I Analyzed 3 Polymarket strategies from my polymarket arbitrage bot performance - Here's What They're Actually Doing
I pulled three accounts showing bot-like behavior (thousands of predictions, repetitive market patterns) and reviewied what they were running purely from the position data.
Here's what I found.
The three accounts
| Account | Joined | Predictions | P&L | Biggest Win |
|---|---|---|---|---|
| @first | Oct 2025 | 19,923 | +$1,484 | $1,006 |
| @second | Apr 2026 | 16,454 | $0 today | $606 |
| @third | Jun 2026 | 4,170 | +$1,342 | $390 |
@first - directional momentum bot
19,923 predictions since October 2025. Eight months running. Every open position I could see was in the "Bitcoin Up or Down - 5 minute" market series, and all of them were single-sided - just Up or just Down, never both.
Entry prices I saw: 33¢, 44¢, 69¢, 76¢, 81¢, 89¢. That spread tells a story. It's not buying at a fixed price - it's reading something about the market and entering wherever the signal fires.
The 33¢ Up entry that returned 202% was almost certainly triggered on a strongly trending BTC candle where the market was badly underpricing the obvious direction. You see heavy momentum, the market is at 33¢ for some reason, you buy it.
Total P&L of +$1,484 across ~20k trades is thin per-trade. A lot of losses mixed with occasional big wins. But it's been running 8 months without stopping - and that's the real signal. Directional strategies are harder to compete away than pure mechanical arb because they require actual market judgment. You can't just copy the approach, you need the signal.
Verdict: Directional BTC 5m momentum scalper. Not arbitrage. Still active after 8 months - the edge hasn't been competed to zero.
@second - the arb bot that finished
16,454 predictions since April 2026. Current positions: zero. Portfolio value: $0. Biggest win: $606.
Zero open positions isn't a bug - it's the signature of the merge/redeem arb strategy done correctly.
Here's how it works. Polymarket binary markets have YES and NO tokens. In an efficient market:
P(YES) + P(NO) = $1.00
When the order book is thin, you sometimes get:
P(YES) = $0.47 P(NO) = $0.49 ───────────────── Sum = $0.96 ← $0.04 edge per dollar deployed
You buy both YES and NO simultaneously, use Polymarket's merge function to combine them into a complete set, then redeem for $1.00. The direction BTC moves is irrelevant - you get $1.00 either way. Profit is locked at entry.
bowbow16 ran this 16,000+ times in about 2 months. Then stopped. That's not a failure - that's what this strategy looks like when the edge gets competed away. As more bots pile in, the spread tightens, the window shrinks to milliseconds, execution risk eats the profit.
Verdict: Textbook merge/redeem intra-market arb. Ran hard, extracted the edge, exited when the spread tightened. Currently inactive.
@third - live arb caught mid-execution
Joined June 2026. Brand new account. 4,170 predictions. When I checked the open positions today, this is what I saw:
Market: Bitcoin Up or Down - June 29, 2:45AM-2:50AM ET Down 43.4¢ 625.1 shares → $309.42 (+$38.11) Up 59.1¢ 595.9 shares → $300.95 (-$51.00)
Both sides. Same market. Same 5-minute window. That's not an accident - that's the intra-market merge/redeem arb running live.
The bot entered both legs simultaneously when the combined price was below $1.00. Then prices moved:
43.4¢ + 59.1¢ = 102.5¢ ← above $1.00, currently underwater ~$12
This happens when there's latency between detecting the arb signal and both orders filling. The spread was probably 95-97¢ when the signal fired, but one leg filled slightly slower and the price moved before the second order landed. This is the execution risk problem in real time, visible in a public profile.
What's important: wolf9478 doesn't care which direction BTC moves. The directional P&L numbers are misleading. The real P&L is (redeem value − entry cost), locked at execution. It's waiting for resolution, then it merges both legs and redeems for $1.00 per complete set.
Verdict: Active intra-market arb bot, caught mid-trade. Slightly underwater from execution slippage. The clearest real-world example of the strategy I've seen.
What this tells you about Polymarket arb in 2026
The arb windows are shrinking. third got slipped 2-3 cents between signal and fill. bowbow16 ran 16k trades then stopped. What used to be seconds-wide is now milliseconds-wide as more bots compete for the same spread.
Pure mechanical arb has a shelf life. second ran April to June and is now idle. eknih has been running since October 2025 and is still going. The directional strategy requires ongoing signal generation that can't just be copied. The arb strategy eventually gets competed to zero by anyone who can see the same spread.
The 5-minute BTC markets are saturated. Every bot I found is concentrated there. Fast resolution, consistent liquidity, clean binary structure - but that also means maximum competition. If you're building something new, less-trafficked multi-outcome markets have wider arb windows that stay open longer.
Execution latency is the real moat. The intra-market arb math is trivial. The API is public. The hard part is getting both legs filled before the spread moves. That's an infrastructure problem, not a strategy problem.
The bot I built from studying this
After looking at these patterns, I built an open-source Polymarket bot in Python implementing 5 strategies:
Strategy 1 - Intra-market merge/redeem (what third is doing)
Buy YES+NO when sum < $1.00, merge, redeem. Risk-free on paper, execution-speed-dependent in practice.
Strategy 2 - Combinatorial arb
Multi-outcome markets (3+ outcomes). All prices should sum to $1.00. When they don't - same approach. Less competition, wider spreads.
Strategy 3 - Cross-platform arb
Compare Polymarket prices against Binance spot. If BTC spot implies 72% probability above $X but the market shows 45%, buy the underpriced side and wait for convergence.
Strategy 4 - Endgame arb
Markets near resolution with 93%+ probability on one outcome. Buy it, collect ~7% return in a few hours. Works best in markets other bots ignore.
Strategy 5 - Momentum/mean-reversion (what eknih appears to run)
Track YES/NO prices as a time series, apply z-score + RSI, enter when overextended. Three timeframes: 5m scalp, 15m swing, 1h position.
Signal ranker composite score:
score = (
profit_score * 0.30 +
confidence * 0.25 +
strategy_prio * 0.20 +
urgency * 0.15 +
risk_reward * 0.10
)
# Priority weights
# Intra-market: 1.00 ← risk-free, always first
# Combinatorial: 0.95
# Endgame: 0.90
# Cross-platform: 0.80
# Momentum: 0.70 ← directional, lowest certainty
If you're building something similar
Don't start with the 5m BTC markets. They're fully arbitraged. Look at multi-outcome markets, longer resolution times, less liquid assets - the spread stays open longer because fewer bots bother.
Measure fill latency, not signal quality. The strategy is easy. The fill latency between both legs of a merge/redeem trade is the variable that determines whether you profit or take slippage.
Track P&L per strategy. If you run 5 strategies, you need to know which ones are carrying the total and which are dragging it.
Run dry-run mode for a week first. Log every signal and simulated fill before touching real capital. Your own orders affect price on thin books - backtests don't show this.
Repo: github.com/casatrick/polymarket-arbitrage-bot - Python, Polymarket CLOB API, WebSocket for live prices. MIT license. Star it if this was useful.
⚠️ Not financial advice. Polymarket markets can be illiquid. You can lose money. The accounts analyzed are public profiles - I'm describing what the position data suggests, not confirmed facts about their implementation.
Comments
Post a Comment