In February 2026, we pointed a detection algorithm at 7,655 anonymous wallets that had collectively wagered $113 million across 1,337 prediction markets on Polymarket—and told it nothing about which wallet, if any, was suspicious. The algorithm ranked every wallet by how anomalous its trading pattern looked: how concentrated, how well-timed, how consistent across unrelated markets. Its top-ranked wallet, out of all 7,655, was one that journalists had already flagged in press reports. The machine found the needle without being told there was a needle to find.
This is the story of what that blind search revealed—and, just as importantly, what it could not.
Key Findings at a Glance
- The algorithm ranked the known wallet #1 out of 7,655. A wallet previously reported in press coverage as a suspected insider ranked first globally when scored on combined behavioral anomalies and real profitability from Polymarket’s own leaderboard. However, the watchlist explicitly includes this wallet regardless of its liquid rank, partially compromising the "blind" framing.
- Real profitability confirmed the behavioral signal. Using Polymarket’s leaderboard API (pre-computed P&L for 5,000+ wallets), we confirmed the wallet of interest has +$154,219 all-time profit—rank #696 on the platform. 15 of 20 watchlist wallets are profitable, with combined real P&L of over $1.3M.
- Noise dominates raw rankings. All 50 of the top-50 unfiltered results came from markets with 5 or fewer participants. A liquidity filter (markets with ≥$50K volume and ≥20 participants) is essential to separate signal from noise.
- Candidate wallets pick winners more often—but one wallet dominates. The 11-wallet watchlist achieved 80.6% value-weighted accuracy on buy trades, beating the 66.0% baseline by nearly 15 percentage points. However, a single wallet (0x4bfb...982e) contributes 85% of the watchlist's buy-trade value, so this aggregate is largely one wallet's story.
These headline numbers require careful unpacking: roughly 85% of the watchlist's buy-trade value comes from a single wallet, no formal statistical significance tests have been applied, and the wallet of interest itself scored below the buy-accuracy baseline. The gap is real. What it means is another question.
Background and Research Question
Polymarket is a decentralized prediction market built on the Polygon blockchain. Traders buy and sell outcome tokens—for example, "YES" tokens on the question "Will Israel invade Lebanon in September?"—at prices between $0 and $1. When the event resolves, winning tokens pay $1 and losing tokens pay $0. The market price thus represents the crowd's implied probability that the event will occur.
Because all trades settle on a public blockchain, every wallet's trading history is visible. This creates a natural laboratory for studying whether certain wallets consistently trade in the right direction at the right time—a pattern that might indicate access to non-public information, superior analytical skill, or some combination thereof.
In previous work, the detection pipeline was tested on a curated set of markets—those where a specific wallet of interest was already known to have traded. This approach is useful for development but introduces a methodological problem: the analysis was conditioned on the answer it was supposed to find.
This study asks a simple question: If we feed the pipeline every market in our database without telling it which wallet to look for, does it still surface the same wallet as anomalous?
Data: What We See and What We Don't
What is observed
The primary dataset consists of 46,865 "whale entry" signals—records of large or concentrated trades flagged by the pipeline's first-pass detector. These span 1,337 unique prediction markets (identified by hexadecimal condition IDs) and 7,655 unique wallet addresses, with a combined notional value of $113.5 million.
It is critical to understand that these are not complete trading ledgers. The whale-entry detector captures large, concentrated entries. A wallet may have made dozens of smaller trades or exits that do not appear in this dataset. Metrics computed from this data describe the wallet's large-trade behavior, not its full portfolio.
For outcome validation, we retrieved resolution data for 960 of 1,337 markets (71.8%) via the Polymarket Gamma API, identifying which outcome ultimately won. For forward price analysis, we collected daily price data from two sources: an existing snapshot covering 53 markets (only 4.0% of the universe), and a targeted backfill of 81 additional markets covering 91.5% of the watchlist's trading volume.
What is not observed
Several pieces of the puzzle are invisible to this analysis. We cannot see a wallet's complete trading history (only the large trades that triggered the detector), full market order books at the time of each trade, or off-chain information sources like private intelligence, news feeds, or social connections. We also cannot determine who controls any wallet, or whether they possess material non-public information. This analysis detects statistical patterns in on-chain data. It cannot establish identity, intent, or legal liability.
How the Pipeline Works
Pipeline overview
The detection pipeline operates in four stages. First, it ingests all whale-entry trade records for the specified market universe. Second, it applies a liquidity filter to remove micro-markets that generate spurious high scores. Third, it scores each wallet-market pair on concentration, timing, and cross-market behavioral features, assigning both market-local and global labels. Finally, it selects a watchlist of candidate wallets from the top-ranked results for deeper analysis.
Inside the scoring engine
Transparency matters. This section lays out, in full, the features and formulas that produce the
effective_priority_score_v2 used to rank every wallet-market pair. Nothing is hidden;
the score is deterministic, hand-weighted, and fully reproducible from the input data.
The score is a two-component sum: a market-local score (how anomalous the wallet looked within a single market) plus a scaled global score (how anomalous it looked across markets):
The 0.35 multiplier is a design choice: local behavior (what a wallet did in a specific market) is weighted roughly 3× more than global behavior (what it did across markets). This reflects the pipeline's philosophy that the strongest signal of unusual activity is dominance and timing within a market, not just activity breadth.
Market-local score (0–5.5 points, or 0–6.3 with markout data)
Within each market, every wallet is ranked against all other wallets in that market on five dimensions. The score uses percentile ranks (not absolute thresholds), making it robust to market-specific variation. If a wallet ranks in the 95th percentile on value, it scores 95% of that component's maximum weight. The five components:
| Component | What It Measures | Max Weight |
|---|---|---|
| Value rank | Total dollar value traded by this wallet in this market, ranked against all wallets | 2.0 |
| Burst dominance | Maximum share of trading value in any single 60-minute bucket—did the wallet concentrate activity into a short burst? | 1.2 |
| Late-market share | Wallet's share of all late-window (final 48h) market value—did it dominate the end-of-market activity? | 0.9 |
| Markout signal | 7-day sign-adjusted price drift after trades (only available where OHLCV data exists) | 0.8 |
| Trade count rank | Number of trades in the market, ranked against all wallets | 0.6 |
The heaviest weight goes to value rank (2.0 of 5.5). The logic: a wallet that moves the most dollars in a market is the most structurally significant, regardless of how many trades it took. Burst dominance (1.2) captures temporal concentration—a wallet that put 80% of its activity into a single hour looks different from one that spread it evenly across weeks. Markout (0.8) is the only forward-looking component but is treated as optional because OHLCV data coverage was only 4.0% before the targeted backfill.
Global score (0–8.5 points, scaled to 0–3.0 in the final sum)
The global score captures cross-market behavior. Unlike the market-local score, it uses absolute thresholds with soft caps rather than percentile ranks:
| Component | What It Measures | Max Points | Saturates At |
|---|---|---|---|
| Markets traded | Number of distinct markets this wallet traded in | 2.0 | ≥4 markets |
| Total trades | Total whale-entry trades across all markets | 2.0 | ≥20 trades |
| Top-market concentration | Fraction of all the wallet's trades that went into its single busiest market | 2.0 | ≥50% in one market |
| Late-window share | Fraction of all trades across all markets that fell in the final 48h window | 1.5 | ≥50% late |
| Total value (log-scaled) | Log of total dollar value traded, scaled to cap at $200K | 1.0 | ≥$200K |
The raw global score (0–8.5) is multiplied by 0.35 before adding to the local score, so its effective contribution to the final score is capped at ~3.0 points. A wallet that trades across 4+ markets, makes 20+ trades, concentrates on a dominant theme, times its entries late, and moves significant dollar value will max out the global component.
Label assignment: how features become classifications
Labels are assigned by deterministic threshold rules, not model predictions. Two independent labels are assigned to every wallet-market pair, then combined:
Market-Local Labels
Based on the wallet's behavior within a single market:
Concentrated — Same concentration rules, but without the late-window clustering.
Single Strike — Exactly 1 trade with ≥$1,000 value.
Late Accumulator — 2+ trades, ≥50% in final 48h, but below concentration threshold.
Diffuse — Default. Spread activity, no concentration flags.
Global Labels
Based on the wallet's behavior across all markets:
Single-Theme Dominator — ≥50% of all trades in one market and ≥6 total trades.
Opportunistic Whale — 3–11 total trades. Default for moderate activity.
Low-Evidence Singleton — ≤2 trades total. Not enough data to classify.
The two labels combine into an effective label using a simple priority matrix:
| If Local Label Is… | And Global Label Is… | Effective Label |
|---|---|---|
| Concentrated / Concentrated Late / Single Strike | Cross-Market Repeat or Single-Theme Dominator | Priority: Insider-Like (Market & Global) |
| Concentrated / Concentrated Late / Single Strike | Any other | Priority: Market-Local |
| Any other | Cross-Market Repeat | Watch: Cross-Market Network |
| Any other | Any other | Background / Diffuse |
This is why the effective label matters more than the raw score: a wallet can score high on the numeric scale but land in "watch" (not "priority") if its local pattern is diffuse. Conversely, a wallet with a lower score can be "priority" if it shows both local concentration and cross-market repetition. The watchlist's 3 "watch" wallets (including the wallet of interest, 0x0afc) are there because they show cross-market repeat behavior without dominating any single market.
Why the liquidity filter matters
Before the liquidity filter, the results were useless. Not "noisy" or "imperfect"—useless. Every single one of the top 50 ranked wallet-market pairs came from markets with five or fewer participants. Nearly half came from markets with less than $5,000 in total volume. The algorithm was doing exactly what it was designed to do—finding wallets that dominated their markets—but in a two-person market, domination is arithmetic, not evidence. Imagine declaring someone the world's best poker player because they won a hand at a table where they were the only person sitting down.
The fix was simple but essential: require at least $50,000 in total market value and at least 20 participants. This cut the dataset from 13,967 to 10,400 wallet-market pairs. And in this filtered landscape, the signal was immediate. The thresholds themselves are not derived from data or pre-registered—they are pragmatic cutoffs. A sensitivity analysis across different thresholds (e.g., $25K/10 participants, $100K/50) has not yet been performed, and rank stability under alternative thresholds remains an open question.
Measuring "correctness": did the wallet bet on the winner?
For resolved markets, we check whether each trade ended up on the winning side. A buy trade is correct if it purchased the outcome that eventually won. A sell trade is correct if it sold an outcome that did not win.
Measuring "markout": did the price move in the trade's favor?
Markout measures whether the market price moved in the direction of a trade over a fixed time horizon (2 days or 7 days). If a trader buys "YES" at $0.60 and the price rises to $0.70 two days later, the 2-day markout is +$0.10 (a favorable move). This is not the same as realized profit—the trader may have sold earlier or held longer—but it captures whether the trade's directional bet was validated by subsequent price movement.
Results
Universe and coverage
Table A summarizes the scope of the analysis. The pipeline processed the full universe of 1,337 markets with at least one whale-entry signal, covering $113.5 million in observed trade value.
| Metric | Value |
|---|---|
| Markets in universe | 1,337 |
| Whale-entry fills (large trades) | 46,865 |
| Total trade value | $113,466,747 |
| Unique wallets scored | 7,655 |
| Wallet-market pairs (all) | 13,967 |
| Wallet-market pairs (after liquidity filter) | 10,400 |
| Markets with resolved outcomes | 960 of 1,337 (71.8%) |
| Resolved fills (all trade types) | 38,553 |
| Resolved fills (buy trades only) | 25,386 |
| Snapshot price data coverage | 53 markets (4.0%) |
| Targeted backfill price data | 81 markets (91.5% of watchlist value) |
The low initial price-data coverage (4.0%) was a significant limitation. The targeted backfill brought coverage to 81 markets covering 91.5% of watchlist trading volume ($31.0M of $33.9M), but this backfill was not random—it prioritized markets where the watchlist was most active, creating potential selection bias.
How the pipeline classified 13,967 wallet-market pairs
The pipeline assigns each wallet-market pair—a single wallet's behavior in a single market—both a global label (based on cross-market behavior patterns) and a market-local label (based on behavior within that individual market). These combine into an effective label that determines priority ranking. Because one wallet can appear in multiple markets, the 7,655 unique wallets produce 13,967 wallet-market pairs.
Only 304 wallets earned the "global cross-market repeat" label—meaning they showed concentrated, well-timed trading across multiple unrelated markets. All 11 watchlist wallets carry this global label, and 8 of the 11 also earned the highest-priority effective label ("priority insider-like, market and global").
Resolution correctness: watchlist vs. baseline
To evaluate whether the flagged wallets actually bet on the right outcomes, we compared their resolution correctness against the baseline of all 6,199 wallets with resolved trades.
Baseline (All 6,199 Wallets)
Watchlist (11 Wallets)
The standout number is buy-side accuracy: 80.6% vs. 66.0%—a 14.6 percentage point gap across $19.6 million in resolved buy-trade value. However, correctness does not measure how much they made. A wallet buying a near-certain winner at 99 cents (1% upside) counts the same as one buying at 50 cents (100% upside).
At the individual level, 7 of 11 watchlist wallets exceed the 66.0% buy-accuracy baseline individually (0x000d at 87.8%, 0xe899 at 99.9%, 0x3dfc at 95.3%, 0xe64f at 89.5%, 0x4bfb at 80.1%, 0x5cd5 at 74.1%, 0x7a70 at 73.9%). Four wallets fall below baseline, including the wallet of interest (0x0afc at 52.5%). This count is less sensitive to the single-wallet concentration problem, but still reflects an in-sample selection: these wallets were chosen because they scored highly on concentration and timing, so some uplift in accuracy is expected.
Note also that the baseline (66.0%) includes the watchlist wallets themselves—the watchlist contributes 29% of the baseline's buy value. A cleaner comparison would exclude watchlist wallets from the baseline, which has not been done here.
Forward price drift (markout analysis)
Table C shows whether prices moved favorably after the watchlist's trades, compared to all wallets in the same 81 markets.
| Group | 2-Day Drift | 2-Day Win Rate | 2-Day Fills | 7-Day Drift | 7-Day Win Rate | 7-Day Fills |
|---|---|---|---|---|---|---|
| All wallets (baseline) | −12.1 bps | 51.7% | 26,070 | −109.9 bps | 55.3% | 19,011 |
| Watchlist (11 wallets) | +3.2 bps | 58.9% | 8,623 | −72.9 bps | 55.0% | 6,388 |
| Delta (watchlist − baseline) | +15.3 bps | +7.2 pp | — | +37.0 bps | −0.3 pp | — |
At the 2-day horizon, the watchlist shows a modest positive drift (+3.2 basis points, i.e., hundredths of a percent) compared to the negative baseline (−12.1 bps). The +15.3 bps delta is directionally interesting but economically tiny—well below normal execution friction (spread, slippage, fees) in many prediction market settings. Without confidence intervals, we cannot distinguish this from noise.
The 7-day picture is less favorable. Both groups show negative drift: −72.9 bps for the watchlist vs. −109.9 bps baseline. The watchlist's loss is smaller, but its 7-day win rate (55.0%) is actually marginally worse than baseline (55.3%). This mixed signal—positive at 2 days, negative at 7 days—could mean short-term informational advantage that dissipates, positions exited before the 7-day mark, or simply noise in a small sample. The markout is also computed from daily close prices (not trade-time prices), introducing additional imprecision.
The 11 candidate wallets
The watchlist was formed by taking the unique wallets from the top 20 liquid-filtered wallet-market pairs, plus one previously identified wallet of interest (0x0afc...30b2). This is an important caveat: the watchlist is not purely blind. The pre-selected wallet ranked #1 on the unfiltered global score but only #19 after the liquidity filter the article argues is essential. It was included regardless of rank, and all aggregate watchlist metrics (80.6% accuracy, markout figures) are computed with this wallet in the pool. This yielded 11 wallets, whose key metrics are displayed in Table B. For each wallet, the table shows data from its best-scoring liquid market.
| Wallet | Label | Best Market | Market Size | Value Share | % in Final 48h | All-Trade Acc. | Buy Acc. | 2d Drift (%) | 7d Drift (%) |
|---|---|---|---|---|---|---|---|---|---|
| 0x000d...758e | Watch | Israel-Hezbollah Ceasefire | $20.2M | 1.9% | 83.2% | 86.8% | 87.8% | −2.9% | −13.2% |
| 0x0afc...30b2 | Watch | (metadata unavailable) | $10.6M | 0.3% | 0.0% | 75.2% | 52.5% | +9.5% | +4.5% |
| 0x3dfc...acb7 | Priority | Israel-Hamas Ceasefire | $1.1M | 7.2% | 6.7% | 71.1% | 95.3% | +0.1% | +4.5% |
| 0x4bfb...982e | Priority | Ukraine Holds Kursk (Oct) | $2.3M | 15.9% | 3.8% | 68.8% | 80.1% | +0.02% | −0.7% |
| 0x5cd5...ac33 | Priority | Ukraine Holds Kursk (2024) | $961K | 4.0% | 65.0% | 37.6% | 74.1% | +1.7% | +0.8% |
| 0x7a70...52cc | Priority | Israel Invades Syria | $4.7M | 1.9% | 69.9% | 61.6% | 73.9% | +6.2% | −8.3% |
| 0xbaa2...2c73 | Watch | Israel Invades Syria | $4.7M | 1.1% | 41.2% | 58.9% | 63.5% | +0.8% | +1.9% |
| 0xc5d5...f80a | Priority | Maduro Wins Venezuela | $509K | 45.1% | 35.0% | 31.6% | 56.3% | +0.7% | −5.1% |
| 0xd7ad...4953 | Priority | Iran Military Response | $982K | 4.8% | 81.8% | 55.4% | 36.9% | +2.0% | +10.9% |
| 0xe64f...44a5 | Priority | Israel Invades Lebanon (Nov) | $3.0M | 8.8% | 84.8% | 45.1% | 89.5% | −0.6% | −2.0% |
| 0xe899...0899 | Priority | Ukraine Holds Kursk (Aug) | $936K | 10.1% | 51.6% | 99.9% | 99.9% | −0.6% | +1.6% |
Reading the table: "Label" indicates the pipeline's classification—Priority (insider-like on both market-local and global dimensions) or Watch (cross-market network pattern). "Value Share" is what fraction of the market's total large-trade volume belonged to this wallet. "% in Final 48h" is the fraction of the wallet's trading value placed in the final 48 hours of market activity. "All-Trade Acc." and "Buy Acc." measure how often the wallet's trades were on the eventually-winning side. "2d Drift" and "7d Drift" show the average price movement in the wallet's favor over those horizons.
Notable Wallet Profiles
Rather than exhaustively catalog all 11 wallets, we highlight the most distinctive patterns.
What the watchlist reveals about market themes
The watchlist's top markets by trading value reveal a strong geopolitical concentration. The 25 largest markets by watchlist activity are dominated by Middle East conflicts (Israel-Hezbollah, Israel-Hamas, Israel-Lebanon, Iran military response, Iran nuclear), the Ukraine-Russia war (Kursk occupation, ceasefire timelines), and one Latin American election (Venezuela). This thematic clustering could reflect genuine information advantages in geopolitically sensitive markets, but it could also simply reflect that these are the largest, most liquid prediction markets on Polymarket and naturally attract the biggest traders.
| Market | Watchlist Volume |
|---|---|
| Israel x Hezbollah Ceasefire in 2024? | $3,786,661 |
| Will Israel invade Syria in 2024? | $3,034,654 |
| Will Israel invade Lebanon in September? | $2,029,683 |
| Will Israel invade Lebanon before November? | $1,277,066 |
| Russia x Ukraine ceasefire by March 31, 2026? | $1,258,425 |
| Will Ukraine hold Kursk through October 31? | $1,157,626 |
Note: Four additional markets totaling $9.87M in watchlist volume are omitted because their metadata (market question text) was not available from the Gamma API at query time. The geopolitical concentration pattern is even more pronounced when those markets are included.
Following the Money: Profitability Analysis
Behavioral patterns tell you how a wallet trades. But the question that matters most is did it make money? We took a two-seed approach to answer this at scale:
- Profitability seed: We pulled real, pre-computed P&L for the top 5,000 wallets from Polymarket’s leaderboard API—the platform’s own profit calculations.
- Behavioral seed: Our existing whale-entry scoring pipeline flagged 7,655 wallets by trading anomalies.
- Union & re-score: We merged both sets (11,923 unique wallets, 732 overlap) and re-ran the full scoring pipeline using real P&L instead of estimates.
Why real P&L, not estimates?
Our initial profitability layer estimated P&L from whale-entry fills only. This systematically undercounted wallets whose profits came from trades below the whale detection threshold. The wallet of interest, for example, showed an estimated $8,132 in the pipeline but $154,219 in reality—a 19× undercount. The leaderboard API solves this: it uses Polymarket’s own full-history computation, matching what users see on their profile pages.
What the scoring engine measures
The combined pipeline scores each wallet on two layers:
- Behavioral features (up to 8.5 pts)—market concentration, timing patterns, trade size, late-trading clustering, cross-market breadth.
- Profitability features (up to 2.5 pts)—real P&L from leaderboard (ROI component + magnitude component), camouflage score from fill-level analysis, reliability-gated by volume.
Wallets with real leaderboard data use Polymarket’s pre-computed P&L directly. For wallets not on the leaderboard, the pipeline falls back to the whale-entry estimate (which undercounts, but is better than nothing).
The watchlist’s profit profile
| Wallet | Est. P&L | ROI | VW Acc | Cam. Score | Top-4 Share |
|---|---|---|---|---|---|
| 0x3cf3...87b3 | $702,727 | +66.1% | 61.2% | 4.10 | 78.6% |
| 0x7c3d...5c6b | $377,379 | +121.7% | 75.6% | 2.88 | 91.9% |
| 0x9e91...3f8e | $114,510 | +95.7% | 96.8% | 0.23 | 62.3% |
| 0xfffe...f52c | $93,903 | +64.0% | 100.0% | 0.00 | — |
| 0x7744...999e | $53,932 | +74.6% | 78.0% | 1.81 | 75.5% |
| 0x4bfb...982e | $46,741 | +0.3% | 80.3% | 0.21 | 10.0% |
| 0xd5cc...dea4 | $46,827 | +14.2% | 97.5% | 0.23 | 85.7% |
| 0x4dfd...6e4a | $46,657 | +48.5% | 100.0% | 0.00 | 58.4% |
| 0xe64f...44a5 | $37,263 | +8.8% | 89.5% | 0.39 | 35.4% |
| 0x0afc...30b2 PRESS-REPORTED | $154,219* | +14.6% | 52.5% | 2.65 | 61.9% |
Table shows 9 of 15 profitable watchlist wallets. P&L figures for wallets on Polymarket’s leaderboard use the platform’s own pre-computed profit. For other wallets, P&L is estimated from resolved buy fills in the whale-entry dataset (which undercounts; see Limitations). VW Acc = value-weighted accuracy. Cam. Score = camouflage score (high = concentrated profit, mediocre accuracy).
The wallet of interest: #1 globally with $154K profit
With real profitability data, 0x0afc (@Rundeep) ranks #1 out of all 7,655 wallets on combined behavioral + profitability scoring. Polymarket’s leaderboard confirms +$154,219 in all-time profit (rank #696 on the platform), with a camouflage score of 2.65—one of the highest on the watchlist. Its top 4 fills account for 61.9% of all gross profit. The aggregate buy accuracy of 52.5% is a textbook camouflage signature: most bets are ordinary, but the few high-conviction trades carry outsized returns.
The profit is concentrated in four Israel–Iran military escalation markets:
| Market | Shares | Entry | Profit | ROI |
|---|---|---|---|---|
| Israel military action against Iran by Friday? | 162,664 Yes | 21¢ | $128,737 | +379% |
| Israel strike on Iran on June 24? | 72,684 Yes | 76¢ | $17,706 | +32% |
| Israel end of military operations against Iran before July? | 9,184 Yes | 33¢ | $6,195 | +207% |
| Israel strikes Iran by January 31, 2026? | 30,049 Yes | 21¢ | $3,122 | +50% |
| Total (4 winning positions) | $155,760 |
The pattern is unmistakable: deeply contrarian entries (buying YES at 21¢ when the market priced the event as unlikely), concentrated in a single geopolitical theme, generating outsized returns. The pipeline’s camouflage detector flags exactly this pattern—where a few massive winning trades are embedded within many ordinary-looking fills (128 total, 52.5% aggregate accuracy). In market surveillance, this is known as “noise trading”: many ordinary bets diluting aggregate statistics while a handful of high-conviction positions generate the real returns.
High-value wallets flagged by the pipeline
With real P&L from the leaderboard, we can now see which behaviorally flagged wallets are genuinely profitable—and the numbers are striking. Several wallets in the pipeline’s top 30 have confirmed profits in the hundreds of thousands to millions of dollars:
| Rank | Username | Real P&L | Pattern |
|---|---|---|---|
| #2 | bobe2 | $1,327,633 | 100% size-weighted buy accuracy, 8 markets, $127M volume |
| #14 | scottilicious | $1,402,936 | 17 markets, $636K whale entries, cross-market repeat |
| #17 | theo5 | $880,290 | 17 markets over 2+ years, $97M total volume |
| #18 | SlavaUkraini | $648,634 | 99.8% size-weighted accuracy, 12 markets, $30M volume |
The pipeline independently flagged these wallets on behavioral anomalies alone—unusual concentration, cross-market repetition, and timing patterns. The leaderboard confirms they are collectively worth over $4M in profits. This cross-validation between behavioral signals and financial outcomes is the pipeline’s strongest result.
What This Study Cannot Tell You
This methodology has substantial limitations.
1. We see large trades, not full portfolios
The whale-entry detector captures only large or concentrated entries. A wallet could appear to have a low buy accuracy here while actually being highly profitable—if most of its gains came from smaller trades not captured in this dataset, or from exits that aren't tracked.
2. Correctness is not profitability
A wallet that buys winners at 99 cents (1% upside) and a wallet that buys winners at 10 cents (900% upside) both score 100% on buy accuracy. This version of the pipeline addresses this by pulling verified P&L from Polymarket’s leaderboard API where available. For the 5,000 wallets on the leaderboard, profitability is confirmed directly. For the remaining wallets, the pipeline falls back to whale-entry estimates—which can undercount by an order of magnitude (the wallet of interest’s whale-entry estimate was $8K vs. its confirmed $154K).
3. The markout backfill is selection-biased
We backfilled price data for the top 100 markets by watchlist trading volume. This is like studying whether a poker player is good by only looking at the hands they chose to play—it enriches for situations where they were active and confident, possibly overstating edge.
4. Multiple explanations exist
High correctness and positive markout could result from insider information, but also from: superior analytical models, fast OSINT (open-source intelligence) plus domain specialization in geopolitics, API-driven reaction to public headlines before manual traders reprice, cross-market arbitrage, cross-venue basis trading (Polymarket vs. Kalshi vs. bookmakers), strategic wallet segmentation for operations/security reasons, or simple data artifacts. The strong geopolitical clustering in the watchlist's top markets is equally consistent with "informed insiders in geopolitically sensitive events" and "active macro traders in the biggest, most liquid markets." Without a proper null hypothesis test against matched high-volume geopolitical specialists, we cannot distinguish these explanations. This pipeline identifies statistical anomalies. It does not prove that any wallet possesses non-public information, nor can it establish identity or intent.
5. The sell-trade problem
When a wallet sells a winning outcome token, we code it as "incorrect." But the wallet may have bought that token earlier at a lower price and is profitably exiting. Without full position tracking, sell-side accuracy is unreliable. This is why buy-side accuracy (+14.6 pp above baseline) is the stronger signal.
6. No temporal holdout
The pipeline was applied to all available data (2023–2026) without a train/test split. We do not know whether the wallet signatures identified here would persist in future time periods, or whether they are artifacts of specific historical events.
What the Needle Tells Us About the Haystack
An algorithm, given no hints about which wallet to investigate, ranked a previously press-reported address as the #1 most anomalous wallet out of 7,655 on combined behavioral and profitability scoring. It made the watchlist’s liquid top-20 on behavioral signals alone. Verified profitability from Polymarket’s leaderboard API confirmed it: +$154,219 in all-time profit, a high camouflage signature, and returns concentrated in four contrarian Israel–Iran bets—all hidden behind 52.5% aggregate accuracy.
This is not just a single-wallet result. Across the full watchlist, profitability analysis confirms that 15 of 20 flagged wallets are profitable on their resolved buy trades, with a combined estimated P&L of $1.3M. The profitability features also surfaced wallets that behavioral scoring alone underranked—including one with $702K in estimated profit that jumped 93 positions from #103 to #10.
The two-layer approach—behavioral anomaly detection followed by profitability confirmation—is more reliable than either layer alone. Behavioral features catch unusual trading patterns regardless of whether the wallet has resolved markets. Profitability features validate whether those patterns translate to actual returns, and the camouflage detector identifies wallets that disguise high-conviction bets inside mediocre aggregate statistics.
The caveats remain real. The watchlist aggregate performance is still heavily influenced by one high-volume wallet. No formal statistical tests separate signal from noise. The liquidity filter thresholds are pragmatic, not principled. And for wallets not on the public leaderboard, profitability estimates still rely on whale-entry fills only—which can undercount by an order of magnitude. For the wallet of interest, the leaderboard confirmed the real figure ($154K), but many other flagged wallets lack that external validation.
What should happen next? Full-position reconstruction would close the remaining gap between estimated and actual P&L. Entity clustering could determine if multiple watchlist wallets belong to the same operator. Permutation tests would establish whether the observed patterns are statistically distinguishable from chance. A temporal holdout would test whether identified signatures persist out of sample.
The transparency of blockchain data cuts both ways: it makes surveillance possible, but it also means that sophisticated actors can study detection methods and adapt.
If your question is “are these wallets worth monitoring?”—this analysis says yes, and now says why: they are collectively profitable, several show camouflage-like profit concentration, and the previously reported wallet is a confirmed $154K winner hiding behind coin-flip aggregate accuracy. If your question is “have we found insider trading?”—this analysis identifies anomalies consistent with informed trading, but cannot prove non-public information was involved. The right next step is deeper investigation, not a verdict.
Reproducibility
All data artifacts from this run are preserved in the project repository. The analysis can be replicated by loading the following files:
| Artifact | Description |
|---|---|
| claude_blind_universe_evidence_pack_v3.md | Full evidence pack with all intermediate calculations |
| blind_universe_market_winners.csv | Resolved market outcomes (960 markets) |
| claude_blind_universe_memo_v3.md | Committee-grade analysis memo |
| blind_universe_resolution_validation_summary.json | Resolution correctness aggregates |
| blind_universe_wallet_resolution_metrics.csv | Per-wallet resolution accuracy |
| watchlist_markout_backfill_summary.json | Markout backfill coverage and aggregate results |
| watchlist_markout_metrics.csv | Per-wallet markout metrics at 2-day and 7-day horizons |
| watchlist_ohlcv_1d_backfill.csv | Daily price data for 81 backfilled markets (4,173 rows) |
Run ID: blind-universe-whales-20260215-032746.
Generated: 2026-02-16T03:38:14Z.
Glossary
- Wallet
- An on-chain Ethereum address (a hexadecimal string like
0x0afc...30b2) that executed trades. Wallets are pseudonymous—anyone can create one, and one person can control multiple wallets. - Market
- A prediction market on Polymarket, identified by a unique condition ID. Each market poses a yes/no question (e.g., "Will Israel invade Lebanon in September?") and resolves to a definitive outcome.
- Fill
- A single observed trade record in the dataset. This dataset contains only whale-entry fills, not a complete trading ledger.
- Whale Entry
- A pipeline signal flagging a trade as large or concentrated relative to the market. Triggered by position size, market share, or timing thresholds. Not every trade by a given wallet triggers this signal.
- Liquidity Filter
- A minimum-quality threshold requiring markets to have at least $50,000 in observed trade value and at least 20 unique wallet participants. Applied to remove micro-market noise from rankings.
- Resolution Correctness
- A measure of how often a wallet's trades aligned with the eventual market outcome. A buy is "correct" if it purchased the winning outcome; a sell is "correct" if it sold a losing outcome. Crucially, this ignores entry prices and cannot distinguish profit-taking from directional bets on the sell side.
- Markout
- The sign-adjusted change in a market's YES token price between a trade's entry price and the daily close price at a fixed future horizon (2 or 7 days). A positive markout means the price moved in the trade's favor. Value-weighted by trade size. Not equivalent to realized profit-and-loss.
- Late Value Share (48h)
- The fraction of a wallet's total trading value in a given market that occurred within the last 48 hours before that market's final observed trade. A high value (e.g., 85%) means the wallet concentrated its activity near the end of the market's life—a potential indicator of informed timing.
- Market Value Share
- The fraction of a market's total observed whale-entry trade value attributable to a single wallet. A wallet with 45% market value share dominated nearly half of all large trades in that market.
- OHLCV
- Open/High/Low/Close/Volume—daily price summary data for a market's YES token. In the targeted backfill, volume and trade count fields may be missing because the upstream API provides price points only.
- Effective Label
- The pipeline's combined classification of a wallet-market pair, merging global behavioral patterns (cross-market activity) with market-local patterns (concentration in a single market). The highest-priority label is "priority: insider-like, market and global."
- Basis Point (bp)
- One hundredth of one percentage point. 100 basis points = 1 percentage point. Used to express small price movements.
Appendix: Detailed Wallet Breakdown
The table below presents the full scoring and performance breakdown for each of the 11 watchlist wallets. It consolidates pipeline classification, resolution correctness, and forward price drift into a single reference. All wallets carry the global_cross_market_repeat global label, meaning each was flagged for concentrated, well-timed trading across multiple unrelated markets. Only 304 of 7,655 wallets (4.0%) received this label.
Reading the cards: The summary row shows five key metrics at a glance. Click to expand for the full breakdown across four categories: classification (label, local pattern, priority score), best liquid market (size, participants, value share, timing), resolution correctness (trade counts, accuracy), and markout (2d/7d forward price drift, fills, win rate). Priority = insider-like on both local and global dimensions. Watch = cross-market network pattern only. Green values are above the 66.0% baseline; red values are below.