70% of the profits from the head wallet come from bots, but AI has not yet taken over the prediction market

Author: Jeff, IOSG

Abstract

The core data predicting market bot panic is quite intuitive: 5% of wallets on Polymarket that appear to be bots contribute 75% of the platform's trading volume. 823 wallets have each netted over $100,000 since January 2025, collectively withdrawing $131 million in profits from Polymarket. Among the top 20 wallets by profit, 14 are classified as bots (Stacy Muur leaderboard inspection). A study from the University of Toronto (covering 2.4 million users and $67 billion in trading volume since 2022) found that 68.8% of users are in a loss position, while the top 1% of users captured 76.5% of all profits.

The narrative derived from this is: prediction markets are a wealth transfer machine, and bots are its operators. The data is accurate, but the framework has a half bias. Core Points 1. The core flaw in the bot narrative is equating "trading volume concentration" with "capital plunder." The 5% of wallets on Polymarket contributing 75% of trading volume only indicates the distribution of account activity and cannot directly conclude that retail funds are being extracted by bots.

2. Data at the group level is more persuasive. AI agent wallets have a return rate of about 37%, while human wallets only range from 7% to 13%. The 3-4 times group-level disparity is real evidence of structural advantage; the fact that 14 out of the top 20 wallets by profit are bots (Stacy Muur leaderboard inspection) belongs to the right tail projection of that distribution and is not independent evidence.

3. The advantage of bots lies in structural dimensions, not judgment dimensions. The three types of markets dominated by bots—price delay arbitrage, real-time sports game state automation, and cross-platform combination arbitrage—share the commonality of not requiring subjective judgments about real-world events. Once market outcomes depend on the comprehensive processing of multi-source information, the advantages of bots systematically weaken.

4. The category structure of Polymarket has shifted from "politics 42%" to "sports 50%" in the past 12 months, with the fastest-growing category being long-cycle event markets, where bots do not have a structural advantage, indicating a clear trend towards retail participation on the platform.

5. Forward-looking judgment: The proportion of bots will continue to rise as deployment costs decrease, but the scale of capital extraction from humans by bots will peak before the bot proportion does—because the mutual erosion speed among bots is faster than their erosion speed on human accounts.

6. Investment strategy: Equity opportunities at the platform level (Kalshi + Polymarket combined 97%+ share) are basically closed; value opportunities are migrating to the L2 agent infrastructure layer (Olas / Valory model) and the venue-agnostic middle layer, while C-end bot products and L3 data/pricing layers do not fit venture.

1. Market Size Exceeds Bot Panic

Three quantitative anchors define the scope of this report's discussion.

First, Bernstein revised the prediction market sector size for 2026E to $240 billion on April 14, 2026, and the path to $1 trillion by 2030 has become a consensus on the sell-side.

Second, the combined YTD trading volume of Kalshi and Polymarket broke $60 billion in mid-April 2026, exceeding the total of $51 billion for the entire year of 2025.

Third, Robinhood launched over 1,000 Kalshi contracts, with the platform's 1 million+ customers cumulatively trading 9 billion contracts. Robinhood's prediction market business has an ARR of about $350 million, $150 million for the full year of 2025, and an estimated $586 million for 2026, making it the fastest-growing product line of the company.

These data points collectively point to one conclusion: prediction markets are no longer a single crypto-native sector; their attributes are closer to a TradFi distribution problem. The "retail being plundered" group assumed in the bot narrative is not crypto users, but retail investors entering through traditional brokerage channels.

This leads to the contextual bias of bot panic: the sector is not automated to extract value, but rather mainstream finance injects traffic at a pace far exceeding any automated extraction speed.

2. The Truly Important Data: 37% vs. 10%

The most cited data point in the bot narrative has sample selection bias.

The source data for "14 bots in the top 20 by profit" is predicated on a small sample sorted by profitability. This sample can only reflect the occupancy of bots in the right tail of the distribution and cannot be used to infer the superiority or inferiority at the group level.

Group-level data (source: Polystrat / Valory disclosure, cross-verified with on-chain analysis data from multiple Polymarket sources):

The 3-4 times difference in win rates at the group level is a true reflection of the structural advantage of bots. The statistic of 14/20 in the profit leaderboard should be understood as downstream performance of that win rate distribution, rather than independent causal evidence.

3. In Which Markets Do Bots Succeed

The extraction scale of bots is highly concentrated in the following three types of markets. The commonality among these three is that they do not require subjective judgments about real-world outcomes, relying instead on delays or pricing advantages related to the platform's matching engine. Price Feed Delay Arbitrage Representative case: Wallet 0x8dxd, which turned $313 into $437,600 in January 2026, only trading 15-minute BTC rise and fall contracts, with a win rate of 98%.

Strategy principle: Monitor the spot prices on Binance and Coinbase, and build positions when Polymarket quotes lag behind CEX. Polymarket introduced a taker fee for 15-minute crypto contracts on January 7, 2026 (with a peak probability around 3% near 50%), specifically neutralizing this strategy. The cumulative win rate of this wallet has since dropped to 54.7%.

Conclusion: The advantage of bots in price feed markets is real but limited to extremely narrow time windows and significantly compressed by the introduction of friction costs by the platform. Real-time Sports Game State Automation Data source: Wallet classification from the cancun2026 team on Polymarket (Dune query 6648075, https://dune.com/queries/6648075, past 7 days, as of 2026-05-11).

Source of advantage: Bots react to game events significantly faster than retail users using live streams (30 seconds delay). Additionally, trading terminals like Kreo and PolyCop open this advantage to non-programmer users through copy-trade and auto-follow features, thus the measured bot share includes human funds routed through bots. Cross-platform Combination Arbitrage Data source: IMDEA Networks paper "Unraveling the Probabilistic Forest: Arbitrage in Prediction Markets" (AFT 2025, dspace.networks.imdea.org/handle/20.500.12761/1941).

The study covers approximately $40 million in arbitrage extraction on Polymarket from April 2024 to April 2025, primarily consisting of two modes: one is the rebalancing of YES/NO shares within the same market; the other is cross-platform combination trading (entering when buying YES on Polymarket and buying NO on Kalshi when the implied probabilities of both are less than $1). This mode has rigid requirements for multi-platform infrastructure and compresses as the matching engines of each platform converge.

4. Areas Where Human Accounts Succeed and Their Limitations

The categories with the lowest proportion of bots are not because "retail picks more accurately," but because "the profitability of these markets depends on the ability to integrate multi-source real-world information," which is where automation continuously places humans at a structural disadvantage.

Two independent studies confirm this judgment.

On-chain behavior research by Joshua Della Vedova (University of San Diego) (jdellavedova.com) indicates that retail users have a higher frequency of picking winning outcomes than bots; the advantage of bots lies in execution—when retail buys YES at $0.72, bots have already entered at $0.55, yielding a floating profit of $0.17 per share.

A working paper from the University of Toronto / HEC Montréal / ESSEC (Akey et al., SSRN 6443103, March 18, 2026) points out that 56% of losing users' order prices fall into extreme ranges (<10¢ or >90¢), while only 28% of the top 0.1% of profitable users place orders in extreme ranges. The typical behavior of losing users is "chasing small probabilities at 5 cents for 20 times" or "chasing certainty at 95 cents," while profitable users typically build positions in the mid-range of the probability curve.

Both studies point to: the judgment ability of retail is generally underestimated, but the timing of execution and order structure is systematically weak.

5. Forward Path: Four Forces Determine the Direction of Bot/Human Dynamics

The key variable in the next 12-24 months is not the current bot/human ratio, but its evolutionary direction. This report identifies four forces, which do not align in direction. Further Collapse of Bot Deployment Costs Coding agents like Claude Code, Codex, open-source frameworks like Hermes, and Polymarket's own open-sourced Polymarket Agents framework under the MIT license collectively lower the engineering threshold for strategies like 0x8dxd from "serious projects" to "weekend prototypes." Copy-trade services further connect human capital to bot infrastructure, mechanically amplifying the measured bot share. Bot Individual Profitability Eroded by Peers The 823 profitable bot wallets are the right tail of a larger group of losing bots. An increase in the number of wallets with similar strategies means that the profitable window for each bot narrows. The 98% win rate of 0x8dxd is structurally non-replicable, not due to inefficiency disappearing, but because of peer competition + platform fee adjustments. The scale of capital extraction from humans by bots is likely to peak before the bot proportion does. Platform Category Structure Tilts Towards Retail Polymarket's category composition in April 2026: sports 50%, crypto 24%, politics 16%, others 10%. The composition in the same period of 2025: sports 29%, crypto 12%, politics 42%.

The absolute value of sports trading volume has increased 11 times year-on-year. The new volume mainly falls in long-cycle event markets, where retail holds an absolute advantage. Bernstein expects the share of sports in sector trading volume to drop from the current 62% to 31% by 2030, filled by economic, political, and corporate event contracts—this structural migration will further expand the exposure of categories where bots do not have an advantage. Different Platforms Naturally Divert by Category Hyperliquid's HIP-4 launched on May 2, 2026, offering daily BTC binary contracts, zero opening fees, USDH collateral, and a unified perpetual/spot market deployment mechanism (1 million HYPE per slot, approximately $42.76 million at current prices).

This is a typical market type where bot advantages are isolated and launched separately. Day-1 trading volume mainly comes from arbitrage funds, consistent with the historical distribution of BTC binary contracts. If HIP-4 subsequently expands into sports and political markets and integrates reliable oracles, its bot share may converge to Polymarket levels; at the current stage, its role is to isolate bot-friendly traffic to an independent platform, further drifting Polymarket's category structure towards retail.

6. Platform Landscape and Valuation Snapshot (Mid-2026)

▲ Source: Bernstein note (April 14, 2026), Polymarket / Kalshi public disclosures, HIP-4 launch announcement Conclusion: The combined share of Kalshi + Polymarket is 97%+, and equity opportunities at the platform level are basically closed for venture check sizes. Investable value is migrating to both above (trading terminal, quantitative strategy services, agent infra) and below (capital efficiency, arbitration, oracle) the platform layer.

7. Risk Warning

Risk 1: Regulatory tail risks. The three bills submitted by Schiff (DEATH BETS Act, Public Integrity Act, Prediction Markets Are Gambling Act), the TRO against Kalshi in Nevada, and criminal charges in Arizona in March 2026 constitute a tug-of-war between federal and state levels. Kalshi's 89% concentration of sports revenue is the most exposed business line, and sports or war/death-related contracts face a realistic probability of an all-category ban.

Risk 2: Oracle and arbitration failure risks. Polymarket integrated Chainlink for price markets in 2025, but subjective markets still rely on UMA. UMA's current token economy generates only about $600,000 in economic flow annually, corresponding to an FDV of $37 million; after MOOV2, proposer rewards have been narrowed to about 37 whitelisted addresses, most of which are Polymarket affiliates. Any controversial high-exposure ruling may trigger a reassessment of trust in the entire sector.

Risk 3: Risk of reversal in sports share. The growth of Polymarket's sports business in 2026 has seasonality (driven by NBA, NFL Super Bowl). If the sports share retracts, the overall dynamic of "bot share rising + retail expansion" may reverse.

8. Implications for Builders and Investors

The bot competition is essentially a question: in the $240 billion prediction market sector predicted by Bernstein for 2026, which layer captures value. There are four layers, each with different value densities.

L1 --- Agent trading products. Strategy advantages diminish, and C-end automated trading bears compliance risks. This layer is not recommended for standalone bets.

L2 --- Agent infrastructure (Olas / Valory model). An economic model that collects tolls, allowing any winning agent to earn fees. This layer is the cleanest investable option.

L3 --- AI-native data, pricing, market creation. Most are absorbed by internal platform teams or taken by existing Web2 incumbents (Kensho, Bloomberg, Dataminr). The remaining investable window is narrow.

L4 --- Arbitration and resolution. Current economic flow is real but small in scale. To become a Tier 1 venture target, a token model redesign is needed, and this design is currently not on the public roadmap.

Directions worth tracking at the edge layer:

• PM-DeFi composability (Morpho collateralizing PM positions, currently 2x leverage, roadmap 4-5x, affecting capital efficiency)

• Trading terminals and copy-trade services (like Kreo)

• PM-native quantitative institutions

• New market primitives (impact markets, futarchy, conditional markets)

Conclusion: Bots Win Categories, Humans Win Markets, Platforms Win Structures

Bots have not taken over prediction markets. The saturation of bots is in specific market types; the ratio of trading volume between any platform's bots and humans is essentially a downstream result of that platform's market type composition. The headline data of "5% wallets / 75% trading volume" confuses trading volume concentration with capital plunder. The main growth driver for Polymarket in 2026 comes from sports markets where bots do not have a structural advantage, and the $131 million extraction by bots mainly occurs in short-window crypto markets where retail participation is low.

Future winning platforms must possess three capabilities: accommodating multiple types of markets under credible arbitration conditions, appropriately balancing bot and human traffic, and retaining cross-category users. Polymarket currently occupies this position: Bitget's Q1 2026 research shows that multi-category users exhibit organic growth trends, with the average number of categories per user rising from 1.45 to 2.34, and active days increasing from 2.5 to 9.9.

Bots remain within their structural advantage range; human capital operating bots will continue to migrate to the next event; ultimately, the platforms that can accommodate both types of traffic across the most market types at appropriate ratios will prevail.