Why You Can't Spot Arbitrage Fast Enough on Prediction Markets (And How Bots Win at Milliseconds)

Sophisticated bots extracted roughly $40 million in arbitrage profits from Polymarket between April 2024 and April 2025, according to a peer-reviewed analysis of 86 million bids across 17,218 markets (arXiv 2508.03474, Aug 2025). Yet only about 1% of detectable election-arb opportunities were actually executed — the rest expired in the seconds it would take a human to react (arXiv 2508.03474, Aug 2025). If you're trying to spot cross-platform arbs by eye, you're not slow. You're playing a different sport.

**Key Takeaways** - Bots extracted $40M in Polymarket arbitrage profits between Apr 2024 and Apr 2025; only ~1% of detectable opportunities were captured ([arXiv 2508.03474](https://arxiv.org/abs/2508.03474), Aug 2025) - The average cross-platform arbitrage window collapsed from 12.3 seconds in 2024 to 2.7 seconds in 2026 — faster than a human can even verify a price across two browser tabs - A Dublin-colocated bot has roughly **80–180 milliseconds** of total latency budget; a typical adult's simple visual reaction time is around **190 ms** just to perceive a price change ([Wikipedia on mental chronometry](https://en.wikipedia.org/wiki/Mental_chronometry)) - The cost ladder runs from $5/mo (Hetzner + free RPC, hobby) to $50K+/mo (NY4 colocation) — and the price tier maps almost perfectly to which arb windows you can catch - Retail still wins on slower windows (political news, minute-to-hour arbs), thin markets where bot capacity is capped, and as a maker collecting Polymarket's up-to-3.15% dynamic fees ([Finance Magnates](https://www.financemagnates.com/cryptocurrency/polymarket-introduces-dynamic-fees-to-curb-latency-arbitrage-in-short-term-crypto-markets/), Jan 2026)

For the full mechanics of how cross-platform arb works at a strategy level, see our companion post on automating cross-platform arbitrage between prediction markets. This post is specifically about the speed dimension — where the milliseconds go, why humans can't compete on raw latency, and what retail can still profitably do.

The $40M Opportunity and the 1% Capture Rate

The single most important paper on prediction market arbitrage to date — published in August 2025 by researchers at IMDEA Networks — analyzed every Polymarket bid placed between April 2024 and April 2025. They found three categories of arb worth a combined $40M:

• $10.58M from single-condition arbitrage (YES + NO across platforms summing to less than $1.00)
• $23.28M from market-rebalancing arbitrage (within a single platform, exploiting redundant outcome pricing)
• $95,156 from rarer combinatorial arbitrage (across multiple correlated markets) (arXiv 2508.03474, Aug 2025)

The top arbitrageur netted $2,009,631 across 4,049 transactions — about $496 average per trade. The top 10 accounts captured around $9.1 million combined (arXiv 2508.03474, Aug 2025). These weren't anomalies. They were the predictable output of bots with the right infrastructure running for 12 months.

But the most damning number is the capture rate: out of every detectable arb the researchers could identify in U.S. election markets, only about 1% was actually executed by anyone. The other 99% expired before any trader — human or bot — could fill both legs (arXiv 2508.03474, Aug 2025).

That's the headline. There's roughly 99x more theoretical arbitrage profit available than gets captured, and the bottleneck is purely speed.

The 2.7-Second Window (And Why It Keeps Shrinking)

Cross-platform arbitrage windows between Kalshi and Polymarket collapsed from an average of 12.3 seconds in 2024 to 2.7 seconds in Q1 2026 (Medium / ILLUMINATION, Q1 2026). The peer-reviewed cross-check comes from a separate arXiv paper analyzing 75 million NBA market orderbook snapshots: in-game single-market arb opportunities persist a median of just 3.6 seconds (arXiv 2605.00864, 2026).

Same order of magnitude. Both numbers tell you the same thing: opportunities now decay faster than the human visual system can update.

Why is the window shrinking? Because every additional sub-100ms bot on the network competes for the same opportunity. Polymarket itself has been actively hiring HFT C++ engineers with lock-free programming and multithreading experience throughout 2026 (eFinancialCareers, 2026). The platforms themselves are professionalizing.

**[ORIGINAL DATA]** Combining the arXiv $40M figure with the 1% capture rate implies the *latent* arbitrage pool on Polymarket alone is closer to **$4 billion per year** at current liquidity levels. Most of that never gets touched because nobody is fast enough.

Where the Milliseconds Go: A Bot's Latency Budget

A profitable cross-platform arbitrage requires the bot to:

1. Receive a price update from one venue
2. Parse it and update its internal orderbook
3. Recompute whether the arb is still profitable after fees and slippage
4. Sign and submit one or both legs of the trade

Each step has a known latency cost. Here's the breakdown for a Dublin-colocated bot trading Polymarket — which runs its CLOB on AWS eu-west-2 (London) based on community traceroute analysis (NYC Servers, 2026):

A few things to notice from this stack:

• Polymarket's own internal processing dominates — about 70–85 ms regardless of where your bot lives (QuantVPS, 2026). You cannot optimize that part. It's the floor.
• Cloud region matters more than CPU speed. A bot in NY adds ~70–80 ms in network round-trips alone. Singapore adds 200+ ms. A retail trader on a New York Comcast connection is starting the race 80 ms behind a Hetzner Dublin VPS (NYC Servers, 2026).
• REST polling is dead. Crypto trading guides put REST polling at 50–150 ms baseline before any real network transfer. WebSocket subscriptions push that to 5–15 ms (CoinAPI, 2025). If your bot polls a REST endpoint every second, by the time it sees a quote the bots on WebSocket have already filled the arb.
• Polygon RPC latency varies by 4x between providers. QuickNode hits ~86 ms global average, Ankr ~164 ms, Alchemy ~207 ms global (with sub-50 ms via their Cortex routing) (Uniblock, 2025). For chain-settled platforms, RPC choice is part of the latency stack.

Add it up. A well-built Dublin bot lands an order in roughly 80–180 ms from the moment a CEX price changes. A retail trader sees the price change on their browser approximately 190 ms after their visual system processes it (Wikipedia on mental chronometry), and by then the bot has already filled and exited.

Why Humans Literally Can't See These Opportunities

The numbers above assume you've already noticed the arb. That's the easy part. The hard part is verifying it across two browser tabs.

A typical college-age adult's simple visual reaction time is around 190 ms. Elite sprinters at the Beijing Olympics averaged 166–169 ms, with one-in-a-thousand best-of starts hitting 109–121 ms (Wikipedia on mental chronometry). That's just seeing a number change. Now add tab-switching to verify the price on the other platform — Harvard Business Review research cited in 2022 found knowledge workers switch between apps and tabs roughly 1,200 times per day at ~2 seconds per switch (HBR via TechCrunch, 2022). A realistic manual cross-platform verification — see price on Polymarket, switch to Kalshi tab, find the equivalent contract, eyeball the spread — takes 5 to 15 seconds.

In that 5-second window, the 2.7-second arb has already opened and closed twice.

That's why Polymarket bots average 89 trades per active day vs 2.2 for human wallets based on Della Vedova's bot-detection methodology, where any wallet making 50+ trades on any day or 1,000+ trades total qualifies as a bot (Bloomberg via Della Vedova / Univ of San Diego, Apr 2026). It's not that bots are smarter. It's that they get 40x more chances at the same opportunity.

For the deeper picture on adverse selection and why retail systematically gets picked off, see our breakdown on why prediction market trades get picked off.

The Cost of Speed: What Each Tier Actually Buys You

Most "you need infrastructure" advice is useless because it doesn't quantify what each spending tier actually gets you. Here's the real ladder:

Sources: Hetzner (2026), QuantVPS VPS comparison (2026), QuickNode pricing (Mar 2026), NYC Servers (2026).

A few things worth knowing about the ladder:

• Hetzner does not have a Dublin or London point-of-presence — only Nuremberg, Helsinki, Falkenstein, and Singapore as of 2026 (Hetzner Cloud). For Polymarket's AWS eu-west-2 backend, that adds ~10–15 ms vs a true Dublin/London VPS. For Kalshi (US-hosted), Hetzner's European footprint is slower than NY-based hosting.
• NY4 colocation gives 0.36 ms latency to Polygon RPC endpoints in some setups, LD4 hits 0.56 ms (NYC Servers, 2026). That's institutional territory but quantifies what the top tier buys.
• Premium RPC matters more than people realize. The 4x latency gap between QuickNode (86 ms) and Alchemy global (207 ms) is the difference between catching and missing arbs that depend on settlement-confirmation latency (Uniblock, 2025).
• Sub-100ms bots reportedly capture roughly 73% of all arb profits in 2026 (Medium / ILLUMINATION, Q1 2026). The institutional tier dominates. The "serious retail" tier picks up scraps but real ones — political and news-driven arbs where the window is measured in seconds-to-minutes, not milliseconds.

**The honest framing:** You don't need to outrun Susquehanna. You need to figure out which arb windows are wide enough for your latency budget — and trade only those.

If you're trying to figure out which platform sits where on this curve, our comparison of prediction market bot platforms breaks down the infrastructure assumptions each one makes.

The Maker/Taker Pivot (The Biggest 2026 Shift Nobody's Writing About)

Here's a change most arb content hasn't caught up to. In January 2026, Polymarket introduced dynamic taker fees that can hit 3.15% on 50-cent contracts for fast-moving 5-minute, 15-minute, and hourly crypto markets — explicitly to curb latency arbitrage (Finance Magnates, Jan 7, 2026). The platform also replaced its initial 500 ms maker advantage with a 250 ms delay across short-duration crypto contracts (Yahoo Finance, 2026).

The strategic implication: the highest-edge moves in 2026 are no longer about being a faster taker. They're about being a maker that collects rebates.

That's a different game entirely. Maker bots don't have to win the speed race — they have to post quotes that are tight enough to win flow but wide enough to survive adverse selection. Bartlett and O'Hara's Stanford Law study of 41.6 million Kalshi trades published in April 2026 found that market makers earn about 2x more per contract in single-name markets than in broad markets specifically because they widen spreads to compensate for informed flow (Stanford Law, Apr 2026). The same math applies to Polymarket makers competing for rebate flow.

For retail, this opens a door. If you can't compete on taker latency, you can compete on maker discipline — using tools like TurbineFi Studio to post and manage quotes systematically rather than chasing fills.

What Retail Can Actually Win At in 2026

The doom-loop framing — "bots win, you lose, go home" — misses where retail still has real edge:

1. Slow-window arbitrages. Political news arbs, late-breaking sports lineup changes, and regulatory announcements often have arb windows of 5 to 60 minutes, not seconds. Bots optimize for the sub-3-second windows because that's where most volume is. The longer windows get less algorithmic attention. A retail trader with a good alert system can compete here.

2. Thin liquidity. The NBA arbitrage paper documented that 76.9% of combinatorial opportunities were constrained to an average of just 14.8 shares of executable size because of orderbook depth limits (arXiv 2605.00864, 2026). Big bots are capital-inefficient at $15-per-trade max size. Retail with $1K–$10K bankroll can profitably scoop these. They're not glamorous, but they're left on the table.

3. Cross-platform calendar arbs. When Kalshi and Polymarket list the same event with different resolution dates (e.g., "by June 30" vs "by end of Q2"), the bots avoid them because of resolution-risk uncertainty. A retail trader who reads the rules carefully can take both sides where the bots can't.

4. Maker rebates on quiet markets. Polymarket's dynamic fee structure rewards quote provision on slower markets. A retail trader posting limit orders on political and economic markets — and using strategy automation to manage them — captures rebate flow without needing institutional speed.

5. Semi-automated alert systems. You don't need to execute in 100 ms. You need to be alerted in 100 ms. A bot that watches both venues and pings you on Telegram when a profitable cross-platform spread opens lets you make the human judgment call on whether to fire — and on slow-window arbs, that's enough.

The arXiv $40M figure cited at the top of this post breaks down into $10.58M from cross-platform single-condition arbs (mostly bot-captured) and $23.28M from rebalancing arbs within Polymarket (arXiv 2508.03474, Aug 2025). The cross-platform piece is where bots dominate. The within-platform rebalancing is where retail discipline still pays — if you know which markets to watch.

That's where our toolset comes in. TurbineFi Studio lets you build alerts and semi-automated triggers in plain English, without writing the WebSocket-handling and orderbook-reconstruction code yourself. It's not going to make you a sub-100 ms bot. It will let you compete in the categories where speed isn't the only edge.

A Concrete Day-One Setup for Retail (No Coding)

If you want to start catching the arbs that aren't already lost to bots, here's the minimum viable stack:

1. A $25/mo Dublin or Frankfurt DigitalOcean droplet — closer to Polymarket's eu-west-2 backend than any US-based home connection.
2. WebSocket subscriptions to Kalshi and Polymarket — never REST polling. Your alerts need to be event-driven, not on a 1-second timer.
3. A semantic event mapper — the same Kalshi contract and Polymarket contract often have slightly different wording. Mismatched mapping is the #1 cause of "I lost on a 'guaranteed' arb."
4. A two-tier alert system — fast (sub-second) for crypto/sports markets; slow (1–5 minute) for political/economic markets where the window is longer and the false-positive cost is lower.
5. A pre-set execution budget per opportunity — most retail arbs blow up because the trader sees the spread, panics, and overbuys. Predefine the position size and stick to it.

You can build all five yourself in two to four weeks of engineering work, or TurbineFi Studio bundles them as a hosted product so you start trading the same day.

Frequently Asked Questions

How fast does a bot need to be to capture cross-platform arbitrage?

Approximately sub-100 ms end-to-end to catch the average 2.7-second cross-platform window, with around 73% of arb profits going to bots in that tier (Medium / ILLUMINATION, 2026). Sub-200 ms still catches a meaningful share but increasingly competes with the institutional bots. Above 500 ms, you're effectively only catching the slow-window arbs (political news, calendar mismatches).

What does it cost to run a competitive arb bot in 2026?

For a Dublin-hosted retail setup, expect $75–125/month total — DigitalOcean Dublin droplet ($24), QuickNode Build plan ($49), and basic alerting. Pro retail with bare-metal hosting and QuickNode Accelerate ($249) runs $300–600/month. True institutional NY4/LD4 colocation with QuickNode Business ($999) or Enterprise plus dedicated RPC is $5,000–50,000+/month (NYC Servers, QuickNode, 2026).

Can a human realistically catch any cross-platform arbs manually?

On the 2.7-second windows, no — a single browser tab switch takes ~2,000 ms and full cross-platform price verification takes 5–15 seconds (HBR via TechCrunch, 2022). But on political news arbs with 5–60 minute windows, manual execution still works if your alerts trigger fast enough. The trick is to let bots watch and humans execute.

Is there a tool that automatically tracks arbitrage across prediction markets?

Yes — multiple. TurbineFi Studio handles the WebSocket subscriptions, semantic event mapping, and alert layer in a no-code interface. Standalone scanners like ArbBets and EventArb exist too. The key differentiator is whether the tool handles resolution-risk reconciliation, atomic execution, and fill verification — most don't, which is how partial-fill losses wipe out small-account traders.

Why does Polymarket process orders so much slower than the network round trip?

Polymarket's CLOB does 70–85 ms of internal processing on top of any network latency (QuantVPS, 2026). That includes order validation, EIP-712 signature verification, orderbook matching, and writing to Polygon. It's the largest fixed cost in any Polymarket bot's latency stack and you can't optimize past it.

The Bottom Line

The $40M figure is the headline. The 1% capture rate is the story. There's roughly 99x more arbitrage profit available on prediction markets than gets captured today, and the only thing standing between that profit and the rest of the market is latency infrastructure.

If you're trying to compete on raw speed against sub-100 ms bots with co-located NY4 servers and dedicated Polygon validators, the math says you'll lose. But the math also says:

• The institutional tier captures ~73% of arb profit — meaning 27% is still left on the table for everyone else
• The longer windows (political, economic, sports lineup) are still retail-playable
• The maker-rebate shift in 2026 changes the optimization target entirely
• Thin-liquidity markets are bot-capacity-limited and favor smaller accounts

The traders who profit from prediction market arbitrage in 2026 aren't the ones who think they can build a faster bot than Polymarket's HFT engineering team. They're the ones who pick the arb categories where speed isn't the only edge — and then automate them properly. TurbineFi Studio is built specifically for that workflow.

---

Disclaimer: This post is for educational and informational purposes only and is not financial, legal, or trading advice. Prediction market trading involves substantial risk of loss. Past performance does not guarantee future results. Latency figures and infrastructure costs cited are approximate and change as platforms evolve; verify against current vendor documentation before deploying capital. The arbitrage statistics referenced here come from peer-reviewed and industry analyses linked inline; the underlying methodologies and full disclosures are available at those sources.