 are verified against primary sources (94.8 percent of foundation-model incidents). It is released alongside a replication package.
-
A transaction-level MEV measurement, not a contract-heuristic one. Rather than classifying contracts by pool counts or block adjacency, the paper applies explicit rules over raw event logs to identify atomic arbitrage, sandwich attacks, and lending liquidations, then attributes each to the executing address. This yields 5,041,034 deduplicated strategy transactions by 8,516 addresses over 768 days.
-
A design that separates development from execution for the same participants. Because blockchains record software deployment and trade execution on one ledger, the author can test whether the same addresses that pause new contract releases during an AI outage keep trading with bots that are already deployed.
-
A cross-chain and multi-definition robustness check. Five alternative operator/bot definitions from the MEV literature, a matched-episode design with two reference calendars, a Base Layer-2 panel of 1,099,593,191 swaps across 29,120 private executors, and bounded Solana verification all test whether the results depend on one detection algorithm or one chain.
Main Findings
-
Entry expanded where code alone suffices. Distinct active arbitrage addresses roughly quadrupled between 2024Q4 (914) and 2026Q3 (3,749), and distinct active liquidators grew by about a third (98 to 134 distinct addresses across the two end quarters). Distinct active sandwich addresses, which require private order flow and builder access, fell by more than a third (207 to 133).
-
Execution stayed concentrated. Across all strategies, 8,516 addresses produced 5,041,034 verified transactions, and the top fifty addresses executed about two thirds of verified trades while deploying under two percent of top-level contracts. In arbitrage the top-fifty share fell from 80.32 percent in 2024Q4 to 50.25 percent in 2026Q3, yet the top ten percent still executed 82.79 percent of 2026Q3 arbitrages. In sandwich attacks the top-fifty share was 99.52 percent in 2024Q4 and 98.14 percent in 2026Q3.
-
AI outages stall new executor releases on Ethereum. Addresses releasing a private executor fall to a rate ratio of 0.67 on the day of an eight-hour disruption (and 0.68 for executor creations during OpenAI outage hours), and to 0.66 at lag two, before rebounding on day three — the pattern of a multi-day development pipeline that clears in a backlog.
-
The same pattern appears on Base. Operator-signed top-level creations fall to 0.76 during OpenAI outage hours, and executor-deploying operators fall to 0.88 at lag two.
-
Deployed bots keep trading. Non-trading contract creations, automated factory clones, constructor-executed trades, and live arbitrage and sandwich execution across both chains show no measurable change during outage windows.
-
Deployment is broadly distributed even where execution is not. Deploying addresses more than tripled between 2025Q1 and 2026Q3, while creations per deploying address fell to about a sixth and total quarterly creations fell by roughly two fifths. The weekend share of creations fell from about a quarter to about a fifth. More than three fifths of inspected 2026 runtime bytecodes have distinct code hashes, with no minimal proxies among them.
-
In liquidation cohorts, the smaller group builds while leaders execute. Pooling all in-window episodes, the smaller group deploys about three quarters of outage-window creations but executes under a tenth of outage-window liquidations.
-
Most vendor outage reports say nothing about cause. Among foundation-model developer disruptions with a stated cause, upstream cloud dependencies account for 27 percent, deployment or configuration changes 20 percent, capacity overload 19 percent, software bugs 15 percent, and network routing 14 percent — but 95 percent of vendor status reports omit root causes entirely.
Methodology in Plain English
Step 1 — Build the outage record. The author pulls incident reports from official Statuspage.io and incident.io endpoints, vendor postmortems, Wayback Machine snapshots, StatusGator and IsDown mirrors, and news or developer-forum reports. Hourly re-filed tickets (such as 35 consecutive hourly tickets from Moonshot Kimi during one 10 July 2026 outage) are merged back into single incidents. Onsets are anchored to the earliest reported start of disruption. The analysis window is restricted to the 17 providers that train and serve proprietary foundation models and to disruptions lasting 60 minutes to 24 hours, giving 709 disruptions and 2,586.5 exposure hours, of which 668 fall inside the 15,600-hour on-chain target span and cover 2,215 non-overlapping outage hours.
Step 2 — Identify strategies from raw logs. Each strategy gets an explicit rule. An arbitrage must be a closed loop of at least two swaps across at least two distinct pools inside one transaction, with each hop feeding the next hop's input token and the closing hop returning at least the initial input. A sandwich must have front-run, victim, and back-run in that order in one block and pool, with the back-run input between 90 and 110 percent of the front-run output, and with attacker legs sharing a sender or an independently established common executor. A liquidation requires a decoded LiquidationCall (Aave V2/V3, resolved via a canonical address registry because the versions share one signature) or LiquidateBorrow (Compound V2) event with strictly positive integers. The rules favor precision over recall, so arbitrage counts are explicitly lower bounds.
Step 3 — Classify creations by what they later do. A contract creation is not treated as a trading release until its on-chain role is established: persistent private executors that go on to execute identified MEV transactions, broader swap-receiving trading contracts, single-use constructor-executed trades, and non-trading contracts. On Base the paper further separates top-level creations signed by an operator key, distinct operators deploying an executor, and capped per-operator creations from uncapped internal factory spawns.
Step 4 — Estimate the outage response. Daily counts of deploying addresses are regressed on outage hours at lags zero through three using a PPML model with day-of-week and calendar-month fixed effects, clustered by Monday-anchored calendar week and checked against a 28-day block bootstrap. For an eight-hour disruption, each lag has rate ratio exp(8·β_k). An hourly within-day PPML model with UTC-date and hour-of-day fixed effects is run across all 15,600 UTC hours. A ratio of the executor rate ratio to the benchmark (non-trading) rate ratio tests whether trading software responds differently from the other contracts the same operators deploy.
Step 5 — Cross-check with matched episodes. Each outage is paired with same-duration, same-clock windows one to five days before onset and on the same weekday one, two, and three weeks earlier. Reference windows contaminated by another outage or a public cloud disruption are dropped. Confidence intervals come from 4,000 paired block-bootstrap draws. Participant rankings for the liquidation cohorts are frozen 28 days before onset, and cohort membership and labels stay fixed across the incident and reference windows. The smaller group's share of liquidations uses ranks set in a 90-day discovery window, with the top 50 addresses assigned to the leading group.
Why This Matters
Impact on research. Prior AI outage event studies have relied on one or two providers without root-cause categories. AIDA extends cloud-outage forensics to foundation-model infrastructure and gives the literature a reusable, multi-vendor, primary-source-verified panel. More importantly, the design resolves an identification problem that aggregate price-and-volume event studies cannot: with only market data, you cannot tell whether a disruption hits the building of trading software, the execution of live trades, or both. Observing both margins for the same addresses separates them cleanly.
Real-world applications.
- Operational risk management at trading firms and market makers — quantifying how much of a desk's strategy-refresh capacity depends on a handful of external AI vendors, and how quickly a backlog clears after an outage.
- Vendor concentration and procurement decisions — the registry's per-provider disruption hours (Anthropic 303 disruptions over 838.2 hours, OpenAI 175 over 887.0 hours, nine Chinese vendors 96 over 286.0 hours, six other international vendors 135 over 575.3 hours) support substitution and redundancy planning.
- Exchange and protocol governance — evidence on how concentrated MEV execution is (top fifty addresses at roughly two thirds of verified trades) informs debates about order-flow auctions and builder access.
- Regulatory and supervisory monitoring — a reproducible method for measuring third-party technology dependence in automated market activity.
Industry relevance. The paper's core economic point is that cheaper code lowers the barrier to building MEV contracts but does not change who wins execution. Private order flow, latency, and capital still decide that. That distinction matters for anyone forecasting whether AI tooling will democratize automated trading or simply widen the pool of unsuccessful entrants.
Future Directions
-
Address structure and contract upgrades. The author notes in Section 5 that address structure and contract upgrades bound the measurements — operators often deploy with one key and call with a different hot wallet, so address-level deployment counts cannot fully isolate trading-software releases. Better entity resolution across keys is an open problem.
-
Extending to Solana with a full cohort panel. The paper verifies loader and swap decoding on bounded historical strata on Solana, where program upgrades replace code in place and routers invoke concentrated-liquidity pools through cross-program invocations, and establishes the archive volume a full panel would require — but does not estimate one.
-
Recovering candidate MEV the sequential rules exclude. Graph-cycle and multi-victim candidates are recorded separately (348,284 and 401,045 respectively) rather than counted, and a few dozen transactions with token graphs too dense for the 100,000-state cycle cap are left unresolved. Tightening detection without sacrificing precision would raise the measured lower bounds.
-
Venue coverage. An audit of one unfiltered day of Ethereum logs (15 June 2026) shows transactions swapping only on unlisted pools — chiefly Fluid DEX, newer Curve crypto pools, DODO, Balancer V1 and V3, and one further concentrated-liquidity design — amount to under two percent of the selected count, but about four percent of selected transactions carry a leg the decoder cannot read, so a cycle through it is counted as broken.
Target Audience
Empirical market-microstructure and crypto-asset researchers will find the transaction-level MEV measurement and the development-versus-execution design most useful. Economists studying AI adoption, task automation, and technology dependence will value the multi-vendor outage registry and the event-study machinery. Quantitative traders, MEV searchers, and exchange or protocol researchers will care about the concentration results and what actually predicts execution. Risk managers and infrastructure planners at automated trading firms — and anyone deciding how much to depend on a single foundation-model provider — are the most direct industry audience.
Authors’ abstract
Generative AI coding tools have lowered the cost of writing trading software, but no systematic dataset records when these services fail. We introduce the AI Incident and Disruption Archive (AIDA), a multi-vendor registry of 5,868 deduplicated outage incidents across 30 providers from 2025 to 2026, with foundation-model incidents verified against primary status-page or postmortem URLs. Pairing AIDA with 5,041,034 verified maximal extractable value (MEV) transactions by 8,516…
The full abstract on arXiv