eagL / scanner failure modes

The failure modes of prediction-market arbitrage scanners

Arbitrage scanners rarely lie by inventing prices. They lie by omission: each failure mode below is a real cost the scanner declined to model, and each one alone is regularly larger than the spread being advertised.

1. Screen prices for real fills

The headline spread assumes you fill the entire trade at the displayed top of book. The displayed level holds for a handful of contracts; past it the price walks against you level by level. Any scanner quoting edge without a size attached is quoting the edge available to roughly one contract.

2. Flat fee haircuts

Venues do not charge a flat percentage. Kalshi's fee is a formula in price and size, rounded up to the cent. PredictIt takes its cut from profit at settlement, so it scales with how right you were. Coinbase charges taker basis points on notional. A scanner that subtracts a uniform 2% gets every one of these wrong, in different directions, on the same trade.

3. Mismatched contracts

The most expensive false positive is a pair that is not the same event: strikes that differ, settlement sources that differ, expiry times that differ by a day. Title text matches; the contracts do not. This is why matching needs a numeric gate in front of the text similarity, so a $100k strike can never pair with a $150k one no matter how similar the words are.

4. Single-leg thinking

A cross-venue trade is two fills or it is nothing. Legging in and watching the second side move away converts an arbitrage into an outright position you never chose. Pricing the pair honestly means asking whether both quotes were still there, simultaneously, after detection, which requires replaying stored quotes rather than trusting the snapshot that generated the alert.

5. Survivorship in the alert log

Scanners grade themselves on the alerts that worked. The honest denominator is every alert fired, evaluated at executable prices, over a fixed window declared in advance. Anything else is a highlight reel. This is also the humbling one: run the honest accounting and most days end with every candidate rejected, which is what a correct scan of an efficient-enough market looks like.

The constructive version

Each failure mode has a fix, and together the fixes are just the method: walk the book, charge each venue's own formula, gate the matching numerically, replay both legs forward, publish nothing below minimum samples. That method is specified on the methodology page and applied in the worked example on the arbitrage page.