{"id":"93c77b84-6a6f-4bc9-bfe5-476f8d6c84cf","arxiv_id":"2607.09491","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Heuristic detection finds 402 M likely one-way arbitrage sequences on Binance (0.94% volume, $31.2 M net profit) and ~2 M on Kraken, with average per-sequence profits under $1 after fees.","lead":"A new heuristic matches anonymized spot trades by quantity and timing to detect one-way arbitrage on Binance and Kraken. It finds hundreds of millions of sequences that make up under 1% of volume yet accumulate tens of millions in estimated net profits while growing faster and less profitable over time.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Thresholds for same-trader OWA inference were selected by maximizing the profitability metric later reported as the main result, creating circular dependence between detection and claims.","rationale":"The Reader correctly isolates the same-trader inference and the profitability-tuned thresholds as the weakest assumption. That circularity is load-bearing for the absolute volume and profit numbers that form the paper’s strongest claim; the existence of a high-speed, low-margin pattern and its longitudinal maturation are more robust. No stronger internal inconsistency appears: the heuristics are transparent, the data sets are large, and the authors already flag the lack of ground truth and the conservative fee assumptions. A sensitivity sweep on held-out data would settle whether the absolute figures survive modest re-parameterization; until then CONDITIONAL remains the appropriate verdict. No change to the Reader’s assessment is required.","tokens_in":24118,"tokens_out":582,"duration_ms":5417,"concrete_test":"Re-run the full pipeline on the held-out months (or a random 20 % of the candidate set never used for threshold selection) while sweeping Δq ∈ {0, ±0.05, ±0.1, ±0.2} bps and Δt windows ±50 % around the published values; report how sequence count, volume fraction, and net-profit estimate change. If any of these quantities moves by more than ~25 % under modest threshold perturbation, the absolute figures are unstable and the strongest claim must be restated as a lower-bound or order-of-magnitude estimate.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (402 M sequences, 0.94 % volume, $31.2 M net profit) rests on the inference that pairs of trades with Δq ∈ [-0.1, 0.1] bps and Δt inside the chosen windows (200/400 ms Binance, 1–2 s Kraken) are executed by the same rational OWA trader. Sections 4.3–4.5 explicitly construct these thresholds by inspecting CDFs of volume/gains/losses and profit-dominance plots on a development set and retaining only the parameter region where aggregate gains exceed losses (and volume concentrates). Because the same profitability signal is both the selection criterion and the headline result, the absolute counts and dollar figures are not independent of the detection rule. The paper prioritizes precision over recall and notes the absence of ground truth, yet still reports the tuned numbers as the primary measurement of OWA prevalence and profit. Longitudinal trends (speed-up, profit compression) are less sensitive to the exact cut-offs, but the absolute scale is.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes a quantity-and-timing matching methodology to detect one-way arbitrage (OWA) sequences in anonymized CEX spot trade data. After coarse filtering on coin types, quantity difference, latency, and order type, the authors refine thresholds on a development set by retaining parameter regions where volume and hypothetical gains concentrate and profit dominance is positive (Section 4.4, Figs. 3–5). Applying the resulting cut-offs (Δq ∈ [−0.1, 0.1] bps; latency 200/400 ms on Binance and 1–2 s on Kraken) yields 402 M sequences on Binance (0.94 % of volume, estimated $31.2 M net profit after lowest published fees) and ~2 M on Kraken (0.13 % volume, $975 k). Per-sequence net profits average well under $1; longitudinally, latencies fall and per-sequence gross returns compress. The authors interpret the activity as professional, competitive OWA that helps rebalance order books.","tokens_in":24366,"tokens_out":1171,"duration_ms":9998,"significance":"If the inference is accepted, the work supplies the first large-scale measurement of executed (rather than merely theoretical) OWA on major CEXes, together with two longitudinal data sets that the authors intend to release. The scale (hundreds of millions of sequences, nearly $190 B volume on Binance) and the clear trends toward lower latency and smaller price discrepancies are of genuine interest to market-microstructure and crypto-finance audiences. The methodology itself—matching on quantities and timestamps when trader identifiers are absent—is a reusable contribution for other anonymized order-flow studies. The explicit prioritization of precision over recall and the transparent reporting of fee-scenario ranges are strengths that make the measurement usable even if absolute counts remain somewhat threshold-dependent.","major_comments":[{"comment":"Sections 4.4–4.5 and Figs. 3–5: the quantity-difference window and latency cut-offs are chosen by inspecting the same volume/gains/losses CDFs and profit-dominance plots that later become the headline results. Because the retained region is defined as the region where aggregate gains exceed losses, the absolute counts (402 M sequences), volume share (0.94 %), and net-profit totals ($31.2 M) are not independent of the detection rule. The paper correctly notes the absence of ground truth and the precision-over-recall stance, yet still presents the tuned numbers as the primary measurement of OWA prevalence. At minimum the authors should (i) report sensitivity of the headline figures to modest changes in the cut-offs, (ii) show that the longitudinal trends (speed-up, return compression) survive alternative thresholds, and (iii) clearly separate the development-set selection step from the fin","section":null},{"comment":"Section 5.1 and the fee assumptions: net-profit estimates rest on the lowest published fee tiers (2.4 bps taker / 1.2 bps maker on Binance; 8 bps / 0 bps on Kraken) and on the modeling choice that OWA must cover two fees while the direct-conversion reference price incurs none. The paper itself shows that under this “double-fee” scenario taker/taker cumulative profits on Binance stagnate or decline for long periods while volume remains high—an outcome the authors attribute to possible private fee arrangements. Because the absolute dollar figures are sensitive to these unobservable fee levels, the $31.2 M / $975 k totals should be presented more explicitly as a lower-bound range under stated assumptions rather than as point estimates of realized profit.","section":null}],"minor_comments":[{"comment":"Section 4.3: the initial Δq intervals ([−10, 1] bps Binance, [−40, 1] bps Kraken) are motivated by published fee schedules, yet the final analysis discards all in-band-fee candidates. A short quantitative note on how much volume is thereby excluded would help readers gauge the precision–recall trade-off.","section":null},{"comment":"Figures 7 and 12: the quarterly box-plots of gross returns would be clearer if the y-axis scales were identical across maker/taker and across exchanges, facilitating direct visual comparison of the claimed compression.","section":null},{"comment":"Section 3.1: the mark-to-market procedure for Binance (BTC VWAP then Kraken BTC/USD) is reasonable but should note any residual basis risk when the two exchanges temporarily diverge.","section":null},{"comment":"Table 2 and the surrounding text: the final latency thresholds (200/400 ms Binance, 1–2 s Kraken) are rounded upward from the profit-dominance zero-crossings; stating the exact zero-crossing values alongside the rounded cut-offs would improve reproducibility.","section":null}],"recommendation":"major_revision","confidential_remarks":"The circularity concern raised by the stress-test is real and load-bearing for the absolute scale claims, but it is fixable by sensitivity analysis and clearer separation of development versus evaluation. The longitudinal trends appear more robust and are the paper’s most interesting contribution. I would not reject on novelty or scope grounds; the measurement fills a genuine gap. If the authors supply the requested robustness checks, the paper would be suitable for the journal."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The real news here is that they actually recovered executed one-way arbitrage sequences from fully anonymized Binance and Kraken spot data at multi-year scale—something the opportunity-only CEX papers and the address-linked DEX papers never did. 402 M sequences on Binance (0.94 % of volume) and ~2 M on Kraken is a new empirical object, and the longitudinal picture (latency collapsing toward 10 ms, per-sequence gross returns shrinking) is consistent across both venues and both maker/taker modes.\n\nWhat they did well: the matching logic is transparent (quantity match + short latency + anchor–non-anchor–anchor + end-with-taker), they ship the full candidate sets and the final filtered sets, and they are explicit that they prioritize precision over recall and have no ground-truth trader IDs. The profit-dominance plots (Fig. 5) and the CDFs of volume/gains/losses give a clear visual of why they cut where they cut. The fee sensitivity analysis (gross vs double lowest-tier) is also honest; they show the plateau in taker/taker net profits and discuss private rebates.\n\nThe soft spot the stress-test flags is real but not fatal. They did tune Δq ∈ [−0.1, 0.1] bps and the latency windows on a development set by looking at the same profitability signal they later report. That makes the absolute dollar figures ($31.2 M / $975 k) and the exact sequence counts partly by construction. Longitudinal trends and the qualitative claim that OWA is live, competitive, and compressing are far less sensitive to the precise cut-offs. Fees remain an assumption (lowest published tiers), and they cannot see failed attempts or multi-leg variants. All of that is stated in the limitations; they do not oversell.\n\nThis is for market-microstructure people, exchange operators, and anyone who needs to know what HFT actually looks like inside the dominant CEXes. The citation pattern is clean—they correctly position against both the theoretical CEX opportunity papers and the on-chain DEX arbitrage literature. Math is elementary but appropriate; data handling is careful.\n\nI would send it to peer review. The contribution is real enough, the caveats are visible enough, and the data sets will be useful even if referees force a more conservative framing of the absolute numbers. Worth reading and worth citing for the measurement itself.","headline":"First large-scale recovery of executed OWA from anonymized CEX trades; absolute counts are threshold-dependent, but the existence and maturation trends look solid.","tokens_in":24964,"tokens_out":589,"would_cite":true,"duration_ms":8035,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"One-way arbitrage is real on big crypto exchanges: 402 million sequences on Binance alone, but average net profit is pennies.","keywords":["one-way arbitrage","centralized exchanges","spot trade data","cryptocurrency markets","algorithmic trading detection","price discrepancies","Binance","Kraken"],"falsifier":"If an exchange later released even a small sample of trades tagged by trader identifier, one could count how many of the paper’s matched pairs actually share the same identifier; a low match rate would falsify the detection method.","tokens_in":25007,"feed_emoji":"₿","tokens_out":654,"duration_ms":6050,"temperature":0.7,"pith_summary":"Centralized crypto exchanges publish anonymized spot trades with no trader IDs, so no one knew whether algorithmic strategies such as one-way arbitrage (converting A to B via a cheaper intermediate coin I) actually occur. This paper supplies a practical detection method that matches pairs of trades by nearly identical intermediate quantities and very short time gaps, then keeps only the parameter ranges that are profitable in aggregate. Applied to five-plus years of Binance data and nine years of Kraken data, the method recovers hundreds of millions of likely sequences that account for roughly one percent of Binance volume and a tenth of a percent of Kraken volume. Aggregate net profits after the lowest published fees reach tens of millions of dollars, yet the typical sequence earns well under a dollar; over time the trades have become both faster and less profitable. The result shows that pricing discrepancies are routine on CEXes and that professional, high-speed bots already harvest them at scale.","feed_headline":"402 million one-way arbitrage trades found on Binance","feed_subtitle":"Net profit is only pennies per trade, yet totals $31 million; speed is rising as margins shrink","key_machinery":"Quantity-and-timing matching refined by aggregate profitability: candidate pairs must exchange nearly the same amount of the intermediate coin (within 0.1 basis points) inside a short latency window (200–400 ms on Binance, 1–2 s on Kraken), and only the latency bins that remain net-positive after fees are retained.","core_discovery":"One-way arbitrage is an actively used trading strategy on major centralized exchanges: the authors recover 402 million likely sequences on Binance (0.94 percent of total volume, roughly $31 million net profit after lowest published fees) and almost two million on Kraken (0.13 percent volume, $975 thousand), with individual sequences averaging less than one dollar net and with clear trends toward lower latency and smaller price gaps.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["402M one-way arbitrage sequences found on Binance","One-way arbitrage yields $31M total on Binance after fees","CEX one-way arb: 402M sequences, sub-$1 avg profit each","Measuring OWA in the wild: 402M Binance trades, rising speed","OWA on Binance and Kraken: thin margins, growing latency edge"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The method treats any pair of trades that match the chosen quantity and latency windows and that are profitable in bulk as belonging to the same rational one-way-arbitrage trader; there is no ground-truth identity data to check that claim.","fun_headline_variants_meta":{"raw":{"variants":["402M one-way arbitrage sequences found on Binance","One-way arbitrage yields $31M total on Binance after fees","CEX one-way arb: 402M sequences, sub-$1 avg profit each","Measuring OWA in the wild: 402M Binance trades, rising speed","OWA on Binance and Kraken: thin margins, growing latency edge"]},"model":"grok-4.5","effort":"low","cost_usd":0.003648,"raw_usage":{"total_tokens":1188,"prompt_tokens":774,"num_sources_used":0,"completion_tokens":89,"cost_in_usd_ticks":36480000,"prompt_tokens_details":{"text_tokens":774,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":325,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":774,"tokens_out":89,"duration_ms":3827,"temperature":1.0,"reasoning_tokens":325,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T02:38:14.948465+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"If an exchange later released even a small sample of trades tagged by trader identifier, one could count how many of the paper’s matched pairs actually share the same identifier; a low match rate would falsify the detection method.","supporting_citations":[],"review_version":1}