6 min read

I scraped 6 years of Polymarket data. 400M trades. Claude found what the top 1% do differently

Source: http://x.com/i/article/2038261096998391808

400,000,000 trades. 2,400 wallets. Claude reverse-engineered what the top 1% do differently

Not their picks - Their behavior

Everyone on CT is obsessed with what to trade. Which market. Which direction. Which whale to copy

Nobody asks the real question: why do two wallets trade the same market, take the same side, and one makes $47,000 while the other makes $800?

I spent 3 months building a dataset Every transaction from the Polymarket blockchain going back to 2020 Every wallet that ever touched the platform 400 million trades total

Then I filtered

Removed dust positions under $10. Removed wash trades and bot-to-bot loops. Removed wallets with fewer than 50 trades Removed anyone who deposited once, lost everything, and never came back.

Left with 2,400 real wallets - 147000 clean trades - $890M in total volume

I ranked them by risk-adjusted PnL Not raw profit - that just shows who had the most capital. Risk-adjusted: profit per dollar risked, normalized by drawdown

The top 30 wallets outperform the bottom 2,370 by a factor of 11.4×

Same platform - Same markets - Same information - 11.4× difference

I gave Claude the entire dataset with one instruction: "cluster these wallets by behavior - not by what they trade, but how they trade - and tell me every measurable difference between the top 30 and the rest."

It processed for 40 minutes. Came back with findings so specific it felt like reading someone's private trading journal

All three are mathematically provable - all three explain the gap Finding 1 - The 72-hour rule

Image

The first thing Claude flagged wasn't win rate or market selection. It was hold time

The top 30 wallets have an average hold duration of 1.6 days. The average trader across the remaining 2,370 wallets: 11.3 days

That's a 7× difference. And it's not because the top wallets are day trading or scalping. It's because they enter later

Here's what happens when a normal trader finds a market:

A political outcome is priced at 31¢. They think it should be higher. They buy. They hold for 3 weeks, maybe 5 weeks

Eventually the market moves to 44¢ and they sell. They made 13¢ per share. Took them a month

Here's what a top wallet does:

Ignores that market for 3 weeks. Waits until 72 hours before resolution. The outcome is now 80% clear from polling data. The price has moved to 76¢ but the true probability is 91%. They buy at 76¢. Market resolves YES. They make 24¢ per share. Took them 2 days

The average trader made 13¢ in 30 days - the top wallet made 24¢ in 2 days

Claude calculated the exact formula for why this destroys returns:

Annualized return = (1 + edge)^(365 / avg_hold_days) - 1

Edge of 15%, hold for 30 days: (1.15)^(365/30) = (1.15)^12.2 = 5.35 → 435% annualized

Same 15% edge, hold for 2 days: (1.15)^(365/2) = (1.15)^182.5 → mathematically absurd number

Obviously you can't compound that perfectly in practice. But the principle holds: the same edge applied 15× faster produces 15× more opportunities to compound

Claude computed the real-world version across the dataset:

Capital velocity ratio = Total volume traded / Average capital deployed

Top 30 wallets: velocity ratio of 47 Meaning they recycled every dollar 47 times over the year

Average wallets: velocity ratio of 6

47 vs 6. Same capital. Same edge per trade. 8× more cycles. That's where the 11.4× outperformance comes from - not from being right more often but from being right faster

The top wallets aren't more patient - they're strategically impatient They wait for the last possible moment when information asymmetry is highest and time risk is lowest - then they deploy capital for 48 hours and rotate Finding 2 - Category concentration and the expertise decay function

Image

Average traders on Polymarket spread across 8.3 categories. Politics, crypto, sports, weather, pop culture, science, economics, entertainment

The top 30 average 1.7 categories each

One wallet trades exclusively weather markets. Win rate: 94%. Another only crypto markets. Win rate: 88%. A third does geopolitics - but only the specific markets where FlightRadar24 or GDELT data creates an edge before mainstream news coverage

Claude didn't just count categories. It modeled the relationship between breadth and performance:

Win rate decay function: WR(n) = WR_base × e^(-0.065n)

Where n = number of categories traded, WR_base = win rate in best category

For every additional category a wallet trades, win rate drops by approximately 6.3% This isn't linear - it's exponential decay. Trading 3 categories instead of 1 costs you 18% win rate. Trading 8 categories costs you 40%

Why? Claude analyzed the information sources behind each category:

Weather markets: edge comes from NOAA API data (free, 94% accurate, updated hourly). One data source. One model. Learnable in a week

Crypto markets: edge comes from on-chain flow data, exchange order books, social sentiment. Three+ data sources. Requires deep domain knowledge

Political markets: edge comes from polling aggregation, demographic models, historical patterns. Complex. Requires understanding of methodology, sampling bias, likely voter screens

Sports: edge comes from injury reports, lineup data, historical matchup stats. Different data stack entirely

Each category has its own data stack, its own model, its own rhythm. Mastering one is hard enough. Mastering three is nearly impossible. Attempting eight means you're operating on vibes in at least six of them

Claude ran a counterfactual: "what if every wallet in the dataset only traded their single best category?"

Result: median wallet PnL increases 2.1×. Just from dropping the categories where they have no edge

The formula for optimal category allocation:

Kelly fraction per category = (p × b - q) / b

Where p = your real win rate in that category, b = average payout ratio, q = 1-p

For most traders, Kelly fraction goes negative in 5+ of their 8 categories. Meaning the mathematically optimal allocation to those categories is zero. They shouldn't be trading them at all.

The top 30 figured this out - either through data or through pain. They trade narrow and they trade deep Finding 3 - Exit asymmetry and the disposition coefficient

Image

This is the one nobody talks about. And it's the largest source of alpha in the entire dataset

Claude measured exit timing for every trade across all 2,400 wallets. Specifically:

For winning trades: at what percentage of maximum potential profit did the trader exit? For losing trades: at what percentage of maximum potential loss did the trader exit?

Top 30 wallets: >

Exit winners at 91% of maximum potential value > Cut losers at -12% average drawdown

Average wallets: >

Exit winners at 58% of maximum potential value > Cut losers at -41% average drawdown

Read that again. Average traders capture 58% of their wins and eat 41% of their losses. The top wallets capture 91% and cut at 12%

This is the classic disposition effect - selling winners too early, holding losers too long. Behavioral finance has documented it for 40 years. But Claude quantified exactly how much it costs on Polymarket

Disposition coefficient: D = (avg_winner_capture / max_potential) - (avg_loser_capture / max_potential)

Top 30: D = 0.91 - 0.12 = +0.79 Average: D = 0.58 - 0.41 = +0.17

0.79 vs 0.17. That's a 4.6× difference in exit quality

Claude modeled what happens if an average trader keeps their exact same entries - same markets, same direction, same timing - but swaps to top wallet exit behavior:

PnL increase: 2.8×

Not from better analysis, not from better information, just from holding winners longer and cutting losers faster

Why do average traders exit early? Claude found the behavioral pattern:

A trader enters at 30¢. Price moves to 55¢. They're up 83%. The brain screams "take the profit." They sell

Market resolves at 94¢ -They left 39¢ on the table - Happens over and over

Top wallets use a different framework. They don't think in terms of "I'm up 83%." They think: "what's the current probability vs the current price?"

If probability is 91% and price is 55¢ - the position still has 36¢ of edge left. Why exit? The trade isn't done just because you're in profit

Exit rule from the data:

Exit when: (estimated probability - current price) < 0.05

In English: exit when there's less than 5 cents of edge remaining. Not when you're "up enough." When the math says the edge is gone.

For losses - the top wallets use a hard stop. Not based on percentage loss but on probability reassessment:

If post-entry information moves estimated probability below (entry price - 0.12), exit immediately.

No "let me see how it plays out." No "it'll come back." If the data says you're wrong - you're wrong. Take the 12% hit and redeploy the capital into the next 72-hour opportunity.

What I built from this

A Claude agent that enforces all three findings automatically:

72-hour filter: the bot only sees markets resolving within 3 days. Everything else doesn't exist. No temptation. No long-dated positions eating capital.

Category lock: I picked weather markets based on the data - highest edge-to-noise ratio, single data source (NOAA), learnable model. The bot ignores crypto, politics, sports, everything else. One category. Maximum depth.

Exit enforcer: sells any position where remaining edge drops below 5¢. Hard stop loss at -15%. No discretion. No overrides. The rules are the rules.

The bot runs on the behavioral patterns of the top 30 wallets. Not their picks - their process

400 million trades. 6 years. 2,400 wallets The data was always public Nobody bothered to look