Crypto on the Clock
An Empirical Study of Short-Term Crypto Markets Microstructure on Polymarket and Kalshi
Inside the fast-settling market structure behind Kalshi and Polymarket. Together, the two venues now run a roughly $7.8B market in near-term crypto bets since 2026 Jan to end of June, fast, binary contracts on where an asset’s price lands 5, 15, or 60 minutes out, settling in minutes or hours. We examined the microstructure of these markets and traced the economic incentives that drive the volume: maker rebates on one side, and trading patterns consistent with settlement-driven strategies on the other.
Analysis covers 2026 activity through June 2026 using Polymarket on-chain trades, Kalshi trade data, and second-level Binance spot data for the settlement study. Partial periods are labeled in the relevant charts.
- The category is real, and Polymarket leads it. Roughly $7.8B of the venues’ combined $10.07B in crypto volume traded in contracts measured in minutes or hours. Polymarket traded $5.59B, versus $4.48B on Kalshi, though Kalshi gained crypto share every month.
- The two platforms made different design choices. Both venues’ volume centers on BTC among assets with live markets, but Polymarket offers micro-ticket trading in 5-minute markets while Kalshi’s shortest contract is 15-minute.
- More volume does not necessarily mean more fee capture. Polymarket moved about 1.25x Kalshi’s crypto dollars but shows about $53.9M in observed per-fill fees over Jan. 5-June 22. On rate, Kalshi’s modeled average fee runs about 2.8x Polymarket’s observed rate (2.74% versus 0.96%). That is a fee-rate gap, distinct from the 3.6x volume-share figure, and it is directional because it sets Kalshi’s modeled fees against Polymarket’s observed fees rather than like-for-like revenue.
- Polymarket’s 5-minute market is bot-driven, with footprints consistent with settlement-driven trading. Bot-like wallets account for roughly 86% of taker dollars and appear on both sides of about 79% of volume. The classification is behavioral, not confirmation that any specific wallet is automated.
- Echoing a framework first published by Stanford University researchers, Binance volume spikes repeatedly align with Polymarket settlement. In still-even cycles, Binance spot volume rises roughly 12–17× in the final seconds, while price moves into the close and then partially reverses. The pattern is consistent with settlement-driven trading, but does not show that the Chainlink benchmark was moved.
- Trading outlasts wallets. By week four, roughly 20% of wallets remain active, compared with about 39% of trades and 30% of dollars. The product retains an active trading core far better than it retains the broader wallet cohort.
Short-term crypto markets are rising, but Kalshi and Polymarket are building different products.
Kalshi and Polymarket are both expanding prediction markets beyond traditional event contracts, and crypto has become one of the clearest examples of that shift. Sports still anchors the broader category, with the majority of activity on Kalshi at roughly 79% of volume and the largest category on Polymarket with roughly 48% of volume share. But crypto is developing into a different kind of prediction-market product: recurring, short-duration markets that let traders take a binary view on price movement over a defined window.
That structure matters. These contracts do not behave like the slower, event-driven markets most people associate with prediction markets, where the question resolves around an election, a policy decision, or a company outcome. They also are not perpetuals or options. They are exchange-traded binary markets that settle in minutes or hours, making them closer to a short-term trading product than a traditional forecasting market.
That short-duration activity is already the center of crypto trading on both venues, not a side experiment. An estimated $7.8B of the two venues’ combined $10.07B of crypto volume over the January-to-June window traded in contracts measured in minutes and hours. Polymarket’s 5-minute product alone reached $2.1B by May 25, less than four months after first appearing in the duration series. Demand for the category is not the open question.
The open question is how different design choices shape different market structures — user behavior, economics, and strategies. Kalshi and Polymarket are building around similar demand, but they structure the category differently across contract length, asset breadth, trading cadence, and fee structure. Those choices shape what the trades look like, who is trading, and who earns.
Polymarket leads in crypto markets. Kalshi is gaining ground.
Across the past 6 months window from 2026 Jan to June, Polymarket traded $5.6B of crypto, about 25% more than Kalshi’s $4.5B. Yet Kalshi is the larger platform with higher volume overall with all market categories combined. The difference is venue mix: Kalshi is overwhelmingly driven by sports, while Polymarket’s activity is spread more evenly across sports, crypto, and politics.
Kalshi: sports-heavy venue, growing crypto segment
| Rank | Category | Volume | Dollar share | Trade count |
|---|---|---|---|---|
| 1 | sports | $27.17B | 79.49% | 371.7M |
| 2 | crypto | $4.48B | 13.11% | 142.0M |
| 3 | exotics | $1.17B | 3.42% | 25.7M |
| 4+ | all other categories | $1.36B | 3.98% | – |
Polymarket: larger crypto market, less concentrated mix
| Rank | Category | Volume | Dollar share | Trade count |
|---|---|---|---|---|
| 1 | sports | $10.57B | 48.24% | 119.0M |
| 2 | crypto | $5.59B | 25.51% | 588.2M |
| 3 | politics | $4.27B | 19.49% | 38.8M |
| 4+ | all other categories | $1.48B | 6.75% | – |
Shares use the displayed category totals as the denominator; trade counts are venue-reported trade rows, and 4+ remainder rows do not have per-category trade-count breakouts.
The weekly category mix shows that contrast in motion. Kalshi remains a sports-first venue, but crypto grows from a thin strip into an increasingly visible part of the exchange. On Polymarket, weekly crypto volume swells into March, recedes into late May, and then shows a June rebound inside a much more balanced venue mix.
That makes the race more dynamic than the cumulative totals suggest. Polymarket still has the larger crypto market, but Kalshi is steadily increasing crypto’s importance within its venue. Kalshi’s crypto share rose every month, from 4.3% in January to 17.9% in June. Polymarket peaked at 34.0% in February and eased to 19.8% by June.
Polymarket leads on scale over the full window; Kalshi has the stronger share momentum within its own exchange. The next comparison is not size but structure: how each venue has built the product.
The same bet becomes a different market when the clock shrinks.
The divergence begins with cadence. Polymarket’s 5-minute contract appears to have found the strongest adoption at the extreme short end of the market, where traders face lower time-in-position risk and can model near-term outcomes more easily. Kalshi starts at 15 minutes, and activity concentrates around 15-minute and hourly BTC contracts. Both venues sell short-horizon crypto, but Polymarket pushes the format closer to continuous trading.
Asset breadth reinforces the difference. Kalshi is almost entirely BTC. Polymarket still revolves around BTC, but ETH, SOL, XRP, and other assets form a meaningful second layer.
That distinction shows up in the tape. Polymarket records many more trades at much smaller typical sizes, while Kalshi records fewer, larger tickets. The distribution makes the difference visible beneath the headline volume.
Kalshi’s typical trade is larger. Its median is about $4.70, roughly one and a half times Polymarket’s $3.05, and its middle 50% runs from $0.96 to $19.79, compared with $1.14 to $6.96 on Polymarket. Kalshi stays ahead in the upper tail. At the 95th percentile, it reaches about $123 against $33.51 on Polymarket. At the 99th, the comparison is $385 against $124. At the 99.9th, it is roughly $1.1K against $967.
The shape is the finding. On the full crypto tape both venues trade small, with median tickets of $3–$5, but Kalshi spreads across a broad $1–$100 range while Polymarket spikes sharply at a few dollars: roughly four in five Polymarket trades are 5- and 15-minute micro-tickets. The occasional whale trade still exists on Polymarket, but it now sits far beyond the 99.9th percentile, swamped by the micro-ticket flood rather than forming a visible fat tail.
That gives the two venues different fee bases. Kalshi processes steadier, mid-sized trades. Polymarket processes a torrent of tiny fills punctuated by occasional whales. More volume does not necessarily produce more fee revenue. The result depends on how each venue prices those trades and how its fee structure interacts with ticket size and contract price.
More volume does not mean more collected fees.
Polymarket traded $5.59B of crypto over the window, about 1.25x Kalshi’s $4.48B, but collected about $53.9M in observed per-fill fees. For Kalshi, the trade data do not report collected fees, so we modeled fees by applying Kalshi’s published fee schedule as a lower-bound assumption. That produces a higher $122.1M modeled fee estimate.
Polymarket’s reported fee total is measured before separate daily rebates and reflects a staggered rollout rather than a single March start: taker fees began on 15-minute crypto markets on Jan. 5, 5-minute crypto markets launched with taker fees on Feb. 12, fees expanded to new all-crypto markets on Mar. 6, and Fee Structure V2 expanded fees across most categories on Mar. 30. Kalshi’s figure is modeled from its fee schedule beginning Feb. 5, so the headline totals compare different fee-active windows. On fee-active weeks only, Polymarket’s average take rate is 0.96%, versus 2.74% for Kalshi, putting Kalshi at roughly 2.8x Polymarket’s average.
Maker volume contracted after Polymarket’s rule changes.
On March 30, Polymarket expanded fees across more market categories. On April 28, it migrated trading to a new exchange system. The cutover changed fee handling, collateral, order signing, and integrations; it also cleared the resting order book, requiring market makers to reconnect and rebuild their quotes.
Weekly crypto volume peaked in March and declined through April, while fee collection continued across the transition. The $53.9M total spans the full period: fees before the cutover reflect the old transaction-refund mechanism, while fees afterward follow the new system.
Breaking that same weekly volume down by pre-migration maker tier shows where the decline came from. The largest maker cohorts pulled back the most, with one step down after the March fee expansion and another around the April 28 cutover.
The timing suggests two separate pressures. The fee expansion may have changed the economics of the flow makers were quoting. The migration then created an operational reset: integrations had to move, resting orders disappeared, and quotes had to be rebuilt. Changes in execution protection may also have made some latency-sensitive strategies less attractive.
Maker tiers are internal buckets based on each maker’s pre-V2 crypto volume and fill count, held fixed after assignment. “Active makers only” excludes zero-volume weeks, and sporadic or unclassified maker spillover rows are excluded.
The timing is suggestive, not conclusive. Because these maker tiers together make up weekly volume, the breakdown does not independently prove that the exchange changes caused makers to leave. It shows a simpler pattern: the larger the pre-migration maker cohort, the sharper the pullback, with the decline steepening around the two changes. That suggests the highest-volume makers were most sensitive to the venue changes, likely because the new structure offered them less favorable fee incentives.
The short-term cycle market is the fee engine.
Collected fees are concentrated in the same products that dominate trading. On Polymarket, the 5-minute and 15-minute contracts account for roughly 90% of observed crypto fees. On Kalshi, the 15-minute-and-under and hourly contracts account for roughly 94% of the modeled fee estimate.
The fastest contracts therefore drive both the activity and the fee base. But the fee and maker data only explain part of the market structure. The next question is who is generating the short-term flow that remains.
25% of wallets drive over 85% of the 5-minute market’s volume on Polymarket.
We further investigated the user base of these short-term crypto markets. Because Kalshi does not publish user-level IDs in public datasets, the discussion below is scoped to Polymarket.
Polymarket’s 5-minute market generates large settled volume, but the flow is not evenly distributed across wallets. The first question is concentration: whether activity is spread across many occasional traders, or concentrated in accounts that trade repeatedly, systematically, and across many markets.
We screen every wallet with at least 25 trades in the 5-minute market. The screen is intentionally simple. It looks for behavior that is difficult to explain as manual trading: very high daily trade counts, highly repetitive ticket sizes, or activity across thousands of distinct markets. This is not a trained classifier and does not prove intent. It is a behavioral label.
A wallet is tagged as bot-like if any of the following hold:
- more than 100 trades per active day;
- 500+ trades at a near-constant ticket size, defined as the P95 ticket sitting within 15% of the median;
- more than 1,000 distinct markets traded.
Wallets that do not trip those rules are then split into sophisticated and retail. A wallet is tagged as sophisticated if it trades large and selectively, either with a $200–$500+ average ticket across relatively few markets, or more than $50k of volume across 30 or fewer markets. Remaining active wallets are tagged as retail.
Classification methodology note
Bot-like behavior is screened first; sophisticated wallets are large and selective; retail-labeled wallets are the remaining active 5-minute takers above the trade floor.
Classification methodology note
Bot-like behavior is screened first; sophisticated wallets are large and selective; retail-labeled wallets are the remaining active 5-minute takers above the trade floor.
| Bot-like | Sophisticated | Retail | |
|---|---|---|---|
| Wallets | 54.8K | 630 | 157.6K |
| Share of wallets | 25.7% | 0.3% | 74.0% |
| Share of volume | 86.3% | 1.0% | 12.7% |
| Avg ticket | $6.80 | $451.19 | $7.29 |
Classification uses 5-minute taker wallets with 25+ trades and applies rules in order, with near-constant tickets defined as p95 within 15% of the median. The 25-trade floor excludes one-off retail, so true venue-wide retail share would be higher and its volume share lower.
Classified this way, the 5-minute market is highly concentrated. Bot-like wallets are a minority of active wallets, but account for most trades and most volume.
The behavior is visible from several angles. By ticket size, wallets with a median trade of $5 or less make up 90.4% of active wallets and produce 96.0% of trades. By cadence, wallets with a median gap of one second or less carry 84.0% of 5-minute volume. By concentration, the top 1,000 wallets account for 43.3% of volume, while the top 10,000 account for 73.1%.
When we look at the two sides of each fill, maker and taker matter. A wallet can post liquidity in one trade and cross the spread in another, so maker/taker is not a permanent wallet type. In the Polymarket fill data, we identify the maker as the wallet on the resting order that was already in the book, and the taker as the wallet whose incoming order crossed that liquidity and completed the fill. We classify each wallet using the same behavioral rules, then look at whether the maker wallet and taker wallet in each trade are bot-like, sophisticated, or retail-labeled.
In the 5-minute market, bot-like wallets appear on both sides of 78.9% of volume. Only 1.4% of volume has no bot-like wallet on either side.
Bots are a natural fit for a product built around five-minute resets, repetitive markets, market-making incentives, and prices tied to assets that trade continuously elsewhere. The product rewards speed, monitoring, and execution more than discretionary decision-making. Taken together, the wallet screen and the maker/taker counterparty mix point to the same conclusion: Polymarket’s 5-minute crypto market has a durable bot-like trading core, with a much thinner retail-labeled layer around it.
However, this leaves room for automated strategies that do more than passively make markets. A bot-like wallet does not sit passively in the order book. It can also cross the spread when speed matters, including in the final seconds before a contract resolves. In particular, building on research published recently by Stanford University, we replicated the paper’s methodology on Polymarket’s extreme short-term 5-minute markets and found patterns consistent with its thesis.
Bitcoin trading on Binance spikes at the close when the Polymarket resolution can still flip.
Polymarket’s 5-minute and 15-minute Bitcoin markets pay based on whether the Chainlink BTC/USD reading at the end of the contract window is at or above the reading at the beginning of that window, as specified in the market rules. Because that benchmark is derived from prices in tradeable crypto markets, a trader can hold a Polymarket position while simultaneously trading Bitcoin elsewhere, and particularly on the venues whose prices are sourced for determining the resolution.
The opportunity is easiest to understand with a simple example. Suppose Bitcoin begins a five-minute window at $100,000 and the Polymarket contract is still trading near 50/50 with seconds left. A trader holding Up can also buy Bitcoin in the underlying market. That trade might be a hedge against the Polymarket position or an arbitrage between the two markets. But if the pressure contributes to a broad enough move in Bitcoin, the closing benchmark could end above its opening level, and which side pays out changes with it. The risk and profit of this strategy are modeled mathematically in the paper mentioned below, which first described the strategy and documented supporting evidence.
The same observed spot trade could therefore reflect hedging, arbitrage, inventory management, or positioning around the settlement itself. Trade data cannot identify the motive. What it can show is whether the underlying market behaves as settlement-driven trading should. Most of that activity could be ordinary hedging, arbitrage, or inventory management. But in principle, a broad enough move in the underlying Bitcoin market could also affect the settlement benchmark and determine which side of the Polymarket contract pays out.
Our analysis asks whether that cross-market incentive leaves a repeatable footprint. We use second-by-second Binance spot data as a proxy for activity in the underlying Bitcoin market, not as Polymarket’s settlement source. If the trading is settlement-driven, it should appear on Polymarket’s clock: in the final seconds, primarily when the outcome can still flip, and then fade after settlement.
Methodology and attribution
This section builds directly upon the framework introduced by David Dai and Ruizhe Jia of Stanford University and Shihao Yu of Singapore Management University.
Methodology and attribution
This section builds directly upon the framework introduced by David Dai and Ruizhe Jia of Stanford University and Shihao Yu of Singapore Management University.
Their paper, “Settlement Manipulation in Prediction Markets” (2026), supplies the central framing, the conditioning on still-even cycles, and the horizon-as-defense test. Their work first formalized the settlement-manipulation hypothesis and proposed the empirical framework for studying it. Rather than reproducing their analysis, we extend it with an independent dataset, additional market microstructure metrics, and several new empirical visualizations using our May 2026 Polymarket and Binance data.
Their core screening question is straightforward: can a feasible trade in the markets underlying a settlement benchmark move that benchmark before the contract resolves? Short contracts are more exposed because they frequently reach the close with the outcome still balanced on a very small price difference. Longer contracts reduce the opportunity because the underlying asset has usually moved far enough from its starting point that the winner is already clear.
Based on Binance 1-second spot data and Polymarket trade-level prices, May 2026, for 5-minute (~8,900 cycles) and 15-minute (~2,900) BTC up/down markets; the drift chart uses BTC/USDT over Dec. 2025–May 2026. “Still-even” means the contract’s implied probability sat between 0.40 and 0.60 in its final minute; the 15-minute still-even sample is small (n≈27). The pattern is consistent with a settlement-driven move; it is a statistical footprint, not a claim about any identified account. Full methodology and figures are in the accompanying settlement-integrity study.
The test: compare contracts that can still flip with those already decided.
Each run of a short-duration contract is one cycle: a five-minute bet closes, and another begins. We align every Polymarket cycle with second-by-second Bitcoin trading on Binance and separate the cycles into two groups.
A cycle is still even when Polymarket’s implied probability remains between 40% and 60% during its final minute. These are the contracts where a small movement in the underlying asset could still change the result. Every other cycle is already effectively decided.
That split gives us a clean test. Ordinary Bitcoin volatility should not care whether a Polymarket contract is still competitive. Settlement-driven trading should:
- appear primarily when the contract remains close to 50/50;
- concentrate in the final seconds before settlement;
- fade or reverse once the contract has resolved; and
- become less common as the contract horizon lengthens.
The burst appears only when the outcome is still live.
The first signal is timing. In still-even cycles, Binance spot volume stays close to baseline through most of the window, then rises sharply in the final ten seconds before settlement. The spike is roughly 12× normal in the 5-minute market and roughly 17× normal in the 15-minute market. Already-decided cycles do not show the same closing burst; they remain much closer to their normal level.
The stale price trend over 5 minutes is why a small late move can decide the payout.
Why can a few seconds of trading matter at all? Because Bitcoin frequently ends a five-minute window very close to where it began.
Across the sample, 56% of 5-minute cycles finish within 0.05% of their opening price, and 81% finish within 0.1%. That does not mean the settlement benchmark is easy to control. It means the contract often resolves on a difference of only a few basis points. A movement that would be ordinary noise over a longer horizon can decide the entire payout over five minutes.
A longer clock makes the exposure rarer.
That narrow margin is what creates the exposure. In the 5-minute market, many cycles finish close enough to the opening price that a small late move could still matter. Longer horizons make that less common because Bitcoin has more time to drift away from its starting level before settlement.
That shows up directly in the still-even sample. About 2.9% of 5-minute cycles remain still-even in the final minute, compared with 0.9% of 15-minute cycles and none of the observed 4-hour cycles. The longer the clock runs, the less often the contract reaches settlement balanced on a final tick.
| Contract | May cycles | Still even | Share still-even |
|---|---|---|---|
| 5-minute | 8,854 | 261 | 2.9% |
| 15-minute | 2,873 | 27 | 0.9% |
| 4-hour | 60 | 0 | 0.0% |
| Selected contracts | 11,787 | 288 | 2.4% |
The same point shows up in the underlying price path. The typical Binance open-to-close move rises from roughly 6 bps over five minutes to roughly 43 bps over four hours. As that move grows, the outcome is more likely to be decided before the closing seconds.
The price moves into settlement, then reverses.
The remaining question is whether the underlying price also moves in the direction that would matter for settlement. It does. In still-even cycles, the Binance price moves into the close and then partially reverses after resolution.
The reversal is important because it separates a generic closing move from a settlement-timed footprint. If the final-second move were just ordinary volatility, there would be no reason for it to fade after the Polymarket contract resolves.
Taken together, the evidence is consistent with settlement-driven trading in the underlying Bitcoin market. The key feature is not merely that volume and price move near the close. It is that the movement arrives on Polymarket’s settlement clock, appears primarily when the outcome remains live, partially reverses after resolution, and becomes rarer as the contract horizon lengthens.
The connection to the bot section is an incentive story, not an identity claim. Ultra-short contracts create a cross-market opportunity that rewards continuous monitoring and fast execution. That may help explain why bot-like wallets dominate the product alongside more conventional incentives such as market making, rebates, and arbitrage.
That attribution carries the same caveat as the wallet screen itself. The bot-like, sophisticated, and retail-labeled tags are behavioral inferences drawn from on-chain trading patterns (ticket size, cadence, and market breadth), not venue-confirmed identities. Pointing the settlement footprint at that bot-like layer describes how the wallets trade; it is not proof of who controls them, nor that any specific account traded to move the benchmark.
That makes settlement exposure a function of market structure. The same short clock that creates more trading cycles, more fee-paying turnover, and more opportunities for automation also leaves more contracts balanced on a narrow final price movement. Longer horizons and harder-to-move settlement benchmarks reduce that exposure.
Bot-like turnover and settlement-sensitive trading can keep a market highly active without building a persistent wallet base over the observed window. Retention tests whether those wallets actually come back.
So how much of this activity actually sticks?
The trading sticks longer than the wallets.
Wallet retention improves as contract horizons lengthen. By week four, only 20.5% of 5-minute wallets are still trading the product, compared with 48.4% for the weekly market. The 1-hour and daily products fall in between, while 5-minute retention declines to roughly 8.5% by week twelve.
The acquisition funnel is shrinking at the same time. The 5-minute product’s weekly new-wallet cohort fell 77.9%, from 23.3K wallets on February 9 to 5.1K on June 22. Retention did not improve enough to offset that decline: pooled week-one retention is 52.9% for 5-minute wallets and 20.5% remain after four weeks, compared with 80.8% week-one retention for the weekly product.
Headcount only captures one version of retention. For the 5-minute product, we also measure how much activity each cohort brings back. Trade retention is the number of trades a cohort generates in a later week, measured as a share of that same cohort’s first-week trades. Dollar retention does the same for traded dollars. In other words, this does not mean the same trades persist. It means the same cohort continues to generate activity.
That distinction matters. After one week, only 52.9% of wallets remain, but those wallets generate roughly 91% of the cohort’s first-week trades and 75% of its first-week dollars. By week four, wallet retention falls to 20.5%, while trade retention is still roughly 39% and dollar retention roughly 30%. By week twelve, the activity advantage fades too, with trade and dollar retention falling into the high-single digits.
That is the retention paradox. Polymarket retains its high-frequency activity engine far better than it retains wallets. The product does not need most wallets to return in order to preserve a meaningful share of trades and dollars, because the wallets that stay are the ones responsible for much of the activity. But that also means headline volume is not evidence of a broad, durable user base.
What we could not answer, and what comes next.
Three questions remain.
Who is trading Kalshi’s crypto markets? The Kalshi trade data used here do not include a persistent user ID or wallet address, so trades cannot be linked across time or attributed to individual participants. That prevents the account-level analysis used for Polymarket, including concentration, bot-like behavior, and repeat trading. Kalshi’s larger tickets and longer contracts look different in aggregate, but the data cannot tell us who is generating that activity.
How durable is Polymarket’s bot-driven core? A minority of bot-like wallets accounts for most 5-minute trading, even as new-wallet cohorts shrink. The open question is whether that automated core remains stable as fees, rebates, and market structure change, and whether the product can broaden beyond it.
Can the settlement footprint be tied to specific positions? Binance trading repeatedly intensifies around Polymarket settlement windows when the outcome remains in play, but the more important forward-looking question is whether the pattern can persist now that it is visible. If the trading is settlement-driven, it would require taking spot-market risk near the settlement threshold, then unwinding immediately after the market closes. Once other traders know to watch that window, they can sell ahead of the expected unwind or lean against it, eroding any edge. The natural next test is whether the footprint fades, migrates, or persists now that it has been described publicly.
Acknowledgment
Thanks to Ruizhe Jia and coauthors for their original research on settlement manipulation in prediction markets and its presentation at the Edge City crypto research camp hosted by Uniswap Foundation, and to James Dai from Surf and Ping Chen, Raymond Yu, and Darren Carter from Pantera Capital for the review.