25 Directions Tested, 4 Survived: A Polymarket Strategy Post-Mortem
Over the past three months, I tested twenty-plus automated trading directions on Polymarket. Whale following, latency arbitrage, news event trading, weather forecast arbitrage, mean reversion, market making, settlement harvesting—basically everything I could think of.
About a dozen made it to live or paper trading. Most got killed during the research phase.
Only 4 survived.
This is a post-mortem. I'll detail the failures—because the anti-patterns themselves have value. They'll help you skip some seductive dead ends. But the survivors won't get much explanation. Reason should be obvious.
Fair warning: this is all provisional. A few months from now, I might be eating my words. Markets evolve, so does thinking.
The Graveyard
Whale Following: The Most Time-Burned Stupid Idea
This consumed more effort than any other direction. Five iterations, all dead on arrival.
First attempt: Large order tracking. Monitor taker orders over $1,000 and follow. Real-world win rate: 42.9%. Might as well flip a coin.
Undeterred, I filtered 287K addresses for the best historical performers and built an elite pool for real-time shadowing. Found two fatal issues.
First: Third-party data sources botch win rate calculations. SPLIT operations underestimate costs. Redeemed positions get incorrectly flagged as "wins." Analyzing 300K addresses, plenty showed 80%+ win rates—until I cross-referenced them against the Polymarket official Data API. Numbers cratered. One address claimed 83.5% win rate. Actual: 50.9%.
Second and more fundamental: many top "whales" are actually market makers. They might hold both YES and NO on the same market. Or hedge across accounts. You see them buying YES and think they're bullish. Actually they're just providing liquidity. You follow them in, and you're really just the other leg of their spread.
Tried Copy Trading variants, computed information coefficients—came to zero. Ran shadow experiments too. Live trading: +$770. Shadow trading: -$148. Directions flipped entirely. These whales' past performance tells you nothing about the future, and you can't even be sure what they're doing.
Five iterations. One road. All dead ends.
BTC 5-Minute Market: Five Ways to Fall Flat
The BTC 5-minute market on Polymarket has some of the best liquidity around. I threw a lot at this. Tried five different plays.
Latency arbitrage. Early PM had orderbook update lag. There was a window after Binance moved where you could front-run. Hit live trading and got slapped with 75.9 basis points of adverse selection—those orders I snagged were mostly deliberate baits left by the maker. Real value already gone. Lost $6.34 and stopped immediately. Later, PM tightened the taker delay config. Window keeps shrinking.
Taker momentum. Follow Binance price direction with FOK orders. Ran 294 live trades. Started with 81% win rate, decayed to 63%. The real killer wasn't the win rate—it was the payoff structure. Win a little, lose a lot. Risk-reward ratio: 0.34x. Add 1.5% taker fees and the whole thing nets $39. Edge too thin to bother.
Reverse mean reversion. Market goes up 5 minutes, so fade it. Result: 4 wins, 10 losses. 29% win rate. Momentum beats mean reversion at this timeframe. Completely wrong direction.
CVD signal. Used cumulative volume delta. Tried momentum and counter-trend versions. One went 5-1 then never fired again. Other went 1-4 then went silent. Too few signals, not enough data to validate anything.
Five plays, zero survivors. The players here are too sharp—pros and quant bots everywhere, crumbs for retail.
Tail Buy-In: 0.2% Win Rate
Buy contracts under $0.05. Bet on low-probability events. Payoffs 20:1, looks juicy.
Tested it: 0.2% win rate. Break-even needs 2.4%. Market prices tail risk more accurately than you'd expect.
Settlement Harvesting: 100% Win Rate, Lost Money
Many PM markets settle through UMA's optimistic arbitration. When the outcome's certain, contract near $1.00 but not yet there—buy and wait for the settlement premium. Should print.
Did two versions. Loose: entry $0.97+, 22 trades, 9-0 record, made $5.09. Strict: entry $0.9975+, 20 trades, 13-0 record, lost $2.32.
Perfect win rate. Every single trade won. Still lost money. At $0.9975 entry, profit per trade under 3 cents. Taker fee 2% kills you. My favorite "corpse": Win rate and profitability are completely different things.
Soccer FLB: Curse of Small Sample Size
Favorite-Longshot Bias is textbook sports betting mispricing—favorites undervalued, longshots inflated. Built a strategy on top-5 European league data: buy favorites in the 0.50-0.85 band.
116 signals, 86 held, 4-9 record.
Maybe the strategy's fine. The problem is soccer: one game a week. Three months? Sample size nowhere close to conclusive. Crypto markets cycle every 5 minutes—three orders of magnitude denser. Plus, PM's sports markets are weird. Liquidity, settlement, none of it looks like traditional books. Can FLB even translate here? Big question mark.
Rebate Farming: You Need Enough Capital
Polymarket pays makers zero fees plus rebate. Built a strategy to spam orders on the 5-minute market and collect rebates.
Redemption cooldown: 30 minutes. Candle cycle: 5 minutes. My $217 can't survive continuous reinvestment. Orphan order ratio exploded. Minimum viable: $400. Sweet spot: $800+. Final loss: $165.75.
Not enough money, can't play. Simple as that.
Market Making: Swimming Naked in a Red Sea
One of my earliest directions. Dual-sided quotes on the Rihanna album release market.
Worked, but competition crushes it—pros faster, bigger banks, better pricing. Market closed, strategy went with it. Later research confirmed: 45% of PM crypto volume is bot-driven. Market making is pure arms race.
News Event Trading: Chain Too Long
Monitor news → AI judges impact → auto-execute. Sexiest direction.
Two iterations. Core problem: from news hit to signal generation to risk management to fill, the pipeline's too long. By the time your order lands on the CLOB, market's already priced it in. Speed and accuracy both matter. Current tech stack can't do both.
Didn't die completely, but needs infrastructure overhaul. Shelved it.
Killed Before Code Was Even Done
Some directions got axed in research:
CTF mint arbitrage. Theory: token mint/split combos create arb windows. Checked 12 markets. CLOB bid-side totals all under 0.04. Taker mode: -97%+ loss.
Binary arb. Someone wrote that arb windows hit 4-8%. Reality: 1-3 seconds, then gone. Needs HFT infrastructure.
BTC daily direction. Analyzed one address, 5,036 trades, 52.6% win rate, Z=1.46, p=0.072. Not statistically significant. No edge.
SPLIT arb. SPLIT→SELL→REDEEM chain. Reverse-engineered an address claiming $38K profit. Leaderboard says: actually down $611.
The Survivors
The strategies that made it—I won't detail params and specifics. Smart money knows why. But some directional feels:
Weather markets are running at break-even right now. Hit the season-transition gotcha, fixed it, been running steady ever since.
Crypto up/down markets are still iterating. Biggest lesson: regime dependence. Win rate looks great in one regime, tanks in another. Knowing when to stop beats knowing when to start.
Niche categories—social media prediction markets for instance—actually generate the most profit. Fewer players, worse pricing, limited capacity but stable.
Also experimenting with the "one strategy, multiple accounts" model to scale returns.
Made some spare change. Light-years from "this pays the bills." But the direction's validated. Now it's just scaling.
The Pattern in Death and Life
Twenty-plus "corpses" show five main causes of death:
Market efficiency trap. Latency arb, whale following, tail buying—all betting on market mistakes. Markets evolve, bots get sharper, your "inefficiency" probably expires. Today's edge is tomorrow's common knowledge.
Payoff structure failure. UMA settlement is the poster child. Win rate means nothing if profits don't cover costs. Slow bleed-out. Ignore win rate. Watch payoff.
Insufficient sample size. Soccer FLB, CVD signals—might be right, but three months proves nothing. Time granularity and capital efficiency just don't sync. Crypto runs 5-minute cycles; you need years to validate slow markets.
Signal chain failure. News event trading—pipeline too long. Speed and accuracy need to both work. Usually can't.
Data illusion. Whale following—your "alpha" might just be statistical noise from the settlement mechanism.
The survivors share one trait: your opponent isn't "the market"—it's casual traders and small players in specific categories. Finding weak opposition beats finding a clever algorithm.
One more thing: no strategy runs on autopilot. No monitor, no regime check = ticking bomb.
Closing Thoughts
92% of traders on Polymarket lose money. Bots capture 73% of arbitrage profits.
I explored twenty-plus directions. Most got shelved. Three months of time cost way outstrips current returns.
But compared to the start? Light-years better. Couldn't execute a single strategy at first. Now I have a few running stable. These "corpses" taught me things no textbook covers—about market efficiency, cost structures, data ghosts. Each shutdown clarifies what works and what doesn't.
Direction's proven. Given time and data, I believe this scales. Prediction markets are still early. Pricing efficiency is rising but nowhere near efficient-market territory. Opportunities exist—just not where you think they are.
Want to try these strategies on Polymarket?
Sign up for Polymarket →