Across the 12,438 wallets Poly Syncer indexes, the top Polymarket traders share nine measurable habits: well-calibrated probability estimates, log-return distributions with positive median and tame tails, position sizes that hug a small fraction of bankroll, deliberate category breadth of two to six domains, fast exits when mid-prices misprice resolution, low single-market HHI, edge-adjusted hit rate above +0.05, sample sizes above 200 trades, and Bayesian-shrunk Sharpe ratios that survive sample-size adjustment. The patterns are statistical, falsifiable, and stable across the chains Polymarket settles on. This piece walks each pattern with the math, names the popular myths that do not survive scrutiny, and shows how we apply Bayesian shrinkage when judging a fresh wallet.
Methodology in one paragraph
The patterns below are computed across the indexed cohort of 12,438 wallets, restricted to wallets with at least 100 closed positions in a trailing 90-day window. The "top" set is the top quintile by composite score (sample size 1,891–2,492 wallets at any 15-minute refresh). Distributional comparisons use bootstrap confidence intervals at 95%; rank correlations use Spearman's ρ. Where we report a pattern as "more common in top wallets" we mean the top quintile vs the rest of the cohort, with bootstrapped 95% CIs that exclude the cohort baseline.
Pattern 1: Calibration that beats the market consensus
Calibration is the alignment between a forecaster's stated probabilities and observed frequencies. A perfectly calibrated forecaster who says "70% YES" is right exactly 70% of the time. The standard scoring rule for calibration is the Brier score; lower is better, with 0 being perfect.
Top Polymarket traders are not necessarily better-calibrated than the market in absolute terms (the market itself is well-calibrated on liquid contracts). But they are better-calibrated where they choose to bet. Their selected entries cluster in the implied-probability range $0.30–$0.55 where the market's calibration is genuinely worse, and they avoid the $0.85+ range where the market is essentially perfect and edges round to zero. The signal is "where they place bets," not "what fraction of all bets are won."
What this means as a copier
If you copy a wallet whose average entry price is $0.41 and whose win rate is 50%, that wallet is producing a 9-point edge from picking spots the market underestimates. If you copy a wallet whose average entry price is $0.79 and whose win rate is 80%, that wallet is producing essentially zero edge with high apparent win rate. Edge-adjusted hit rate makes this visible.
Pattern 2: Log-return distribution shape
We compute log-returns on each closed position: ln(proceeds / cost). The distribution shape carries a lot of information about how a wallet makes money. Top wallets show:
- Median above zero. Half or more of trades are profitable. The median is more informative than the mean, which is dominated by tails.
- Skewness slightly positive. Big winners outnumber big losers but not in a degenerate way. Excess skew (>1.5) is a red flag for outlier-driven scoring.
- Kurtosis moderate. Tails are present (binary contracts produce them) but not so heavy that one trade dominates. Excess kurtosis above 8 across 100+ trades suggests concentration.
Wallets in the top quintile have median log-return roughly +0.03 to +0.07, skew 0.3–0.9, kurtosis 4–7. Wallets that score high but fail this distributional check are flagged as outlier-driven; the winsorized Sharpe column on the leaderboard makes the difference visible.
Pattern 3: Position-size discipline
Position size as a fraction of bankroll is the single most-controllable risk lever a trader has, and the top quintile uses it with discipline. Specifically, we observe:
- Median position size is 0.8–2.4% of estimated bankroll.
- 95th-percentile position size rarely exceeds 6% of bankroll.
- The within-wallet standard deviation of position size is typically < 1.4% of bankroll.
Compare with the bottom-quintile-but-active cohort, where median is 4–9% and 95th percentile regularly exceeds 25%. The discipline is what produces the Sharpe difference, not the directional accuracy. The math is in our Kelly piece.
Pattern 4: Deliberate category breadth
The top quintile is not dominated by single-category specialists, despite the popular narrative. Across the indexed cohort:
| Number of distinct categories traded | Share of top quintile | Share of cohort baseline |
|---|---|---|
| 1 (pure specialist) | 11% | 22% |
| 2–3 | 34% | 28% |
| 4–6 | 41% | 30% |
| 7+ | 14% | 20% |
Top wallets cluster in the 2–6 category band. Pure specialists do exist and are over-represented in some categories (sports player props especially), but as a fraction of the top quintile they are roughly half their fraction of the broader cohort. The 7+ category wallets are also under-represented — spreading too thin appears to dilute edge.
Pattern 5: Fast exits when mid-price misprices resolution
Top wallets do not always hold to expiry. We measure the time-to-exit on positions that were closed before resolution and find that top wallets exit at a median of 38% of the way between entry and resolution time, vs 71% for the broader cohort. They are responding to mid-price moves: when a contract entered at $0.42 has run to $0.74 with a week still to resolution, taking the gain is the right call regardless of whether it would have resolved YES.
Pattern 6: Low single-market HHI
The Herfindahl-Hirschman index across markets is a structural concentration measure. Top quintile median HHI is 0.11 (volume spread fairly evenly across many markets); cohort baseline is 0.31. Wallets above HHI 0.4 are flagged on the leaderboard regardless of composite, because their score is structurally fragile to a single resolution.
Pattern 7: Edge-adjusted hit rate above +0.05
The single most-discriminating metric we have. Top quintile median edge-adjusted hit rate is +0.067 across the rolling 30-day window with sample ≥ 100 trades. The cohort baseline median is +0.004 — statistically indistinguishable from zero. The metric is robust to the shape of the entry-price distribution because it subtracts the mean implied probability, so it cannot be gamed by always trading favourites.
Pattern 8: Sample size as the foundation
This is meta-pattern more than pattern: top wallets earn their position by showing up consistently. Median trades-in-window for the top quintile is 287 (vs 71 cohort baseline). The standard error on win rate at n = 30 is roughly 9 percentage points; at n = 287 it is about 2.9 percentage points. The smaller error is what allows the score to reflect skill rather than reflect noise.
Pattern 9: Bayesian-shrunk Sharpe survives sample-size adjustment
This is the test that separates genuine top wallets from sample-size flukes. We compute a Bayesian shrinkage on raw Sharpe toward the cohort median, with shrinkage weight inversely proportional to sample size. The formula in essence is:
ShrunkSharpe = (n / (n + k)) * RawSharpe + (k / (n + k)) * CohortMedian
where n is the wallet's trade count and k is a tuning constant (we use k = 50). For a 32-trade wallet with raw Sharpe 4.0 and cohort median 1.1, the shrunk value is ~2.2. For a 287-trade wallet with raw Sharpe 2.4 and cohort median 1.1, the shrunk value is ~2.21.
This is an essential statistical trick because it tells you that, despite very different raw Sharpes, the two wallets have about the same evidence for skill — one has a strong but small sample, the other has a moderate but large sample. Top wallets in our top quintile have shrunk Sharpe ≥ 1.8 with high reliability.
Non-patterns (myths that do not survive scrutiny)
It is worth being explicit about what we do not see in the data, despite frequent claims:
"Top wallets all have huge bankrolls"
The correlation between estimated bankroll and composite-score quintile is small (Spearman ρ ~ 0.18). Excellent wallets exist at $5,000 working bankrolls and at $500,000 working bankrolls. Capital is not a substitute for edge.
"Top wallets win by always betting on favourites"
Average entry price for the top quintile is $0.46; for the cohort baseline it is $0.51. The top quintile is, if anything, slightly more contrarian on average. The win-rate-on-favourites strategy is exactly the one edge-adjusted hit rate is built to expose as low-edge.
"Top wallets are insiders"
If insiders dominated the top, we would see a small number of huge concentrated bets, not a large number of well-sized ones. The size and breadth distributions argue strongly against insider-driven scoring. Single-event insider plays do exist but they are filtered out by the 30-trade minimum and the sample-size shrinkage.
"Top wallets all use a Polymarket bot"
Some do. Many do not. Bot users tend to show up in the high-frequency segment of the leaderboard but are not over-represented at the very top of the composite. The distribution of fill timestamps within a wallet's history makes bot use somewhat detectable; we do not flag it because using a Polymarket bot is not a quality signal one way or the other.
"Top wallets always close before resolution"
About 53% of top-quintile positions close before resolution; the rest hold to expiry. There is no universal rule. The pattern is to exit when mid-price misprices, not to exit out of habit.
How we judge a fresh wallet that fits the patterns
A wallet that just crossed the 30-trade minimum and looks pattern-fit (good calibration, low concentration, disciplined sizing) is the most interesting case. The Bayesian shrinkage above prevents premature elevation, but the patterns themselves are early signals. Operationally:
- Compute raw composite. If > +1.4, mark as candidate.
- Apply Bayesian shrinkage. If shrunk composite > +0.7, hold for further data.
- Wait for sample to cross 100 trades. Re-evaluate. If shrunk composite is still > +1.4, the wallet has earned the badge.
This pipeline is the difference between "this wallet looks great so far" and "this wallet is in the top quintile with statistical reliability." The full implementation is in /methodology.
What this means for copy traders
Three practical implications:
- Filter on patterns, not on rank. A rank-7 wallet with HHI 0.6 is structurally weaker than a rank-22 wallet with HHI 0.12. The leaderboard guide's filter recipes are built for this.
- Diversify across patterns, not just across wallets. Pick one calibration-driven specialist, one disciplined-sizing generalist, one fast-exit politics wallet, one steady-accumulator earnings wallet. The patterns are partly uncorrelated; the diversification is real.
- Trust the math, not the narrative. A wallet's name, story, or confidence is irrelevant to its score. The numbers are what survives.
Frequently asked questions
Why use Bayesian shrinkage instead of a simple sample-size minimum?
A hard minimum is a step function: 99 trades is "out," 100 trades is "in." Bayesian shrinkage is continuous: 99 trades and 100 trades produce nearly identical shrunk scores. The continuous version captures evidence accumulation more honestly and avoids the discontinuity of a hard cutoff. We use both: the hard minimum for inclusion in the public leaderboard, the shrinkage for the displayed Sharpe value.
Why is Brier score not in the composite?
Brier score is a calibration measure for a forecaster who outputs explicit probabilities. Polymarket traders output trade decisions, not stated probabilities; we infer their implicit probability from their entry price relative to the contract. Edge-adjusted hit rate is the operational equivalent and is in the composite at 20% weight. Brier-style metrics are computed for backtest comparison but are not the user-facing score.
Are these patterns stable across categories?
Mostly. Position-size discipline and HHI are universal. Edge-adjusted hit rate ranges differ by category (politics wallets average +0.07; sports +0.05; crypto +0.04). Calibration is best in liquid categories with frequent resolutions (sports, earnings) and worst in tail-heavy categories (geopolitics, weather).
How often do top quintile wallets stay in the top quintile?
Across consecutive 30-day windows, the top-quintile retention rate is roughly 58%. About 25% drop to the second quintile, about 12% drop further, and about 5% leave the cohort entirely (insufficient activity). This is markedly better than random (20% retention) but not nearly perfect — another argument for diversification across multiple top-quintile wallets rather than concentration on one.
Where can I get the raw data behind these patterns?
The /developers API exposes per-wallet metrics matching the composite breakdown; the whitepaper documents the schema and the indexer. Aggregate distributions are available on request for academic research.
Where to go next
For the user-facing application of these patterns, see how to read the leaderboard and how to evaluate a single wallet. For the engineering pipeline that produces the metrics, the architecture post and scoring methodology are the references. To put any of this into practice, the setup walkthrough is the path. Plan flow: /dashboard/billing.