70% of top-earning wallets are bots, but AI hasn’t taken over prediction markets yet
TechFlow Selected TechFlow Selected
70% of top-earning wallets are bots, but AI hasn’t taken over prediction markets yet
Prediction markets are wealth-transfer machines, and bots are their operators.
Navigating Web3 tides with focused insightsAuthor: Jeff, IOSG
Executive Summary
The core data fueling panic around prediction market bots is strikingly straightforward: on Polymarket, 5% of wallets classified as “bot-like” account for 75% of trading volume. Since January 2025, 823 wallets have each netted over $100,000, collectively withdrawing $131 million in profits from Polymarket. Of the top 20 wallets on the profit leaderboard, 14 are categorized as bots (per Stacy Muur leaderboard inspection). A University of Toronto study—covering 2.4 million users and $67 billion in trading volume since 2022—found that 68.8% of users are unprofitable, while the top 1% captured 76.5% of all profits.
The resulting narrative is that prediction markets are wealth-transfer machines—and bots are their operators. The data is accurate, but the framing contains a fundamental bias.
Core Thesis
1. The central flaw in the “bot narrative” lies in conflating “trading volume concentration” with “capital extraction.” The fact that 5% of wallets on Polymarket generate 75% of volume reflects only the distribution of account activity—not proof that retail capital is being extracted by bots.
2. Aggregate-level data is more persuasive. AI agent wallets achieve ~37% win rates; human wallets, just 7–13%. A 3–4× gap at the population level is real evidence of structural advantage; meanwhile, the fact that 14 of the top 20 profit-leaderboard wallets are bots (Stacy Muur leaderboard inspection) represents merely the right tail of that distribution—not independent evidence.
3. Bot advantage is structural, not judgmental. Bots dominate three market types: price-feed latency arbitrage, real-time sports event state automation, and cross-platform combinatorial arbitrage. Their commonality? None requires judgment about real-world outcomes themselves. Once market resolution depends on synthesizing multi-source real-world information, bot advantages systematically erode.
4. Polymarket’s category composition has shifted over the past 12 months—from “Politics 42%” to “Sports 50%.” The fastest-growing categories are precisely long-horizon event markets where bots hold no structural edge—confirming a clear platform-wide trend toward retailization.
5. Forward-looking view: Bot share will continue rising as deployment costs fall—but the scale of capital extraction from humans by bots will peak *before* bot share does—because bots cannibalize one another faster than they extract from human accounts.
6. Investment strategy: Equity opportunities at the platform layer (Kalshi + Polymarket collectively hold >97% share) are effectively closed to venture-scale checks. Value migration is underway—to the L2 agent infrastructure layer (e.g., Olas / Valory models) and venue-agnostic middleware. C-end bot products and L3 data / pricing layers lack venture fit.
I. Market Size Outweighs Bot Panic
Three quantitative anchors define the scope of this report.
First, Bernstein revised its forecast for the prediction market sector to $240 billion for 2026E on April 14, 2026—with consensus across sell-side firms now projecting a path to $1 trillion by 2030.
Second, Kalshi and Polymarket combined YTD trading volume surpassed $60 billion in mid-April 2026—already exceeding the full-year 2025 total of $51 billion.
Third, Robinhood launched over 1,000 Kalshi contracts, with its >1M customers having traded 9 billion contracts cumulatively. Robinhood’s prediction market ARR stands at ~$350 million, up from $150 million in 2025 and projected to reach $586 million in 2026E—making it the firm’s fastest-growing product line.
Together, these figures point to one conclusion: Prediction markets are no longer solely a crypto-native sector—they increasingly resemble a TradFi distribution problem. The “retail investors being extracted from,” assumed in the bot narrative, are primarily not crypto-native users, but retail investors entering via traditional brokerage channels.
This implies a contextual bias in bot panic: The sector isn’t being drained of value by automation—it’s being flooded with traffic by mainstream finance at a pace far exceeding any speed of automated extraction.
II. The Real Data That Matters: 37% vs. 10%
The most frequently cited data point in the bot narrative suffers from sample selection bias.
The statistic “14 of the top 20 profit-leaderboard wallets are bots” originates from a small, pre-sorted sample ranked by profitability. This sample reveals only how bots occupy the right tail of the distribution—not whether they outperform humans at the population level.
Population-level data (sources: Polystrat / Valory disclosures, cross-validated with multiple on-chain Polymarket analytics datasets):
A 3–4× win-rate differential at the population level is the authentic signal of bots’ structural advantage. The 14/20 leaderboard statistic should be interpreted as a downstream manifestation of that win-rate distribution—not as standalone causal evidence.
III. Where Bots Win
Bot extraction is highly concentrated in three market types. Their shared trait: none requires subjective judgment about real-world outcomes—instead, they rely on latency or pricing advantages tied to platform matching engines.
Price-feed latency arbitrage
Representative case: wallet 0x8dxd, which turned $313 into $437,600 in just 15 minutes trading BTC up/down binary contracts in January 2026—achieving a 98% win rate.
Strategy mechanics: Monitors spot prices on Binance and Coinbase, entering positions on Polymarket when its quotes lag behind CEXs. Polymarket introduced a taker fee (peaking near 3% at ~50% probability) for 15-minute crypto contracts on January 7, 2026—specifically designed to neutralize this strategy. The wallet’s cumulative win rate has since declined to 54.7%.
Conclusion: Bot advantage in price-feed markets is real—but confined to an extremely narrow time window, and significantly compressed as platforms introduce frictional costs.
Real-time sports event state automation
Data source: cancun2026 team’s Polymarket wallet classification (Dune query 6648075, https://dune.com/queries/6648075, last 7 days, as of May 11, 2026).
Source of advantage: Bots react to in-game events significantly faster than retail users relying on live video streams (~30-second delay). Moreover, trading terminals like Kreo and PolyCop democratize this edge via copy-trade and auto-follow features—meaning measured “bot share” includes human capital routed through bots.
Cross-platform combinatorial arbitrage
Data source: IMDEA Networks paper “Unravelling the Probabilistic Forest: Arbitrage in Prediction Markets” (AFT 2025, dspace.networks.imdea.org/handle/20.500.12761/1941).
The study covers ~$40 million in arbitrage extraction on Polymarket between April 2024 and April 2025, driven mainly by two patterns: (i) rebalancing YES/NO shares within the same market; and (ii) cross-platform combo trades (e.g., buying YES on Polymarket and NO on Kalshi when the sum of their implied probabilities falls below $1). This pattern requires rigid multi-platform infrastructure—and shrinks as matching engines across venues converge.
IV. Where Human Accounts Win—and Their Constraints
Categories with the lowest bot share aren’t those where “retail picks better”—but rather those where profitability hinges on synthesizing multi-source real-world information—a domain where automation remains structurally disadvantaged relative to humans.
Two independent studies corroborate this.
Joshua Della Vedova (University of San Diego)’s on-chain behavioral research (jdellavedova.com) finds: Retail users select winning outcomes at higher frequencies than bots; bots’ edge lies in execution—e.g., entering at $0.55 when retail buys YES at $0.72, locking in $0.17 per share in unrealized gain.
A working paper by researchers from the University of Toronto / HEC Montréal / ESSEC (Akey et al., SSRN 6443103, March 18, 2026) shows: 56% of losing users place orders at extreme ends (<10¢ or >90¢), versus just 28% for the top 0.1% of profitable users. Losing users typically “chase low-probability outcomes at 5¢ for 20× upside” or “pile into near-certainty at 95¢ late”; profitable users predominantly build positions along the middle of the probability curve.
Both studies converge on this insight: Retail judgment is widely underestimated—but execution timing and order structure are systematically weaker.
V. Forward Path: Four Forces Shaping the Bot/Human Dynamic
Over the next 12–24 months, the key variable isn’t today’s bot/human share—but its trajectory. This report identifies four countervailing forces.
Further collapse in bot deployment cost
Coding agents like Claude Code and Codex, open-source frameworks like Hermes, and Polymarket’s own MIT-licensed Polymarket Agents framework have collectively reduced the engineering barrier for strategies like 0x8dxd—from “serious project” to “weekend prototype.” Copy-trade services further route human capital into bot infrastructure—mechanically inflating measured bot share.
Declining per-bot returns due to intra-bot competition
The 823 profitable bot wallets represent the right tail of a much larger cohort of unprofitable bots. As wallet counts employing similar strategies rise, each bot’s profitable window narrows. 0x8dxd’s 98% win rate was structurally non-replicable—not because inefficiency vanished, but due to peer competition + platform fee adjustments. Capital extraction from humans by bots will likely peak *before* bot share does.
Platform category mix shifting toward retail
Polymarket’s April 2026 category composition: Sports 50%, Crypto 24%, Politics 16%, Others 10%. In April 2025: Sports 29%, Crypto 12%, Politics 42%.
Sports trading volume grew 11× YoY in absolute terms. New volume is concentrated in long-horizon event markets—where retail dominates. Bernstein projects sports’ share of total sector volume will decline from 62% today to 31% by 2030—replaced by economics-, politics-, and corporate-event contracts—further expanding the category exposure where bots hold no advantage.
Natural category-based platform fragmentation
Hyperliquid’s HIP-4 launched on May 2, 2026, offering daily BTC binary options, zero opening fees, USDH collateral, unified perpetual/spot settlement, and validator-slashable market deployment (1M HYPE per slot, ~$42.76M at current price).
This is a textbook example of a bot-favorable market type being spun off onto a dedicated platform. Day-1 volume came largely from arbitrage capital—consistent with historical BTC binary contract behavior. If HIP-4 later expands into sports/politics markets and integrates trusted oracles, its bot share may converge toward Polymarket’s; for now, its effect is to isolate bot-friendly traffic onto a separate platform—further drifting Polymarket’s category mix toward retail.
VI. Platform Landscape & Valuation Snapshot (Mid-2026)
▲ Source: Bernstein note (April 14, 2026), Polymarket / Kalshi public disclosures, HIP-4 launch announcement
Conclusion: Kalshi + Polymarket hold >97% combined share. Platform-layer equity opportunities are effectively closed to venture-sized checks. Investable value is migrating upward (to trading terminals, quant strategy services, agent infra) and downward (to capital efficiency, arbitration, oracles).
VII. Risk Disclosures
Risk 1: Regulatory tail risk. Three bills submitted by Schiff (the DEATH BETS Act, Public Integrity Act, and Prediction Markets Are Gambling Act), Nevada’s TRO against Kalshi, and Arizona’s criminal charges filed in March 2026 constitute an active federal–state tug-of-war. Kalshi’s 89% revenue concentration in sports exposes it most acutely—sports or war/death-related contracts face realistic tail risk of blanket bans.
Risk 2: Oracle & arbitration failure risk. Polymarket integrated Chainlink for price-based markets in 2025—but subjective markets still rely on UMA. UMA’s current token economy generates only ~$600K in annual economic flow against an FDV of $37M; post-MOOV2, proposer rewards are restricted to ~37 whitelisted addresses—most affiliated with Polymarket. Any high-profile contested ruling could trigger a sector-wide trust reassessment.
Risk 3: Sports share reversal risk. Polymarket’s 2026 sports growth exhibits seasonality (driven by NBA, NFL Super Bowl). Should sports share retract, the “rising bot share + expanding retail” dynamic could reverse.
VIII. Implications for Builders & Investors
The bot debate is ultimately about one question: Within Bernstein’s $240B 2026 prediction market forecast, which layer captures value? Across four layers, value density varies sharply.
L1 — Agent trading products. Strategy edges decay rapidly; C-end automated trading carries compliance risk. Not recommended as a standalone bet.
L2 — Agent infrastructure (Olas / Valory models). A toll-road economic model—fee capture regardless of which agent wins. This is the cleanest investable layer.
L3 — AI-native data, pricing, market creation. Most is internalized by platform teams—or captured by incumbent Web2 players (Kensho, Bloomberg, Dataminr). Remaining investable windows are narrow.
L4 — Arbitration & resolution. Current economic flow is real but small. To become a Tier 1 venture asset, a redesigned token model is required—which is not on the current public roadmap.
Emerging areas worth tracking:
- PM-DeFi composability (e.g., Morpho collateralizing PM positions—currently 2x leverage, roadmap targets 4–5x, impacting capital efficiency)
- Trading terminals & copy-trade services (e.g., Kreo)
- PM-native quant firms
- Novel market primitives (impact markets, futarchy, conditional markets)
Conclusion: Bots Win Categories, Humans Win Markets, Platforms Win Structure
Bots have not taken over prediction markets. They saturate specific market types—and any platform’s bot-to-human trading volume ratio is fundamentally a downstream reflection of its market-type composition. The headline “5% wallets / 75% volume” conflates volume concentration with capital extraction. Polymarket’s 2026 growth is primarily driven by sports markets—where bots hold no structural edge—while the $131M in bot-extracted profits occurred overwhelmingly in short-window crypto markets where retail participation is itself low.
Future-winning platforms must possess three capabilities: (i) hosting diverse market types under credible arbitration; (ii) accommodating both bot and human traffic at appropriate ratios; and (iii) retaining cross-category users. Polymarket currently occupies this position: Bitget’s Q1 2026 research shows organic multi-category user growth—average categories per user rose from 1.45 to 2.34, and active days per user from 2.5 to 9.9.
Bots remain confined to their structural advantage zones; human capital running bots will continuously migrate to the next event; and ultimate winners will be platforms capable of hosting both flows—across the broadest possible range of market types—at the right balance.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
IOSG Ventures
