Context so you know I'm not just fishing for free alpha:
I built and ran BTC/ETH latency arb, buy the correct side of the 5/15-min up/down binary in the ~2.7s window before Polymarket reprices to a confirmed Binance move. No forecasting; just reading price action faster than the book updates.
Why I shut it down, on measured numbers:
➡️Directional win-rate decayed to roughly coin-flip 50.8% to 56.3% across days of several hundred resolved trades each.
➡️Stratified by edge, only the 7–10% raw-edge band held up (WR 54.2%, Wilson 95% CI [51.8, 56.5]); the 5–7% band was net-negative per trade.
➡️The Jan-2026 taker fee ≈ 0.07·p(1−p) per share bites hardest at p≈0.5 (~3.5% of price per leg, ~7% round-trip) exactly where the strategy lived. The resolution leg flips net-negative after fees.
➡️Net: execution edge, not a forecasting one. Once the market caught up to the public Binance reference and the fee landed, there was no durable directional alpha left for a non-colocated retail bot.
So the question: does anyone have an edge that (a) doesn't decay to zero as competitors arrive, and (b) survives the fee, shown over a real sample with confidence intervals, not a green screenshot?
Genuinely curious about: complement / pair-sum mispricing (YES NO < $1) on thin books, scoring-rule or resolution-source edges, maker capture with honest adverse-selection accounting. If yours is real, I'll trade my full teardown, fee math, WR-by-edge tables, exit-realiability analysis for a comparison of method.