← Leo Labs

I Reverse-Engineered 40 Top Polymarket Addresses. Only 3 Strategies Actually Make Money.

Data API + chain analysis — 40 addresses, 100K+ trades, three strategies

Which strategy actually works for the people making millions on Polymarket?

I pulled the leaderboards for sports and crypto—top 20 addresses in each—and reverse-engineered 40 accounts across 100,000+ trades using Data API + LB API + chain data. Not dashboard screenshots. Every buy, every sell, every REDEEM gets reconstructed into behavior.

After dissecting the data, I found something interesting: regardless of market, profitable addresses fall into three completely different buckets. They're not just running different parameters. They're playing entirely different games.

---

1Type 1: Directional — Pick Right, Hold to Expiration

The most profitable sports strategy is almost stupidly simple. I didn't believe it at first.

Out of 18 effective addresses on sports: 14 only buy, never sell. Hold until expiration. Win → redeem. Lose → zero. No trading.

But the same "buy and hold" generates wildly different returns.

swisstony: $494M traded, 1% return rate, $4.96M net profit. Fully automated. 353 orders in 30 minutes covering the five major leagues. Pennies per trade, but the volume is insane.

majorexploiter: 39% return, max single position $990k. 600+ trades, almost all-in on two Arsenal matches. Big bets win big—$2-3M swings.

One person scales horizontally. One person bets large. Both made millions. Different playbooks, same edge: information advantage in the events they're betting on.

The Leaderboard #1 Is Losing Speed

kch123 tops the sports leaderboard with $10.35M cumulative profit.

But as of mid-March, the last 30 days: -$479k. Win rate past 7 days: 31% (15W-33L). All 14,303 trades are buys. Zero sells. 493 trades per day. 74% have <10 second intervals.

A machine that made $10M is stalling. You won't see that on the leaderboard. You only see it when you dissect the chain data.

I Mislabeled Myself With Bad Heuristics

fengdubiying, sports #13, $3.13M profit.

During my batch analysis, I tagged him as "sell-heavy"—looks like he trades volatility.

Actual data: 93.6% of proceeds come from REDEEMs. Sells account for 6%. Real strategy: concentrated LoL bets. Largest position: $1.58M on T1 vs KT Rolster, 74.4% win rate, 7.5:1 payoff ratio.

Selling is his stop-loss tool, not his primary strategy. Look at the buy-sell ratio on the dashboard and you completely misread what he does.

---

2Type 2: Structural — Make Money Without Predicting

Crypto leaderboard is a different species entirely. Sports = direction betting. Crypto = market making.

The top 5 in crypto break down like: three market-making bots running binary options on price moves, one inventory manager using MERGE, one specialist in fundraising milestone arbitrage (43.3% return).

Retail bets direction. Top players run the house.

How Market Makers Make Money

0x8dxd, BTC 5/15-minute binary options market maker.

94% of trades are symmetric: buying both YES and NO simultaneously. Runs 24/7. Median position size under $6. Buy-in spread: price +/- <$1 on both sides. The spread is the profit. At least three independent addresses run the same pattern.

Another MM address is more extreme: essentially monopolizes liquidity in the Economics category. 982 buys, zero sells, six-figure PnL. Making money off maker rebate + liquidity premium.

Good Code Doesn't Mean Good Profits

You might think market making is a license to print money. There's an open-source Polymarket MM bot on GitHub with solid engineering: WebSocket live feeds, risk management suite (stop-loss + volatility freeze + sleep periods), auto position consolidation. The author admits it doesn't make money.

Why? The pricing logic is penny jumping—quote $0.01 ahead of the best bid. That's just following, not pricing. No original pricing model, no edge.

Engineering rigor doesn't matter. MM profitability depends entirely on your pricing model beating the market.

One more stat worth noting: chain analysis shows over 70% of crypto arbitrage profits on Polymarket go to bots with <100ms latency. Fewer than 8% of wallets are profitable overall. If your bot runs on second-scale latency, you're providing liquidity for the fast players.

---

3Type 3: Knowledge-Based — Small Orders, High Conviction

Type 3 addresses are a completely different animal. Very low trade frequency—maybe two or three positions a month—but each one backed by deep research.

A few examples: one address models weather using public meteorological data, enters only when win probability exceeds 77%, probably two or three positions monthly, five-figure profits per trade. Another account is 89% short (buying NO), monthly-scale holding periods, mediocre win rate, but 9x average payoff per position. A few big winners cover all small losses.

Even more extreme: in FDV (full denomination) markets, does one thing: buy NO at 50-55 cents, cash out at $1. 100% win rate. Not luck—market miss it, so the mispricing exists.

But knowledge-based isn't just "research hard enough and you win." I analyzed one case: someone built a BTC price anomaly probability matrix from 1.37M rows of historical data, backtested perfectly, walked it forward and it collapsed. Markets get efficient fast. Rules that worked last month are arbitraged away this month.

The real edge for knowledge traders is deeper understanding of a specific market category than market pricing reflects, not a more complex model.

---

Comparing All Three Playbooks

---

What I'm Running

Now the personal part.

I'm running multiple lines simultaneously: crypto market making (structural), sports probability pricing (directional), weather data modeling (knowledge-based). None are at kch123's scale—no 493 daily orders, no $494M volume.

After reverse-engineering these 40 addresses, the most important realization: knowing which game you're playing beats optimizing any parameter.

Directional betting without information advantage is guessing, no matter how clean your execution. Structural plays with second-scale latency? You're the one getting extracted. Not motivational nonsense—this is what the data told me.

Now I run each line small-scale first, validate the edge exists, then scale. No rushing. Get one or two categories working before expanding.