Quoting Polymarket for the spread and the maker rewards

The thesis assumed prediction-market books carry huge gaps a patient quoter can sit inside, and that a maker-rewards programme would top up the return. Both halves failed on inspection. The five-minute books are tight, and the rewards field came back empty on every market we would have quoted, so there was nothing to offset the adverse selection a maker eats. That closed the venue on the side we would actually have traded, and it was neither a close call nor a small sample.

The pick-off backtest runs about 590,000 quote evaluations, so this is not an underpowered read. Adverse selection rises with reaction lag as expected, but the five-minute books measure about 1.4 cents wide, not the 40-point gaps the original thesis assumed, and the result holds across a tenfold range of the volatility assumption. Separately, the rewards field came back null on all four markets in scope, so reward-harvesting is ruled out rather than merely unattractive. The taker path was the wrong side of the trade in any case. The write-up exists to record a second lesson: the companion taker gate first read as a clean, decisive no-go in two independent sessions, and it was neither, it was statistically underpowered. The tool was printing a raw mean with no significance attached, which let a weak sample masquerade as a verdict. It now prints the spread of outcomes, a significance statistic and a confidence interval, and returns an explicit underpowered verdict rather than a bare number that can be mistaken for one.