Betting against the crowd on Polymarket

On a two-and-a-half day tape in July this contrarian fade came in at +$0.20 a trade and we called it inverse noise rather than an edge. That was the right call: it was noise. At seventeen times the data the same strategy is −$0.99 a trade and now clears the significance bar that was locked in advance, in the wrong direction. The pre-commitment did its job in both directions. It stopped us trading a positive-looking number that was noise, then killed the strategy once the number became real.

910 takes on the 45-day tape, 19.9% win rate (confidence interval 17.4 to 22.6), −$904.68 net, −$0.9942 average per take, standard deviation $14.88 per take, t=−2.02. The pre-committed criterion was an absolute significance statistic at or above 2 in either direction, so the significant-loss branch fires and the fade is killed on its own written criterion rather than on a judgement call after the fact. For scale on the same tape, the no-skill baseline across 50,609 takes is −$0.8101 a take at t=−17.06, so an unskilled trader bleeds steadily here; the fade bleeds faster. The July reading was +$0.20 a take at t=+0.16, which is why it was not tradeable then either.