Digital markets3 min read

The position that only looked closed

A trading log showed a position closing in under a minute. The exchange told a different story.

There is a particular kind of bug that shows up when one part of a system acts on information that hasn't caught up yet. Nothing necessarily crashes. The records can look tidy while the real work is somewhere else entirely.

Updated November 5, 2026: This article was corrected to reflect that the entry filled and the first close was an accounting record.

I ran into one in a trading system I was working on. The logs said a position had opened and closed in under a minute. The exchange told a different story. This account draws on the internal investigation; the sources below provide general context on asynchronous systems.

The symptom

Some trades were being recorded as a natural close within a minute, even though the target was nowhere near.

That label sounded routine. It wasn't. The first close was an accounting record, not an order that actually closed the position. The fee on that record was an estimate, so it couldn't by itself establish what the exchange charged.

What was actually happening

The system placed a maker-only entry order and registered the position while the order was still resting. In the case traced through the logs, the order then filled after a reprice. The exchange position update arrived later.

In that gap, the position manager checked its snapshot, didn't see the position and concluded it had closed. It removed the position from its own tracking and recorded a natural close. No closing order was sent at that point.

When the actual position appeared in the snapshot, the system saw it as untracked and adopted it into another tier. That tier later closed the real position. The phantom close record was not evidence of a separate completed trade.

Why the label mattered

Natural was a reassuring name for an uncertain result. It made a missing position look like the normal end of a trade, even though the system had not confirmed what happened.

That is the part I keep coming back to: a default value is a claim about the world. If a system has not established the outcome, the record should leave room for that uncertainty instead of making it look resolved.

The kitchen version

You call a dish away. One station hears you; another doesn't. From the first station's point of view, the ticket is moving. But the plate still can't go out because a component was never started.

A callback helps: the receiving station repeats the call. The handoff is not complete just because the message was sent. Someone needs to confirm what happened.

The software problem was different in its details. Here the entry filled, but the position update lagged behind it. The common lesson is to distinguish a missing update from proof that the work is finished.

What changed

The fix brought three existing safeguards from sibling tiers into the swing tier: a check that the position had been observed on the exchange, a longer grace period and a check for pending entries before treating a position as closed.

The investigation recorded the fix as deployed and set out checks for subsequent swing opens. That plan is not, by itself, proof of how every later trade behaved. The useful test is whether the short phantom closes and related adoptions stop recurring in fresh records.

When two parts of a system hand work to each other, keep track of the state and confirm the outcome before treating it as complete. A reassuring label is no substitute for that confirmation.