Data artifacts

Stale quotes, ticker collisions and phantom gaps — what ARBHUB filters out, and what can still slip through.

Why screeners hallucinate

A cross-exchange screener compares numbers produced by eleven different systems with different symbols, conventions and update cadences. Most of the spectacular spreads such a comparison produces are not opportunities — they are artifacts of the comparison itself. Knowing the artifact classes is the difference between reading the matrix and being read by it.

The artifact classes

• Stale quotes. A price that stopped updating minutes or hours ago, compared against a live one, produces a clean-looking gap that is actually just elapsed time. Classic sources: dead REST feeds on tail symbols, degraded venue APIs, halted markets that still publish the last print.
• Synthetic prices. When a venue’s ticker carries no real bid/ask, a mid derived from the last trade stands in — on a thin book that number can be far from anything executable, and spreads computed on it are fiction.
• Ticker collisions. The same symbol naming different tokens on different venues. One venue’s ON is not another venue’s ON; comparing them prints a permanent triple-digit “spread” that no trade can capture.
• Unit-mismatch contracts. 1000X/kX conventions — a contract quoted per 1,000 tokens next to one quoted per token. Without normalization the price gap is a factor of ~1000; with sloppy normalization it is a subtle few percent.
• Stock and index perps. Tokenized-equity perps price against a market that is closed half the day. A crypto-hours venue drifting against a closing print produces multi-percent phantom gaps every evening.

What ARBHUB filters

• Prices carry freshness stamps; quotes older than the staleness threshold are flagged stale and, in the default clean mode, excluded from a row’s reference math rather than compared.
• 1000x/k contract conventions are normalized to per-token prices before any comparison.
• A cross-venue median price gate drops venues whose price deviates several-fold from the rest of the market for that symbol — the ticker-collision signature — and a curated exclusion list removes known same-name-different-token venue listings.
• Rows whose best pair rests on flagged quotes are hidden in clean mode; turning clean mode off shows them with explicit badges instead.

What can still slip through

Filters bound the problem; they do not close it. Near-priced collisions (two tokens that happen to trade close), the first minutes of a feed going stale, thin-book mids inside the plausible range, and closed-market drift on stock perps can all pass every automated check. The manual verification loop stays mandatory: hover the cell, compare venues, open the pair, and look at both live order books before believing a number. If the books do not show the gap, the gap is not there.