built a bot for polymarket temperature markets. two months. finished measuring it this week and it doesn't work
market beats my model by 34% on RPS. i assumed that meant my data was worse, so i pulled NOAA's NBM, which is the free bias corrected forecast the serious people apparently use
market beats that as well. p = 3.8e-9 across 255 city days
then the part i didn't expect. my own model beats NBM. so the modeling was never the problem, which somehow felt worse than if it had been
two things left after that
the METAR floor. live observations that physically rule bins out, since the max can only go up. by the time i see it, 96.7% of those bins are already trading under 0.02. the market is just there first
maker instead of taker to skip the 5% fee. 0.7 to 1.1 cents lost per filled share, 462 days, every CI under zero. not close
all of it paper. never put money in
the part im keeping is the test setup. 2400 tests, mutation testing, calibration that returns nothing when n is too low. thats what stopped me on aug 14 when 4 of 5 criteria came back green and i wanted to ship it
still don't know what the market knows
#polymarket #predictionmarkets #quant