What 470 million trades reveal about Polymarket farmers

If you've spent any time on Polymarket leaderboards, you've seen the wallets at the top with 97% win rates over 1,800 resolved markets and lifetime P&L of $4,200. You probably wondered how someone trades that frequently and accurately for so little money. The answer is they aren't trading. They're farming.

Farmers are wallets running automated patterns to extract token incentives, fee rebates, or reflexive market dynamics. They look like high-performing traders to any system that reads naive on-chain data. They contaminate every leaderboard. They are the single biggest reason copy-trading on Polymarket leaderboards systematically loses money — you copy what looks like edge but is actually a strategy that requires bot infrastructure and operates on tiny margins you can't replicate.

Across 470M trades indexed from the Polygon CTF contract, roughly 16% of wallets that pass naive "smart" thresholds (positive lifetime P&L over 50+ resolved markets, win rate above 60%) are farmers. Filtering them out is what separates a useful smart-money signal from an inflated leaderboard.

This piece walks through the four farming archetypes we observe, the heuristics that catch each, and what the on-chain pattern tells you about why the platform's reported numbers diverge from ours.

The four archetypes

Archetype 1: Token-incentive farmers

Polymarket has historically run volume-rewards programs that pay liquidity providers in tokens for placing limit orders. Farmers respond by:

• Placing limit orders inside the spread on near-resolution markets (e.g. 0.97 / 0.99 on a market trading at 0.98)
• Letting the orders fill against the natural flow
• Rolling positions repeatedly across many markets simultaneously

The trade-level signature is straightforward: average position entry between 0.95 and 0.99 or between 0.01 and 0.05, average per-market profit under $30, hundreds of markets per week. The wallet is hoovering up sub-cent rebates while showing up on every leaderboard with phantom 98% win rates because, technically, every market resolves the way the wallet was positioned (because it was positioned on near-certainties).

Detection: combine entry-price distribution + per-market dollar profit + trade frequency. Wallets that fail any one of those three filters are dropped from the smart-money cohort.

In our data, the largest token-incentive farmer cluster has 1,872 wallets (likely controlled by a much smaller number of operators) responsible for ~3.4% of all platform volume. Without filtering, they would dominate the win-rate leaderboards.

Archetype 2: Mint-and-merge arbitrageurs

Polymarket's CTF contract supports two atomic operations beyond regular buying:

• MINT — pay $1 USDC, receive 1 YES token + 1 NO token. Equivalent to buying both sides at $0.50 each.
• MERGE — return 1 YES + 1 NO, receive $1 USDC. Equivalent to closing both sides at $0.50 each.

If the binary YES + NO order book trades at 0.49 + 0.49 = $0.98 (because of a momentary order-book imbalance), an arbitrageur can:

1. Buy YES on book at $0.49
2. Buy NO on book at $0.49
3. MERGE YES + NO into $1.00 USDC

Net gain: $0.02 per pair across however many pairs the arb supports. This is a pure on-chain arbitrage with no directional view. In our data, arbitrageurs running this loop typically rotate dozens of markets per minute, holding positions for an average of 47 seconds before settlement.

Naive systems read these wallets as having close to 100% win rate (every position closes profitable, by definition) and astronomical trade frequency. They're not predicting anything; they're cleaning up market microstructure noise. The detection is via the entry/exit timestamp delta — any wallet with median holding period below 5 minutes across 100+ markets is automatically reclassified.

Archetype 3: NegRisk converters

NegRisk is the multi-outcome market wrapper Polymarket uses for things like "Who will win the 2028 GOP nomination?" (10 candidates, 10 binary sub-markets). The CTF supports a CONVERSION operation specific to NegRisk: trade between sister markets without fully closing.

A wallet can convert "Trump YES" into "DeSantis NO" (the sister market in the same NegRisk family) without paying market spread on either leg. This is useful for hedging and rebalancing, but it gets weaponized by farmers to manufacture infinite resolution counts across linked markets.

The naive on-chain interpretation: the wallet "won" both the Trump position AND the DeSantis position simultaneously. Win rate spikes; resolved-market count balloons. We see wallets with 4,200 "resolved markets" in our index, of which only ~340 are actually independent decisions; the rest are NegRisk conversion-driven counts.

Our consolidation logic groups all NegRisk sister-market settlements per underlying event into one decision, regardless of how many on-chain settlement events fired. After consolidation the same wallet's "real" resolved-market count drops from 4,200 to ~340, and win rate moves from 97% to 71%. The 71% is the meaningful number.

Archetype 4: Bot wallets running scripts

These are wallets that don't even pretend to be farming for incentives — they're running directional bots responding to on-chain or off-chain triggers, with no human in the loop. Their signature:

• Trade timing distribution: clustered at exact-second timestamps suggesting time.sleep(60) loops, or millisecond-level reactions to specific events
• Trade-size distribution: bimodal — either tiny ($1-50) or specific round numbers ($1,000.00 / $5,000.00) repeated identically
• Ratio of cancelled to filled orders: often above 60% (real traders are below 15%)
• Activity over time: continuous 24/7 with no human-pattern dropoff during sleep hours

Bot wallets aren't always unprofitable — some are quite well-engineered — but their performance is not transferrable to a human copy-trader because the edge is in the latency or signal source, not the directional view. Our filter flags ~3,400 wallets as bots in this archetype. They're separated from the smart-money cohort because their signal is operational, not informational.

What the unfiltered leaderboard looks like

If you ran a naive smart-money filter on Polymarket data — lifetime P&L > $200K, win rate > 60%, resolved markets > 50 — you'd produce a leaderboard of ~14,200 wallets. Of those:

So roughly 1 in 7 wallets at the "top" of any naive Polymarket leaderboard is a farmer or bot whose signal is not transferable to human trading. If you copy-traded the top 100 unfiltered, you'd be following 14-15 wallets that aren't really making predictions, dragging your performance down toward their structural margin (which is sub-1% per trade).

How we filter

The composite filter is a multi-step pipeline applied to every wallet that crosses the lifetime-P&L threshold:

1. Entry-price distribution check: median entry price must fall between 0.15 and 0.85. Wallets clustered above 0.95 or below 0.05 are flagged token-incentive farmers.
2. Holding-period check: median holding period must be above 6 minutes. Wallets below this are flagged arbitrageurs.
3. NegRisk consolidation: all sister-market settlements grouped to underlying decision.
4. Per-market profit check: median dollar profit per resolved market must be above $50 (raised from $30 to filter dust trading).
5. Bot signature check: order-cancel ratio + timestamp clustering + trade-size repetition.
6. Category-balance check: a wallet that has only ever traded one specific market repeatedly (e.g. 800 trades on the same Bitcoin-price market across multiple monthly resolutions) is flagged as a strategy bot, not a discretionary trader.

Each check has its own threshold tuned on labeled data. A wallet failing any one check is excluded from smart-money classification. The aggregate effect: 16% of naive-pass wallets get reclassified, and the smart-money win rate moves from a misleading 78% (with farmers) to a more honest 67% (without).

Why this matters

Two practical consequences:

For copy-traders: you cannot copy a farmer or arbitrageur, even if they show up at the top of a leaderboard. Their edge is in operational infrastructure (high-frequency, low-margin, 24/7 automation) that requires capital + tech you don't have. Following them produces near-zero returns offset by gas costs and platform fees. Stop using unfiltered leaderboards.

For traders calibrating their own performance: the meaningful peer group is the 12,000 wallets that pass farmer filtering, not the 14,200 that show up on raw lists. Comparing your win rate to a leaderboard average that includes 9% token-incentive farmers is comparing yourself to people running an entirely different strategy.

OrcaLayer's leaderboard (orcalayer.com/leaderboard) applies the full filter pipeline before ranking. Wallet pages (orcalayer.com/wallet/0x...) show the per-wallet breakdown — including the farmer flags and the consolidated NegRisk-adjusted history — so you can see exactly why a wallet is or isn't classified as smart.

The 470M trades on Polymarket are a remarkable dataset. Without farmer filtering, they tell you noise. With it, they tell you who is actually predicting the world correctly with money on the line. That second signal is the entire point of the platform.