Three Polymarket backtest traps that create a fake edge
Empty-book mid prices summed to 3.0 against 1.03 on liquid events. Selecting by final volume lifted a 0-0 rate from 5.5% to 16-22%. Using leaky timestamps showed gains of +126% and +$250 that vanished once fixed.
What we asked
While testing two ideas on historical Polymarket data, cheap outcomes just before resolution and temperature markets priced against public weather forecasts, we got attractive results that turned out to be false. Each came from a data trap, not from the market. This article lists three of them, the false number each produced and the number after the fix.
The tail test covered 12,561 closed markets from 1 September to 13 October 2026. The weather test covered 1,977 closed daily temperature events across 51 cities, from 23 September to 14 October 2026, with 21,747 baskets (outcomes) and $46.5 million of volume.
Trap 1: the mid price of an empty book
The mid price is halfway between the best bid and the best ask. On a basket nobody trades, the best bid can be 0.001 and the ask far higher, which puts the mid at 0.3-0.5. Nobody can trade at that price.
For a temperature event, the mids of all baskets should sum to about 1. We measured 3.0 on events with under $5,000 of volume and 1.03 on events with $50,000 or more. Our first weather run, which treated those mids as prices, showed +$2,000 on buying NO. After limiting the test to events of at least $20,000 and baskets of at least $2,000 (about 960 events), the +$2,000 disappeared and market prices proved well calibrated: a price of 0.25 corresponded to a 30% win rate, and 0.54 to 55%.
Trap 2: selecting markets by final volume
In the tail test, we kept only markets with at least $10,000 of volume. Among football Under 0.5 markets (the bet that the match ends 0-0), that gave a 0-0 rate of 16-22% at prices of 4.5-7.5 cents. But volume accumulates during the match, while the score is still 0-0, so a high final volume partly selects matches that stayed 0-0.
Without a threshold, 0-0 occurred in 1.9% of markets under $1,000 of volume, 6.4% between $1,000 and $10,000, and 15.9% at $10,000 or more. In the unselected pool of 2,392 markets one hour before kickoff, it was 9.7% at 6.8 cents. In markets where buyers had traded, it was 5.5% at 6.3-6.5 cents, an expected value of -11% to -15%.
Trap 3: timestamps that sit inside the event
Two examples.
For esports markets, the recorded end date was 6 hours after the match started, so a snapshot taken "before the end" fell inside the game. Together with a second error, that most of the wins were Under 0.5 outcomes (0-0 draws), this produced sports tails at 5 cents winning 12%, an expected value of +126%. Measuring from the scheduled start instead, match outcomes priced 3-6 cents won 4-5%, with an expected value of about 0 to -20% after a 1 cent spread.
For weather, we used the archived "latest forecast run" for a past date. That run includes forecasts issued after the moment of entry. A same-day forecast showed +$250, per $1 staked on each trade. We discarded it. The comparable test with the previous day's forecast, entering six hours into the observation day, lost money: -$350 (error margin 69) buying YES, 721 trades, and -$72 (29) buying NO, 1,378 trades.
The numbers side by side
| Trap | False result | After the fix |
|---|---|---|
| Empty-book mid | Sum of mids 3.0; +$2,000 on NO | Sum 1.03 on liquid events; market calibrated |
| Final-volume selection | 0-0 in 16-22% of markets at 4.5-7.5c | 5.5% at 6.3-6.5c with buyers present; -11% to -15% |
| Esports end date | +126% at about 5c | Match outcomes at 3-6c win 4-5%; about 0 to -20% |
| Forecast run after entry | +$250 | Excluded; previous-day forecast -$350 and -$72 |
Limits of this measurement
-
The weather figures depend on a model of the forecast error calibrated by city, and the forecast point is a geocoded city centre, not a weather station.
-
We did not record the full data period for each corrected sub-test.
-
The timestamp fix for the forecast archive was an exclusion, so there is no corrected "same-day" number.
What it means for a trader
Each trap made a false edge look significant. A backtest result worth trusting uses prices that someone traded at, selects markets using only information known at the entry time, and uses timestamps that do not include what happened afterwards. We found no edge in either test once these were applied.
Sources
- sourcePolymarket Gamma API
- sourcePolymarket CLOB API
- sourcePolymarket Data API
- sourceOpen-Meteo
Polytrading is an analytics tool. It does not trade for you, hold your keys, or promise returns.