The 'MEV Moment' in Market Predictions: Betting on Ups and Downs or Creating Them

Polymarket and Kalshi have taken different product paths.

Written by: Ally , Danning , from Pantera

Key Findings

This analysis covers the period from January to June 2026, including on-chain transaction data from Polymarket, trading data from Kalshi, and Binance's second-level spot data used for settlement research. Incomplete periods are noted in the corresponding charts. The trading volume data in this article is defined as taker volume rather than nominal volume.

• Short-term crypto contract trading reflects real demand, with Kalshi's monthly taker volume surpassing Polymarket. Of the combined $10.07 billion in crypto asset price prediction market (taker) volume across both platforms, approximately $7.8 billion comes from contracts measured in minutes or hours. In the first half of 2026, Polymarket recorded $5.59 billion, while Kalshi reached $4.48 billion—however, Kalshi's crypto share is increasing every month.
• Polymarket and Kalshi have taken different product paths. Both platforms primarily trade BTC, but Polymarket focuses on 5-minute markets that allow small bets, while Kalshi's shortest contracts are 15 minutes.
• High trading volume does not equate to high fee revenue. Polymarket's crypto transaction volume is about 1.25 times that of Kalshi, but the measured fees over six months were only about $53.9 million. Based on fee rates, Kalshi's modeled average fee rate is approximately 2.8 times that of Polymarket's measured rate (2.74% vs. 0.96%). Note that this is a rate difference and is separate from the earlier 3.6 times share change; this comparison serves only directional reference—comparing Kalshi's modeled fees with Polymarket's measured fees is not a like-for-like revenue comparison.
• Polymarket's 5-minute market is dominated by bots and shows signs of manipulation. Suspected bot wallets contributed approximately 86% of the taker transaction volume and appeared on both buy and sell sides in about 79% of transactions. This classification is based on trading behavior and does not confirm any specific wallet being controlled by a program.
• As revealed by a Stanford University study, the surge in Binance's trading volume coincides repeatedly with Polymarket's settlement timing. During periods when outcomes are uncertain (i.e., when BTC prices are close to the opening price), Binance's spot trading volume amplifies by about 12-17 times in the last few seconds, with prices moving towards the settlement direction before partially retracing post-settlement.
• Trading activity is more 'sticky' than users. In Polymarket's 5-minute market, wallet retention declines rapidly: by the fourth week, only about 20% of wallets are still trading. However, the remaining traders show a much slower decline in activity—retaining about 39% of transaction counts and 30% of transaction amounts during the same period.

Short-Term Crypto Market

The short-term crypto market is rising, but Kalshi and Polymarket are offering two different products

Both Kalshi and Polymarket are expanding the boundaries of prediction markets beyond traditional event contracts, with the crypto segment being the most typical example of this shift. Sports remain the fundamental base of the entire industry—sports account for about 79% of trading volume on Kalshi and are also the largest category on Polymarket, accounting for about 48%. However, the crypto market is evolving into a completely different type of prediction market product: a rolling open short-term market that allows traders to make binary bets on price movements within a defined time window.

This structural difference is crucial. The behavioral logic of these contracts is entirely different from the slow-paced event markets typically associated with elections, policies, or corporate events; they are neither perpetual contracts nor options, but binary markets that settle within minutes to hours through exchange matching—more akin to a short-term trading product than a prediction tool.

During the statistical window, approximately $7.8 billion of the $10.07 billion in crypto transactions across both platforms came from minute-level and hour-level contracts; Polymarket's 5-minute product reached a market of $2.1 billion in less than four months since its launch. The question of whether there is demand for this category is no longer in doubt.

The real question is: how do different design choices shape different market structures—user behavior, economic models, and trading strategies? Kalshi and Polymarket face similar demands but have made different choices regarding contract cycles, asset coverage, trading rhythms, and fee structures. These choices determine what the transactions look like, who is trading, and who is profiting.

Competitive Landscape

Polymarket leads in the crypto market, while Kalshi is closing in

From January to June 2026, Polymarket's crypto transactions totaled $5.6 billion, nearly double Kalshi's ($4.48 billion). However, when aggregating all categories, Kalshi has the larger total transaction volume. The difference arises from category structure: Kalshi's transactions are overwhelmingly dominated by sports, while Polymarket's activities are more evenly distributed among sports, crypto, and politics.

Kalshi: Sports-focused, with the crypto segment steadily growing

Polymarket: Larger crypto market, more balanced category structure

The summary row for '4+' in the table is derived from the proportion of amounts; these categories do not have separate transaction count data.

By expanding the category structure weekly, we can observe the dynamic evolution of this comparison: Kalshi remains a sports-prioritized platform, but crypto trading has grown from a narrow band to an increasingly prominent part of Kalshi; Polymarket's weekly crypto transactions fell back after peaking in March, while its overall category structure has always been more balanced.

Weekly category structure of the two platforms: Kalshi (top) is dominated by sports, with crypto steadily rising; Polymarket (bottom) has a balanced structure, with crypto falling back after peaking in March

Thus, this competition is more variable than cumulative totals suggest. Polymarket's crypto market remains larger, but Kalshi is steadily increasing the weight of crypto on its platform: its crypto share has risen monthly from 4.3% in January to 17.9% in June; Polymarket peaked in February (34.0%) and fell back to 19.8% by June.

Two opposing curves: Kalshi's crypto share rises monthly, while Polymarket falls back after peaking in February

Throughout the window period, Polymarket wins on scale; Kalshi wins on internal share momentum. The next competition will not be about volume, but about structure: what each platform has made of this product.

Product Path Divergence

As time shortens, the same bet transforms into a different market

Divergence begins with trading rhythm. Polymarket's 5-minute contracts have found the strongest demand at the shortest end of the market—shorter holding times reduce the time risk for traders, and extremely short-term outcomes are easier to model. Kalshi, on the other hand, starts at 15 minutes, focusing on 15-minute and hourly BTC contracts. Both sell short-term crypto contracts, but Polymarket has pushed this format to the limits of almost continuous trading.

Asset coverage further widens the gap. Kalshi has almost only BTC; Polymarket also centers on BTC but has significant second-tier assets like ETH, SOL, and XRP.

Weekly crypto transactions split by contract duration: short-term contracts are the main battlefield for both sides

Weekly crypto transactions split by asset: Kalshi (top) is almost entirely BTC, while Polymarket (bottom) has a clear multi-asset layer

This difference is directly reflected in the transaction data: Polymarket has many more transactions with smaller individual amounts; Kalshi has fewer transactions with larger individual amounts. Looking at the distribution reveals the real differences beneath the total volume.

Kalshi's typical single bet is larger: the median is about $4.70, approximately 1.5 times that of Polymarket ($3.05); the middle 50% of transactions fall between $0.96 and $19.79, while Polymarket ranges from $1.14 to $6.96. The upper tail of the distribution also shows Kalshi leading: the 95th percentile is about $123 compared to $33.51, the 99th percentile is $385 compared to $124, and the 99.9th percentile is about $1,100 compared to $967.

Distribution of single transaction amounts in total crypto transactions (logarithmic scale): both sides have small transactions, but Kalshi is larger from the median to the tail

Looking at the total transactions, both platforms have small single bets (median $3-$5), but Kalshi's transactions are spread across a wide range of $1-$100, while Polymarket clusters around a few dollars: about four-fifths of Polymarket transactions are small orders in the 5-minute and 15-minute markets. Occasionally, there are large transactions on Polymarket, but they have now been pushed beyond the 99.9th percentile, drowned in a flood of small orders, and do not even constitute a visible long tail.

Thus, the fee structures of the two platforms are completely different: Kalshi handles more stable medium-sized orders, while Polymarket deals with a large number of small orders mixed with occasional large ones. Larger transaction volumes do not necessarily lead to higher fees—the outcome depends on how the platform prices these transactions and how the fee structure interacts with the single transaction amounts and contract prices.

Platform Revenue

Larger Transaction Volume ≠ Higher Fees

During the window period, Polymarket's crypto transactions totaled $5.59 billion, about 1.25 times that of Kalshi ($4.48 billion), but the measured fees per transaction were only about $53.9 million. Kalshi does not disclose actual fee revenue, but modeling based on its publicly available fee schedule (as a lower limit assumption) yields a modeled estimate of $122.1 million—higher instead.

It should be noted that Polymarket's total fee amount is the figure before daily rebates are deducted, and its fees are phased in rather than implemented all at once in March—on January 5, the 15-minute crypto market began charging fees; on February 12, the 5-minute market went live and started charging fees; on March 6, fees expanded to all newly launched crypto markets; and on March 30, Fee Structure V2 extended fees to most categories. Kalshi's figures are modeled based on the fee schedule starting from February 5. The fee windows corresponding to the two totals are not consistent. Looking only at those weeks when fees were effective: Polymarket's average comprehensive fee rate was 0.96%, while Kalshi's was 2.74%—about 2.8 times that of the former.

After Polymarket's Rule Adjustments, Market Maker Transaction Volume Shrinks Significantly

On March 30, Polymarket expanded fees to more categories; on April 28, it migrated trading to a new exchange system. This switch changed the handling of fees, margins, order signatures, and various external integrations, and cleared the order book—market makers had to reconnect and rebuild their quotes.

Crypto weekly transaction volumes peaked in March and have been declining since April, while fees have continued to accrue before and after the migration. The total of $53.9 million spans the entire period: fees before the switch correspond to the old trading rebate mechanism, while those after are calculated according to the new system.

Polymarket weekly crypto transaction volume and measured fees: transactions peaked in March and declined, while fees continuously rose between the two nodes of fee reform and migration

By breaking down the same weekly transaction volume according to the scale of market makers before the migration, we can see where the decline comes from: the largest market maker group shrinks the most—after the fee expansion in March, it dropped to the next level, and again dropped to the next level before and after the system switch on April 28.

The timing indicates two independent pressures: the fee expansion may have changed the economics of the flow of quotes from market makers; the system migration represents an operational reset—related integrations need to be migrated again, existing orders were cleared, and quotes must be rebuilt. Changes in the execution protection mechanism may also have made some delay-sensitive strategies less attractive.

Weekly market maker transaction volume broken down by market maker scale: the larger the tier, the sharper the pullback, with two dashed lines indicating two levels of decline at the fee reform announcement and V2 launch

Market maker tiers (Mega--Micro) are self-defined: based on each market maker's transaction volume and number of transactions before V2 (up to April 28), and remain unchanged thereafter; "active market makers only" refers to those excluding weeks with zero transactions. This is not the official tiering of the platform.

This temporal overlap is indicative, not conclusive. These market maker tiers add up to the weekly transaction volume itself, so this breakdown cannot independently prove that platform changes drove away market makers. It presents a more fundamental pattern: the larger the market maker before migration, the more severe the contraction, and the decline accelerates significantly before and after the two changes. This indicates that the market makers with the largest transaction volumes are most sensitive to platform changes—the most likely reason is that the new structure worsened their fee incentives.

Short-Cycle Markets Are Fee Engines

Fees are concentrated on the same batch of dominant transaction products: on Polymarket, the 5-minute and 15-minute contracts contributed about 90% of the measured crypto fees; on Kalshi, contracts of 15 minutes or less, plus hourly contracts, accounted for about 94% of the modeled fee estimates.

Polymarket measured fees broken down by contract duration: 5-minute and 15-minute contracts contribute about 90%

Kalshi modeled fees broken down by contract duration: contracts of 15 minutes or less and hourly contracts account for about 94%

The fastest contracts drive both activity and support fees. However, fees and market maker data only explain part of the market structure. The next question is: who is behind these short-term flows?

Trader Profiles

Polymarket 5-Minute Market: 25% of Wallets Contribute Over 85% of Transaction Volume

We further examine the user composition of these short-cycle markets. Since Kalshi's public data does not contain user-level IDs, the following analysis is only for Polymarket.

Polymarket's 5-minute market generates significant settlement transaction volume, but this flow is unevenly distributed among wallets. The first question to answer is concentration: is the activity dispersed among a large number of occasional betting retail investors, or concentrated among accounts that trade repeatedly, systematically, and across many markets?

We screened all wallets in the 5-minute market with at least 25 transactions. The screening rules were deliberately kept simple, focusing on behaviors that are difficult to explain through manual trading: extremely high daily transaction counts, highly repetitive single transaction amounts, or activities spanning thousands of markets. This is not a trained classifier and cannot prove intent—it is merely a behavioral label. Wallets meeting any of the following conditions are marked as suspected bots (bot-like):

• A single active day has more than 100 transactions;
• There are over 500 transactions with nearly constant amounts (defined as the 95th percentile single transaction amount falling within ±15% of the median);
• Trading has occurred in over 1,000 different markets.

Wallets that do not trigger the above rules are further divided into "sophisticated traders" and "retail". Sophisticated traders are characterized by "large and accurate bets": either maintaining an average single transaction of over $200-$500 in relatively few markets, or accumulating over $50,000 in transactions across no more than 30 markets. The remaining active wallets are marked as retail.

Classification Method Explanation

Suspected bot behavior is screened first; sophisticated wallet addresses are large and concentrated; retail is what remains after passing the transaction threshold of active 5-minute wallets.

The statistical subjects are 5-minute wallets with over 25 transactions. The rules are checked in order: wallets that meet both bot and sophisticated standards are counted in the bot category. The near-constant single transaction rule uses the 95th percentile single transaction amount falling within ±15% of the median. The 25-transaction threshold excludes one-time retail investors, so the actual proportion of retail wallets across the platform will be higher, and their transaction volume proportion will be lower.

By this measure, the 5-minute market is highly concentrated: suspected bot wallets are a minority among active wallets, yet they account for the vast majority of transaction counts and transaction amounts.

A Quarter of Wallets Contribute to Nearly 90% of Transactions: Analysis of Wallet Behavior by Transaction Count and Volume

This behavior is evident from any perspective. By transaction amount: wallets with a median transaction amount of no more than $5 account for 90.4% of active wallets and generate 96.0% of transactions. By frequency: wallets with a median transaction interval of no more than 1 second account for 84.0% of the 5-minute transaction volume. By concentration: the top 1,000 wallets account for 43.3% of the transaction volume, while the top 10,000 account for 73.1%.

Small Transactions Support the Entire Market: Transaction Count and Volume by Median Transaction Amount per Wallet

Wallets Placing Orders in Seconds Dominate Transactions: Wallets with Median Intervals ≤ 1 Second Account for 84% of Transaction Volume

Not Just One Whale, but an Entire Ecosystem: Cumulative Transaction Volume Proportion of Top Wallets

Looking at both sides of each transaction. Both makers and takers are worth examining—one wallet can provide liquidity in one transaction and actively take orders in the next, so the maker/taker identity is not fixed for a wallet. In the Polymarket transaction data, we define the maker as the wallet that has an order resting on the book, and the taker as the wallet that crosses the spread to facilitate the transaction; then we classify each wallet using the same behavioral rules and count which category the maker and taker belong to in each transaction.

The result is: in the 5-minute market, 78.9% of the transaction volume involves suspected bot wallets on both sides; transactions without bots on either side account for only 1.4%.

Bot Counterparties are Also Bots: 5-Minute Transaction Volume Split by Trader Type on Both Sides

Human Participation Increases as Time Periods Lengthen: Comparison of Bot Participation Across Different Duration Products

The Same Behavioral Characteristics are Imprinted Across the Entire Short Cycle Ladder: Proportions of Second-Level Orders and Small Transactions in Various Duration Products

For a product that resets every five minutes, is highly repetitive, has market-making incentives, and trades assets priced based on continuous trading in other venues, bots are natural players. This product rewards speed, monitoring, and execution, rather than subjective judgment. When examining wallet screening and the structure of counterparties on both sides together, the conclusion points to the same place: Polymarket's 5-minute crypto market has a solid core of bot trading, with only a thin layer of retail traders on the outside.

However, this also leaves room for automated strategies—even more manipulative tactics. Suspected bot wallets will not just quietly rest on the order book; when speed is needed, they will also actively take orders, including in the last few seconds before contract settlement. Based on a recent study published by Stanford University, we tested and cross-validated Polymarket's ultra-short 5-minute market and found signs consistent with manipulation.

Market Manipulation

When Polymarket's Settlement Results Can Still Be Reversed, Bitcoin Trading on Binance Surges at Closing Time

According to market rules, Polymarket's 5-minute and 15-minute Bitcoin markets determine outcomes as follows: the Chainlink BTC/USD reading at the end of the contract window must not be lower than the reading at the beginning of the window. Since this benchmark comes from the tradable crypto market price, traders can hold a position in Polymarket while trading Bitcoin elsewhere—especially in the venue that provides data for settlement pricing, where prices can be influenced.

Chainlink's BTC/USD Data Stream aggregates prices from multiple trading platforms, and the specific data sources have not been disclosed. This article uses Binance's BTC/USDT spot price as an observational proxy for market prices, rather than viewing it as the sole input source for the Chainlink oracle. The same trading strategy also applies to the downward (Down) direction.

A simple example illustrates this opportunity. Suppose Bitcoin opens a five-minute window at $100,000, and in the last few seconds, the Polymarket contract is hovering around 50/50. A trader holding a "buy" position can simultaneously buy Bitcoin in the spot market. This buy order could be a hedge or an arbitrage between the two markets; however, if this buying pressure is sufficient to push Bitcoin to make a significant move, it could also raise the closing benchmark above the opening price—directly rewriting which side receives the payout. The aforementioned paper mathematically models the risks and rewards of this strategy and provides detailed data evidence.

Thus, the same spot transaction could be a hedge, arbitrage, or rebalancing, but it could also intentionally influence the price—trading data cannot distinguish the motive. What it can answer is another question: does the behavior of the underlying market exhibit the characteristics of "settlement-driven trading"? Most of this activity may just be ordinary hedging, arbitrage, or rebalancing; but in principle, sufficiently large movements in the underlying Bitcoin market could indeed affect the settlement benchmark, determining which side of the Polymarket contract wins.

Our analysis aims to see if this cross-market incentive leaves repeatable footprints. We use second-by-second Binance spot data as a proxy variable for underlying Bitcoin market activity (note: Binance is not the settlement data source for Polymarket). If the trading is aimed at settlement, it should appear in sync with Polymarket's clock: concentrated in the last few seconds, primarily occurring when the outcome can still be reversed, and dissipating after settlement.

Methodology and Attribution Statement

This section builds directly on the research framework proposed by Stanford University's David Dai, Ruizhe Jia, and Singapore Management University's Shihao Yu.

Their paper "Settlement Manipulation in Prediction Markets" (2026) provides the core framework, the condition screening method for "still-even periods," and the testing approach of "cycle length as defense." This research formalizes the hypothesis of settlement manipulation for the first time and provides an empirical framework for studying it. We did not simply replicate their analysis but expanded it using independent datasets, more microstructure indicators, and multiple new empirical charts based on our May 2026 Polymarket and Binance data.

Their core screening question is straightforward: in the underlying market of the settlement benchmark, can a feasible transaction move this benchmark before contract settlement? Short contracts are more exposed—they often have very small price differences at closing; long contracts have far fewer opportunities because the underlying asset has usually moved away from the starting point, and the outcome is already determined.

Based on Binance's 1-second spot data from May 2026 and Polymarket's transaction prices, covering 5-minute (approximately 8,900 cycles) and 15-minute (approximately 2,900 cycles) BTC up/down markets; drift charts use BTC/USDT data from December 2025 to May 2026. "Still-even" refers to when the implied probability of the contract is between 0.40 and 0.60 in the last minute; the sample size for the 15-minute still-even cases is small (n≈27). The observed patterns are consistent with settlement-driven trading, but these are statistical footprints and do not constitute accusations against any specific account. Complete methodology and charts can be found in the accompanying study on settlement integrity.

Testing Design: Comparing Contracts That Can Still Be Reversed with Those That Are Already Determined

Short-cycle contracts run one cycle (cycle) each time: a five-minute bet ends, and the next begins. We align each Polymarket cycle with Binance's second-by-second Bitcoin transactions and then divide the cycles into two groups.

If the implied probability in Polymarket remains between 40% and 60% in the last minute, this cycle is still-even—small fluctuations in the underlying asset could still rewrite the outcome. The remaining cycles are effectively determined.

This division gives us a clean test. Ordinary Bitcoin fluctuations should not care whether a Polymarket contract still has suspense; however, settlement-driven trading should:

• Occur mainly when the contract is still close to 50/50;
• Concentrate in the last few seconds before settlement;
• Fade or reverse after contract settlement; and
• Become less common as the contract cycle lengthens.

Transaction Volume Surges, Only Occurring When the Outcome is Undetermined

The first signal is the timing. During periods of uncertainty, Binance's spot trading volume remains close to the baseline for most of the window, then surges dramatically in the last ten seconds before settlement: the 5-minute market is about 12 times the normal level, and the 15-minute market is about 17 times. Established periods do not exhibit this closing pulse, and the trading volume remains relatively normal.

Binance BTC spot trading volume per second during the period: the uncertain period (red line) surges in the last few seconds before settlement, while the established period stays close to the baseline

What enables manipulation is the almost static price in five minutes

Why can a few seconds of trading have an impact? Because Bitcoin often returns almost to its starting point at the end of the five-minute window.

In the sample, 56% of the 5-minute closing prices differ from the opening prices by less than 0.05%, and 81% differ by less than 0.1%. This does not mean that the settlement benchmark is easy to manipulate, but rather that these contracts often determine outcomes based on a few basis points of difference—fluctuations that would only be noise over a longer period can decide the entire payout within five minutes.

The shorter the contract, the closer the closing price is to the opening price: distribution of BTC price fluctuations by "opening - closing deviation" across various durations

The longer the clock, the fewer opportunities

It is this extremely narrow price difference that creates opportunities. In the 5-minute market, many closing prices are so close to the opening prices that a small late movement can change the outcome; as the period extends, Bitcoin has more time to drift away from the starting point, and such situations become less frequent.

The samples that remain undecided directly confirm this: about 2.9% of the 5-minute periods remain undecided in the last minute, while only 0.9% of the 15-minute periods do, and none of the observed 4-hour periods remain undecided. The longer the clock runs, the lower the probability that the contract hangs on the last jump.

The underlying price path follows the same rule: the typical fluctuation range from opening to closing on Binance rises from about 6 basis points in five minutes to about 43 basis points in four hours. The greater the fluctuation, the more likely the result has already been settled a few seconds before closing.

The longer the window, the further the price drifts: typical fluctuations (basis points) from opening to closing for Bitcoin across various contract durations

Prices surge towards settlement, then retract

The final question: does the underlying price also move in a direction meaningful for settlement? Yes. During periods of uncertainty, Binance's price moves towards the closing direction and partially retracts after settlement.

Price fluctuation intensity per second during the period: the uncertain period experiences the most intense fluctuations in the last few seconds before settlement

This retraction is crucial—it distinguishes ordinary closing fluctuations from those that occur right at the settlement point. If the changes in the last few seconds are merely typical fluctuations, there is no reason for them to dissipate after the Polymarket contract settles.

Price surges, settles, then falls back: Binance BTC price path aligned with the late direction before and after the settlement point

Putting all the evidence together, they align with the hypothesis that "there are settlement-driven trades in the underlying Bitcoin market." The key is not that the volume and price happen to fluctuate near the closing, but that these changes occur at the time of Polymarket's settlement, mainly during periods of uncertainty, partially retract after settlement, and become less frequent as the contract period extends.

This section's connection to the previous chapter on bots is an incentive logic, not an identity identification. Ultra-short contracts create a reward for around-the-clock monitoring and rapid execution across markets—this may explain why suspected bot wallets dominate this product, alongside conventional incentives like market making, rebates, and arbitrage.

This attribution and wallet screening carry the same limitations: labels like suspected bots, seasoned traders, and retail investors are inferred from on-chain trading patterns (single transaction amounts, rhythm, market breadth), not confirmed identities by the platform. Pointing the settlement footprint towards this automated group describes how these wallets trade; it neither proves who is behind the wallets nor confirms that any specific account trades to move the benchmark.

This turns settlement integrity into a product design issue. The same short clock creates more trading periods, more paid turnover, and more automation space, while also leaving more contracts hanging on extremely narrow final price differences. Longer periods and more stable settlement benchmarks can reduce this exposure.

Bot-driven turnover and settlement-sensitive trading can make a market appear vibrant, but may not necessarily establish a sustainable user base. Will these wallets return? Retention data provides the answer.

User Retention

How much of this excitement can stick around?

Trading lasts longer than wallets

Wallet retention improves as contract periods lengthen. By the fourth week, only 20.5% of wallets trading 5-minute products remain, while weekly markets have 48.4%; 1-hour and daily products are in between; by the twelfth week, retention for 5-minute products drops to about 8.5%.

The shorter the contract, the faster the wallet loss: retention curves for wallets based on "weeks since first trade" across various durations

Meanwhile, the customer acquisition funnel is also narrowing. The weekly new wallets for 5-minute products dropped from 23.3K in the week of February 9 to 5.1K in the week of June 22, a decrease of 77.9%. The improvement in retention is far from enough to offset this decline: only 52.9% of 5-minute wallets return after a week, and 20.5% remain after four weeks; while the first-week retention for weekly products is as high as 80.8%.

The funnel for 5-minute products is shrinking: weekly new wallets (bars) and retention of wallets across periods (lines)

15-minute products: similarly shrinking funnel

1-hour products: smaller scale, same pattern

4-hour products: a slender funnel

Weekly products: a control group with human rhythm, significantly higher first-week retention

But counting heads only reveals one side of retention. For 5-minute products, we also measured how much activity each cohort brings back: trading retention refers to the proportion of transactions generated by a cohort in subsequent weeks compared to its first-week transaction count; monetary retention does the same for transaction amounts. In other words, this does not mean the same trades are still happening, but that the same group of people continues to generate activity.

This distinction is important. After one week, only 52.9% of wallets are still active, but these wallets account for about 91% of the transaction count and 75% of the transaction amount for that cohort's first week; by the fourth week, wallet retention drops to 20.5%, but trading retention remains around 39%, and monetary retention is about 30%; it is not until the twelfth week that the advantage in activity is exhausted, and both trading and monetary retention fall into the high single digits.

Wallets are declining, but trading continues: retention curves for wallets, transaction counts, and transaction amounts for the same cohort of 5-minute products

This is the retention paradox: Polymarket retains its high-frequency activity engine far better than it retains wallets. This product does not need most wallets to return to maintain a significant share of transactions and amounts—because the wallets that remain are precisely the ones generating most of the activity. Conversely, this also means that apparent transaction volume does not prove the existence of a broad and lasting user base.

Unresolved Issues and Future Work

Unanswered Questions and Next Steps

Three Questions Remain

1. Who is trading on Kalshi's crypto market? Kalshi's trading data lacks persistent user IDs or wallet addresses, making it impossible to associate trades across time or to individuals. This prevents us from conducting account-level analyses on Kalshi, such as concentration, bot behavior, and repeat purchases. Kalshi's larger single trades and longer contracts present a different picture in aggregate, but the data does not reveal who is behind these activities.
2. How long can Polymarket's bot core sustain itself? A few suspected bot wallets account for the vast majority of five-minute trades, while the entry of new wallets continues to decline. The unresolved question is: when fees, rebates, and market structure change, will this automated core remain stable? Can this product break through and reach a broader audience?
3. Can settlement footprints be attributed to specific positions? When outcomes are still uncertain, Binance trading has repeatedly surged near Polymarket's settlement windows; however, the more pressing forward-looking question is: now that the pattern has been exposed, can this strategy continue? This trade appears to require bearing spot risk near the settlement threshold and immediately closing after the market closes. Once other traders know to watch this window, they can sell before the expected close or simply do the opposite, eroding this advantage. The natural next step to test is: after public exposure, will this footprint disappear, shift, or still remain profitable?

Acknowledgments: Thanks to Ruizhe Jia and collaborators for their original research on market manipulation, and for their presentation at the Edge City crypto research camp hosted by the Uniswap Foundation; thanks to Surf's James Dai, Pantera Capital's Ping Chen, Raymond Yu, and Darren Carter for their participation in the review.

Data Source: Surf AI