Sniping "stale" quotes after a spot move
Bitcoin jumps on the exchange; the prediction market's quote hasn't moved yet. Buy the lagging side before it reprices — the classic latency trade. The quote wasn't stale. It was right.
The hypothesisPrediction-market quotes lag the spot exchange by seconds. Stream exchange trades; on a sharp spot move, buy the candle side the move favors at its not-yet-repriced ask. A pure speed trade — the bot-taxonomy articles call it the "repricing bot."
The test
A live exchange trade stream detected sharp moves within seconds; on each trigger the probe snapshotted the market's book (measuring our own trigger-to-book latency at ~230ms), logged the favored side's executable ask, then re-polled that ask at 1, 2, 5, 10, and 30 seconds to time the repricing, and settled every entry at resolution. Entries required the quote to be genuinely "stale" — still at its pre-move level — so the probe bought exactly what a latency bot would buy.
The result
−16.8% ROI over 373 settled entries. The diagnostic that explains it: thirty seconds after entry, the market's mid sat on average below the ask we had paid — the "lagging" quote had not been slow, it had been correct, and the spot move was part noise. The only profitable slices (entries that repriced within 10 seconds, +15% on a small subsample) are identifiable only in hindsight — at decision time they look identical to the losers. Fifteen-minute windows were the worst stratum at −61%.
Conclusion
The latency-arbitrage story assumes the market maker is asleep; in practice the maker's quote already prices mean reversion in the spot move itself. What looks like lag is the book correctly refusing to chase noise — and the entries a "stale-quote sniper" can actually get filled on are adversely selected by construction. Speed was never our deficit here: we measured ourselves at ~230ms and still lost. The trade is structurally, not technologically, broken.