Methodology
Every other prediction-market leaderboard sorts by total profit. That rewards exactly one thing: having made money. It says nothing about whether that money came from real skill or from one enormous lucky bet. Sirtio Score is our attempt to separate the two.
Trader A
+$500,000
One long-shot bet on a single election market. Got it right. Total career trades: 3.
Trader B
+$500,000
A steady 8% edge across 1,400 trades over a year, spanning a dozen categories.
Same profit. A raw-PnL leaderboard ranks them identically. Trader B is the one worth paying attention to -- Sirtio Score is built to tell them apart.
Every trader on the leaderboard first had to appear on Polymarket's own public Monthly leaderboard -- their top 100 traders ranked by profit over the trailing 30 days. We don't scan every wallet on Polymarket; we only see traders who Polymarket itself surfaces as recent top performers by that measure.
Once we have those 100 wallets, we look at each one's own trading history -- specifically, resolved positions from the last 90 days -- and compute Sirtio Score from that, independent of how they ranked on Polymarket's monthly list. That means the two windows don't match exactly: a wallet could be a strong monthly performer but a weaker 90-day one, or vice versa within the pool.
We also filter out likely bot / market-making wallets before scoring. A wallet with an extreme number of resolved positions or trade events in the window isn't a human placing predictions -- it's almost certainly a script. Wallets that cross that line have their data collection cut off early rather than fully processed, so a "3,000+" style count for one of these wallets reflects that cutoff, not their true total.
A trader not appearing here doesn't mean they're not skilled -- it may just mean they didn't crack Polymarket's own top-100-by-monthly-profit cutoff. This leaderboard is a snapshot of the strongest recent performers among Polymarket's own top-ranked traders, not a comprehensive ranking of every trader on the platform.
Data refreshes once daily via our automated pipeline -- so rankings reflect the most recent daily snapshot, not real-time trading.
Earlier versions of this score used a "win rate" component -- counting how often a trader's positions resolved in their favor. We removed it entirely. The problem isn't the concept of measuring accuracy; it's that classifying any individual position as cleanly "won" or "lost" depends on Polymarket's settlement and redemption mechanics, which don't always cleanly track position state -- a position a trader exited early, or a market that resolved in an unusual way, can be miscounted. This is a well-documented, independently-observed problem in the prediction-market data community, not something we ran into alone.
Realized PnL sidesteps this entirely: it's a continuous dollar value that's always well-defined once a position closes, regardless of how or why it closed. Nothing in the current model requires classifying a position as won or lost.
Sirtio Score is built on a statistical technique used across fields that face the same underlying problem: estimating someone's true skill from a limited, noisy sample. A handful of trades can look great by chance; a large, consistent track record is much harder to fake.
Rather than scoring a trader's raw average return directly, the model blends it with what's typical across the whole tracked pool -- weighted by how much real history exists for that specific wallet. A trader with only a few resolved positions gets pulled closer to the pool average; a trader with a long, consistent record gets judged much more on their own numbers. The score is then based on how far that blended estimate sits from average, relative to how confident the model actually is in that estimate -- rewarding traders who are both profitable and consistent, not just traders who got lucky once.
We're intentionally not publishing the exact formula, weights, or constants behind this -- see below. What's above is the real shape of how it works, not a simplification hiding something different underneath.
[ formula hidden for now -- coming soon ]
A trader who's statistically indistinguishable from the average tracked trader -- given how much history exists on them -- scores exactly 50. Scores above that reflect a trader the model is increasingly confident is outperforming; scores below reflect the opposite. Tier labels on the leaderboard (Elite, Great, Good, and so on) are set from the real distribution of scores across currently tracked traders, not fixed round numbers -- they get re-checked as more history accumulates.
The exact thresholds are still early -- this model has only been running against real data for a short time. As more resolved positions accumulate across more traders, the tier cutoffs get revisited against the real, current distribution rather than left as a first guess.