Eren Ege ÇelikWorkWritingAboutNowContact
← Writing

Microstructure and Efficiency of Polymarket's 5-Minute Cryptocurrency Binary Markets

An empirical reverse-engineering and edge-search study

June 2026·~15 min·Working paper

Abstract

We study the 5-minute up/down binary markets on Bitcoin and Ethereum operated on Polymarket, resolved against the Chainlink BTC/USD/ETH/USD price feeds. Over roughly 84 hours of recorded order-book and multi-exchange tick data (~978 BTC and ~33 ETH healthy slots), we (i) reverse-engineer the dominant market maker's (MM) pricing model, (ii) characterise the latency structure between external price feeds and the on-chain order book, and (iii) conduct an exhaustive, verifier-first search for a tradeable edge. We find that the MM prices each slot as a Brownian-probit digital option, fair = Φ((F − F₀)/(σ√τ)), with feed F = mean(Coinbase, Binance), and is well calibrated to the Chainlink resolution (Brier 0.2011 vs. our best model's 0.2089 over 404 slots). Every candidate edge we tested failed under out-of-sample, placebo and cost-inclusive verification: directional hold-to-resolution, taker latency arbitrage, spike-scalping, split-sell with profit-lock, passive market making, slot-start cheap-side, dependent-market arbitrage, and six literature-derived hypotheses. The one inefficiency that survives, a feed-to-book reaction lag exceeding 400 ms on 43% of large spikes, is not capturable as a taker because of a ~250 ms server-side execution delay, and is weak as a maker because of adverse selection. We conclude the market is efficient at the retail-accessible level. The second half of the paper is a taxonomy of the statistical artifacts that repeatedly manufactured false positives — look-ahead, spread illusion, trend skew, single-session noise, selection bias — and the verifier protocol that caught them.

1. Introduction

Polymarket lists a continuous series of 5-minute binary markets on BTC and ETH: at each slot boundary a "price to beat" is fixed, and the market resolves UP if the reference price 5 minutes later exceeds it, DN otherwise. Two tokens (UP, DN) trade on a central-limit order book (CLOB) with UP + DN = 1 at settlement. The structural appeal for a latency or microstructure trader is obvious: the resolution is a deterministic function of a public price feed, the horizon is short, and a dominant automated market maker quotes both sides continuously.

This paper documents a multi-week effort to find and monetise a durable edge in this venue, and the (negative) conclusion that none exists at the retail-accessible level. At least five distinct classes of statistical artifact independently produced "edges" of +1.5c to +9c that evaporated under verification. The negative result is the finding; the order of work that produced it is what transfers.

1.1 Central results

  1. The MM is a calibrated Brownian-probit pricer. Its mid is better calibrated to resolution than any model we built (§4).
  2. A real feed→book reaction lag exists (43% of large spikes lag >400 ms; §5) but is uncapturable because of a fixed ~250 ms server-side taker delay; the lag is mechanical: the book follows the feed rather than anticipating it.
  3. Every directional/structural edge is a false positive once costs (the bid–ask spread and taker fee) and proper out-of-sample/placebo controls are applied (§6).
  4. Efficiency is uniform across BTC and ETH and across the slot lifecycle. The only deviations from the model are transient staleness episodes that the latency structure makes unexploitable.

2. Market structure and resolution

  • Instrument. 5-minute binary on BTC (and ETH) up/down. Two ERC-1155 conditional tokens (UP, DN); CTF allows free split (1 USDC → 1 UP + 1 DN) and merge.
  • Resolution oracle. The Chainlink BTC/USD data-stream, not spot exchanges. The "price to beat" (priceToBeat in the Gamma eventMetadata) is the Chainlink tick whose payload timestamp lands exactly on the slot boundary (payload_ts_ms % 300000 == 0). UP wins iff CL[boundary+300s] > CL[boundary].
    • Correctness note. Anchoring resolution to the first observed Chainlink tick after the boundary is wrong: the ~1.2 s payload lag places that tick before the boundary, mis-signing the slot in up to ~p90 $3.66 / max $16 of drift. All resolution labels in this study use the boundary-aligned priceToBeat, verified to $0 against Gamma's eventMetadata.
  • Fees. Taker fee ≈ p·(1−p)·0.07 USDC per share (maximal near 0.50). Maker fee = 0, plus a ~20% rebate; however, 5-minute crypto markets are not in the CLOB liquidity-rewards pool (rewards.rates = None), so maker compensation is the thin rebate only.
  • Execution delay. A constant ~250 ms server-side taker delay applies to all crypto up/down markets across assets and durations (BTC/ETH/SOL/XRP/BNB/DOGE, 5m and 15m). It is network-independent, confirmed by regional-VPS latency probes, so it is a design property rather than a routing artifact.

3. Data and infrastructure

  • Capture topology. An AWS Ireland VPS records the live feeds (screen session); a local laptop runs the live model/UI (the Polymarket data WebSocket is geo-blocked from some regions but reachable from the laptop); a home archive pulls logs via cron+rsync.
  • Feeds. Binance and Coinbase direct WebSockets (sub-second, ms-stamped), Chainlink via Polymarket's real-time data stream (RTDS), plus Kraken and Bitstamp. The CLOB book (top-of-book bid/ask + sizes per token) and the trade-print stream (last_trade_price) are recorded.
  • Latency profile (one-way, to our process). Direct WS ≈ 106 ms; RTDS-relayed Binance ≈ 260 ms; RTDS-relayed Chainlink ≈ 1279 ms. The direct feeds are the only ones fast enough to matter.
  • Recorded resolutions. Two regimes. (a) May event-logs: full sub-second per-source ticks (cb/bn/kr/bs + book + Chainlink), 35 files, ~3.7 GB. (b) June panels: a 1 Hz model panel, augmented mid-study with sub-second trade-prints and raw per-event ticks. A 1 Hz panel under-resolves the sub-second lead-lag dynamics; the full-tick May data is the basis for all latency claims.
  • Coverage. ~978 healthy BTC slots and ~33 ETH slots (~84 h) after data-hygiene filtering (book partitioned by slot boundary; freshness < 8 s; slot-edge blackout t_off ∈ [0,30) ∪ (270,300]).

4. The market-maker pricing model

The dominant MM was reverse-engineered from 380 BTC slots (the "MM model card").

4.1 Functional form

fairUP(t)  =  Φ ⁣(F(t)−F0σ τ),τ=300−toff\text{fair}_{\text{UP}}(t) \;=\; \Phi\!\left(\frac{F(t) - F_0}{\sigma\,\sqrt{\tau}}\right), \qquad \tau = 300 - t_{\text{off}}
  • Feed F = mean(Coinbase, Binance), latency ≈ 300 ms. Alternative inputs (Chainlink relay; median of cb/bs/kr) were falsified by exclusive event tests (e.g. cases where the synthetic median had moved but the Chainlink relay had not, yet the mid moved → input is not the relay; cases where only Coinbase moved while Bitstamp/Kraken lagged ≥800 ms, yet the mid moved → input is not the median).
  • Anchor F₀ = feed-self at slot start. Using the Chainlink anchor instead induces a constant α·σ ≈ −$54 B–CL offset.
  • Brownian probit fit: R² = 0.92. The form is operationally indistinguishable from logistic/piecewise within < 0.1 tick.
  • Spread: 1c floor; widens with volatility and immediately post-spike.

4.2 Volatility

σ is dynamic, not constant: a constant-σ model breaks across the May→June regime. The card's estimate is

σ  ≈  3.41  +  0.40⋅realized15\sigma \;\approx\; 3.41 \;+\; 0.40 \cdot \text{realized}_{15}

and our deployed live approximation is

σ  ≈  2.10  +  0.052⋅range2\sigma \;\approx\; 2.10 \;+\; 0.052 \cdot \text{range}_{2}

(2-minute range outperforms stdev).

The σ mechanism of our "timidity." Backing σ out of the MM mid,

σimp  =  F−F0τ  ⋅  Φ−1(mid)\sigma_{\text{imp}} \;=\; \frac{F - F_0}{\sqrt{\tau}\;\cdot\;\Phi^{-1}(\text{mid})}

(36k points) gives median σ_imp = 5.21 $/√s against our deployed σ = 6.06, a ratio of 0.82. Our σ is too large, which pushes our fair toward 0.50. The MM's implied σ tracks realized variance RV (4.28) rather than jump-robust bipower BV (2.93), at σ_imp ≈ 1.1·RV. So it prices in jumps plus about a 20% premium, and our range proxy over-estimates against it. Matching σ to ~1.1·RV aligns our fair to the MM and yields no edge (§6). It reproduces an already-calibrated quote.

4.3 Calibration to resolution

Over 404 slots, the MM mid is better calibrated than our best fair: Brier 0.2011 (MM) vs 0.2089 (ours); the gap is stable, not noise. The same holds for ETH (MM mid Brier 0.166). The MM's ask is calibrated to the realized win-rate (§6.4). There is no static mispricing to fade.

4.4 Replication and staleness

A frozen card replicates the MM mid to ~oracle-σ accuracy at multi-second horizons. The only systematic deviation is transient staleness: ~19 segments totalling ~9.2 minutes where the quote lags its own model. These are the latency episodes of §5, real but unexploitable.

5. Latency and microstructure

5.1 Feed→book lead

On full-tick May data we measured, per spike (fair move ≥ 4c within 2 s), the time between the feed crossing +2c and the book mid crossing +2c (n = 1014):

  • Feed leads the book on 73% of spikes; median lead 320 ms.
  • 43% of spikes lead by > 400 ms; 26% by > 800 ms. The distribution is fat-tailed: the MM is genuinely slow on a large minority of large spikes.

5.2 Why it is not capturable (taker)

A live test (75 fires) found our fill lands ~350 ms after detection, dominated by the ~250 ms server-side taker delay, against a median book reprice ~213 ms after the trigger. We are ~137 ms too slow at the median. The clean capture window covers only ~13% of fires, and the +2c price cap misses 52% of them, which are the winners.

A capture simulation confirmed the trap. With a forward-looking spike trigger (knowing the move would reach 4c), taker capture shows +1.5c/share at 350 ms latency. But the trigger is look-ahead. With a causal trigger (fire when the past 1 s moved ≥ threshold), net is −0.6c to −1.4c, CI below zero. In real time we cannot select the slow-MM spikes: by the time a move is large enough to identify, the lag is consumed or the move has reverted. (§6/§7.)

5.3 Order-flow imbalance is mechanical, not informed

Queue imbalance QI = (bid_size − ask_size)/(bid_size + ask_size) does not lead the feed: the cross-correlation corr(QI(t), F(t+k)−F(t)) peaks at k = −1.4 s (QI reflects the past feed move), with negligible forward correlation (+0.03) and a clean placebo. The book follows the feed; there is no anticipatory informed flow to ride.

5.4 Exit symmetry is also dead

The exit-side feed lead flattens rather than reverses (≥55% of spikes never retrace ≥$3); the book self-soaks via mean reversion. A reactive trailing stop is the latency-robust exit, but it captures no directional alpha.

6. A catalogue of tested-and-rejected edges

Each row is an edge that looked positive at some stage and was rejected on verification.

#HypothesisApparent signalWhy it died
6.1Directional hold-to-resolution (buy below fair, hold)in-sample +EV at ≥12cMarket efficient; MM Brier < ours; placebo +0.019; cross-log inconsistent.
6.2Taker latency-arb (lift stale ask)book stale 213 msFill 350 ms (250 ms server delay); 52% miss = winners; causal sim −0.6c.
6.3Spike-scalp (+ trailing)+5–6c "reprice"Latency-naive: reprice window ≈ POST latency; realistic net −2.5c.
6.4Reverse favourite-longshot (late-slot favourite under-priced)+9c residual vs mid, OOS-replicated, symmetricSpread illusion: realized win-rate 0.675 ≈ ask 0.679. You pay the ask, not the mid; net at-ask −1.94c. The MM ask is calibrated.
6.5Split-sell + profit-lockbacktest +$1.87/slotOverfit: forward −$1.64; first-60s variant ~breakeven (+$0.2–0.3/slot, noisy).
6.6Passive market making / continuous quotingspread captureAdverse selection; inventory sticks; not in LP-rewards pool. Incumbent-MM PnL ≈ +0.485c/share but CI [−176,+1055]/slot — thin and high-variance; we are a worse pricer.
6.7Slot-start cheap side (MM pinned to 0.50)premise true (flip 7%→1.4%)No edge; MM calibrated at every t_off (multi-session realized−mid = −0.29c).
6.8ETH (less liquid)buy-cheap +0.18; Brier 0.166Trend artifact: side-split asymmetric (cheap-UP +0.30, cheap-DN −0.01); edge collapses 17→29 slots.
6.9Dependent-market Σ=1 arbitrageΣ deviations up to 30×Mis-grouping: non-exclusive prop/advance markets. The one valid negRisk event deviates 2.5c (sub-fee); the negRisk AMM already arbitrages true partitions.

6.10 Literature-derived hypotheses (six, via a multi-agent research workflow)

A parallel literature-research workflow generated six citable, testable hypotheses; five were tested verifier-first (the sixth, a Kou jump-diffusion form change, was predicted by the workflow itself to be replica-refinement, not edge):

  • H1 — surprise sign-flip (continuation at moderate surprise, reversion at high surprise, normalised by σ√τ). Result: no sign-flip; sign(s)·Δmid is monotone-positive (+3c→+16c→+39c) — i.e. the known, uncapturable MM-lag continuation. Placebo-clean, both halves. Dead.
  • H2 — OFI/microprice lead. Result: QI follows the feed (§5.3). Dead.
  • H3 — jump-robust σ. Result: mechanism confirmed (our σ too large; MM ≈ 1.1·RV) but matching σ only reproduces the calibrated quote. Mechanism, not edge.
  • H4 — reverse favourite-longshot at slot-end. Result: spread illusion (§6.4). Dead.
  • H6 — latency-immune post-shock slow correction (enter 2–5 s after the shock, hold). Result: fading the spike loses −6c (CI below zero), worse for larger spikes; momentum also negative. The price is already fair 3 s post-shock. Dead.

7. Methodology: verifier-first research and an artifact taxonomy

A proxy metric — in-sample PnL, residual-vs-mid, forward correlation — is not the verifier. The verifier is out-of-sample, cost-inclusive, placebo-controlled net PnL with a slot-clustered confidence interval. Repeatedly the proxy was positive and the verifier was zero or negative.

7.1 The artifact taxonomy (each independently produced a false edge)

  1. Look-ahead / forward-selection. Defining an event using future data (a spike "that reaches 4c in the next 2 s") and entering at its onset cherry-picks winners. Caught by re-running with a strictly causal (past-window) trigger: +1.5c → −0.6c (§5.2).
  2. Spread illusion. A residual measured against the mid is not tradeable; you transact at the ask/bid. The favourite "under-priced by 9c vs mid" had realized win-rate equal to the ask (§6.4).
  3. Trend skew. A directional drift over the sample makes one side's favourites win more; a pooled statistic reads as an edge. Caught by side-split symmetry, where UP-favourite and DN-favourite must agree. This killed the ETH buy-cheap edge (§6.8).
  4. Single-session noise. "Mid-slot favourites win +4.5c" in one session became −0.29c (perfectly calibrated) across many sessions. Caught by n-stability (does the edge collapse as slots accumulate?).
  5. Selection bias. Resolving on decisive/market-settlement labels rather than the Chainlink anchor manufactures +0.083 where the correct label gives −0.026.

7.2 The verifier protocol

Every candidate must pass, in order: (i) placebo (shuffle the signal → effect → ~0); (ii) side-split symmetry; (iii) n-stability (no collapse as data grows); (iv) chronological OOS (fit on early data, freeze, test on later); (v) cost-inclusive net at the executable price (ask/bid) with the taker fee; (vi) slot-clustered bootstrap CI excluding zero. A point estimate without these is treated as noise.

7.3 Tooling

Reusable single-file analyses: mm_card.py (MM replication), spike_lead.py / latency_capture(_causal).py (latency), jump_sigma.py (σ decomposition), ofi_lead.py (microstructure), surprise_signflip.py, reverse_flb.py / favorite_late.py, post_shock.py, mm_pnl.py (incumbent profitability from trade-prints), calib_framework.py / calib_by_second.py (calibration), plus a slot-review viewer (ui/slot_review.html) with full-tick raw-stream inspection.

8. Synthesis: the 5-minute crypto market is efficient

The evidence converges from every direction:

  • Pricing: the MM is a correctly-specified Brownian-probit digital pricer with realized-variance σ; its mid and ask are calibrated to the Chainlink resolution (§4).
  • Latency: a real feed→book lag exists but is mechanical and uncapturable under the 250 ms server-side taker delay (§5).
  • Microstructure: order flow is reactive, not informed (§5.3).
  • Cross-section: ETH (less liquid) and dependent markets are equally efficient once artifacts are removed (§6.8–6.9).
  • Cost wall: the bid–ask spread plus the p(1−p)·0.07 taker fee is the apparent magnitude of every "edge". The apparent edges are the spread (§6.4).

The only positive quantity we measured is the incumbent MM's thin liquidity-provision margin, ~+0.485c/share on enormous volume with high variance. That is a market-making business we cannot contest from a retail position: worse calibration, no queue priority, no rewards pool.

Conclusion. Within the retail-accessible action space (taker entries, small passive maker orders, free CTF split/merge), Polymarket's 5-minute crypto binaries admit no durable, capturable edge. This is a strong, multiply-confirmed negative result.

9. Implications and future directions

  • The only lever that ever existed was latency. It needs both a fast continuous external feed and an MM that lags it. Both are present in crypto, and the server-side taker delay neutralises them. Event, sports and geopolitical markets have no such feed, so their edge would have to be modelling, which the liquid order books (1c spreads) already price efficiently.
  • Directions requiring an operator decision (not pursued autonomously):
    1. Different venue or horizon. Kalshi, or longer-horizon and lower-liquidity markets where modelling edges may survive. Requires API/auth integration.
    2. Paid market-making. Daily commodity and FX up/down markets are in the rewards pool, so the edge is the rebate rather than direction. Requires MM infrastructure and inventory risk management.
    3. Price-ladder monotonicity arbitrage. Nested "above $X" thresholds must be monotone. Violations, likely only in illiquid rungs, are risk-free but size-limited.
  • What survives the negative result: a validated MM model card, a latency map, and the verifier protocol and artifact taxonomy of §7, which apply to any short-horizon prediction-market edge search.

Appendix A — Key quantitative results

QuantityValueSource
MM functional formΦ((F−F₀)/(σ√τ)), R²=0.92mm_card, 380 slots
Feedmean(Coinbase, Binance), ~300 msexclusive falsification
MM Brier vs ours0.2011 vs 0.2089 (404 slots)calib_check
σ implied (MM) / deployed (ours) / RV / BV5.21 / 6.06 / 4.28 / 2.93jump_sigma, 36k pts
Feed→book lead >400 ms / >800 ms43% / 26% (median 320 ms)spike_lead, 1014 spikes
Taker delay / our fill / book reprice250 ms / 350 ms / 213 mslive, 75 fires
Latency capture: forward / causal+1.5c / −0.6c per sharelatency_capture(_causal)
Reverse-FLB: residual-vs-mid / net-at-ask+9c / −1.94creverse_flb, favorite_late
Split-sell: backtest / forward+$1.87 / −$1.64 per slotss_walkforward
Incumbent MM PnL+0.485c/share, CI [−176,+1055]/slotmm_pnl, 6.2M shares
Taker feep(1−p)·0.07 USDC/share—
Server-side taker delay~250 ms, all crypto up/downmarket scan

Appendix B — Data inventory

  • ~978 BTC + ~33 ETH healthy slots (~84 h) after hygiene filters.
  • May full-tick event-logs: 35 files, ~3.7 GB (per-source ms ticks + book + Chainlink).
  • June panels: 1 Hz model panel + sub-second trade-prints + raw per-event ticks (added 2026-06-17).
  • Resolution label: Chainlink boundary tick (priceToBeat), verified against Gamma eventMetadata.

Appendix C — Glossary

  • Slot — one 5-minute market.
  • t_off — seconds since slot start (0–300); τ = 300 − t_off.
  • F, F₀ — feed price and its slot-start value.
  • fair / mid / ask — model probability / book midpoint / best ask.
  • CL anchor — Chainlink price-to-beat defining resolution.
  • RV / BV — realized variance / jump-robust bipower variation.
  • Verifier-first — accept an edge only after placebo + side-split + n-stability + OOS + cost-inclusive net + slot-clustered CI.

A negative-result paper. What it offers is the rigour of the rejection and a reusable protocol, not a trading strategy.

marketspolymarketmicrostructureempirical
← All writingNewer: The verifier-first protocol