Polymarket Strategies That Don't Work

We replayed market making on every Polymarket football (soccer) 1X2 market for 13 days. It lost $23,055.

That was one of four tests. We took the most popular ideas for making money on Polymarket and ran each one against real data: a recorded order book for every football market since September 17, on-chain trades from our own indexer, a live arbitrage detector, and 35,000 closed markets. None of them survived execution.

All data and the full write-ups are public in the Polymarket arbitrage, market making and copy trading backtests repository on GitHub.

Results: 4 Polymarket strategies tested

Two of them, arbitrage and copy trading, die on the same detail: the 1-second taker delay on sports markets.

Polymarket arbitrage: 369 candidates, $1.83 takeable

A football match on Polymarket has dozens of logically linked markets. "Team A wins" + "Draw" + "Team B wins" must sum to 1. "Nobody scores" is the same event as "total under 0.5". A half total can never exceed the match total. When those relations break, you buy all legs and lock in a profit.

Our live detector ran for about 18 hours and found 369 candidates across 35 events. 262 were confirmed by an immediate REST check, so these were real prices, not feed glitches. The most striking one: a REST check found one team's Yes offered at 0.34 when fair was around 0.55, 1,172 shares on that leg, about 20 c of edge per set after fees. The draw leg refilled 0.2 seconds later, so up to ~400 full sets were available for about half a second. By one second the 0.34 offer was gone.

The catch is a field called secondsDelay. Sports markets on Polymarket carry it: 1 second for football, 0 for NBA. Per the Polymarket docs, a marketable order waits out that delay and "cannot be canceled" while pending. Resting limit orders can still be cancelled. In that second, makers pull their quotes or faster takers grab the legs.

• All 30 candidates with 2 c+ of edge were in 1-second-delay markets.
• We checked 28 of them one second later. All legs were still in the book in 3, worth $1.83 combined.
• In 19, only some legs were left, which turns "risk-free arbitrage" into a $137 naked bet.
• Before kick-off there is no delay, and everything would have filled. Total edge there: $1.55 at up to $10 per leg. Both numbers are pocket change.

Of the 41 legs that vanished, other takers took 24 and makers pulled 17. Even someone who won every single race would have made about $14 over the same ~18-hour run, at up to $10 per leg.

Polymarket market making in live football: -$23k in 13 days

Makers pay no fees on Polymarket and get a 15% rebate on sports. Quoting both sides during a match and collecting the spread sounds like free money.

We replayed it on our recorded order book, which contains 99.9% of the trades that later show up on-chain. The model was conservative: 50 shares per quote, a 200-share position cap, back of the queue, 0.3 s requote, with and without a kill-switch on goals.

Over 13 days it lost $23,055 with the kill-switch and $24,110 without. Faster requoting, sticky quotes and quoting 1 c inside the spread were all negative too.

The reason is adverse selection. Here is every maker fill on the market over a week, not just ours:

Makers at the front of the queue earn +0.04 c in normal play. If you are at the back, you mostly get filled when price is about to move through your level, and most of our replay losses came in normal play, not just around goals.

Polymarket copy trading: insiders exist, copying them fails

This one starts with a mistake we almost published.

We took wallets that bought before goals and simulated copying them 1-5 seconds later. The result was +$4.23 per $10. Then we noticed we had selected exactly the buys that were followed by goals, which you cannot know in real time. Re-tested on every taker buy by those wallets over a week, without knowing the outcome, it came to -$0.41 per $10. If your selection uses what happened next, the backtest is not telling you anything.

Done properly (pick wallets on one period, test them on the next week), 92 wallets made +$2.33 per $10 at their own prices, with t = 6.8 over 10,198 buys, so some participants really do see goals before the market. Copy them 1-5 seconds later and you get -$0.10 per $10.

We looked at one of the best of these wallets trade by trade. 92% of its buys were maker fills, which have no exchange delay. In a quarter of cases its fill was the last trade at the old price, and the ask jumped 9 c in the same millisecond.

Polymarket calibration: is the favorite-longshot bias tradable?

The favorite-longshot bias is a classic in sports betting. If Polymarket longshots are overpriced, you can fit a correction and trade the difference.

We took 35,135 closed markets with trades, split them by time (train before August 11, test after), and fit a Platt correction at 24 h, 6 h, 1 h and 10 min before the last trade.

The correction moves Brier in the fourth decimal, and across all markets the rule makes about zero. There is some bias in the last 10 minutes (markets at 0.05-0.10 resolved Yes only 1.9% of the time), but the amounts are tiny and do not survive fees. This is consistent with the AIA Forecaster report from Bridgewater AIA Labs, where an LLM forecaster scored worse than liquid-market prices (Brier 0.1258 vs 0.1106).

Polymarket bot pitfalls: taker delay, fees, malicious repos

The repo has a separate file with the exchange mechanics we measured:

• taker delay on sports markets, and 250 ms on some crypto and finance up/down markets
• the fee formula: C x feeRate x p x (1 - p), most expensive at 0.50
• the order book WebSocket lags by about 17 ms, on-chain settlement by about 2 s, so the Polygon mempool gives you no head start
• Gamma API pagination breaks with offset after roughly 2,100-2,500 records
• fake "Polymarket bot" repos and npm packages that steal wallet keys, documented by StepSecurity and SafeDep

FAQ

Does Polymarket arbitrage work? Not in live sports. Our detector found 369 candidates in 18 hours. After the 1-second football delay, all legs were still in the book in only 3, worth $1.83.

Why can't you take Polymarket sports arbitrage? Marketable orders wait out secondsDelay (1 s for football) and cannot be cancelled while pending. Makers can cancel their quotes during that second, and faster takers can take the legs first.

Is market making on Polymarket profitable in-play? Not for a queue-back maker. A 13-day replay lost $23,055, with 12 of 13 days negative, because you mostly get filled right before the price moves against you.

Can you copy-trade Polymarket insiders? Not in live football. The wallets made +$2.33 per $10, copying them 1-5 s later made -$0.10. We only tested live football on a horizon of seconds.

What we did not test

This rules out copying informed money in live football, where the edge lasts seconds. It says nothing about markets where news takes hours or days to get priced in, like politics or long-dated markets. That is what our smart-money leaderboard tracks, and it is the next thing we want to measure. For how we separate real smart money from noise, see why a 99% win rate is a trap and our methodology.

We built most of this expecting at least one idea to work. None did, so here are the numbers. The data is CC BY 4.0: rerun it, extend it, and tell us where we got it wrong.

This is research, not financial advice. Results reflect our setup and period (September to October 2026). Polymarket changes fees and delays, so conclusions can go stale.