A $50k trade from a fresh wallet is not the same as a $50k trade from someone with a 90% win rate across 200 markets. The dollar amount is identical. The information content is completely different.
Polymarket sees thousands of trades daily. Most of it is noise. Market makers, people gambling, wash trading. I wanted to find the wallets whose activity actually predicted outcomes.
The 11 factors
I ended up with 11 things that seemed to matter. Each contributes to a score from 0 to 100.
Wallet history
- Age. Older wallets have track records you can analyze.
- Markets traded. More breadth means more data points.
- Win rate. The obvious one.
Trade characteristics
- Position size in absolute terms.
- Size relative to their usual. If someone usually bets $5k and suddenly drops $100k, that's interesting.
- Timing. How close to resolution?
Behavioral signals
- Conviction. Do they add to losing positions or cut and run?
- Contrarian moves. Are they fading the crowd?
- Cluster activity. Do they trade at the same time as other known whales?
Technical
- Wallet type. EOAs behave differently than smart contracts.
- Transaction patterns. Gas usage, timing, activity bursts.
The scoring
const score = weights.reduce((sum, w, i) => (
sum + w * normalizedFactors[i]
), 0);
return Math.min(100, Math.max(0, score));
Nothing fancy. Weighted sum, clamped to 0-100. The weights came from backtesting against historical market resolutions.
What counts as a whale
Depends on the market. On a $50k market, $5k is a lot. On a $10M market, it's a rounding error.
| Market size | Threshold |
|---|---|
| Under $100k | $5,000 |
| $100k to $1M | $15,000 |
| Over $1M | $50,000 |
An example that worked
Wallet 0xabc... dropped $120k on a political market when the price was at 35 cents. The system scored it 87.
Why? 94% historical win rate. 3-year-old wallet. Going against the crowd. Added to the position twice after the initial buy.
Four hours later, the market was at 78 cents.
This doesn't always happen. But when a high-scoring wallet moves, I pay attention.
The code is in the GottaTrackEmAll repo on GitHub if you want to dig into the implementation.