Hold time is the lever nobody tunes
A position closed in seconds and the same position closed in ten minutes are not the same trade. The gap between them is measured in multiples, not percentages, and it is the one thing you fully control.
The shape of the loss
On a live book we split every completed cycle by its cost and looked at the tail. Removing the worst one per cent moved cost per million from 177 to 43. Three quarters of the damage sat in one cycle out of a hundred.
Those cycles all look the same: a position larger than the intended clip, held for hours, closed by crossing the spread. Nothing about the entry distinguishes them. We tested six entry filters against that tail and all six failed, and two of them looked like wins only because the metal was trending, which a split by longs and shorts immediately exposed.
- The loss is not spread evenly. It is concentrated in a thin tail.
- The tail's shape is inventory, not price prediction.
- Entry filters do not reach it; exit and inventory policy do.
The inventory lock, and what it costs
The rule is simple: once a residual position exists above the venue minimum, the side that would grow it goes dark and only the reducing side keeps quoting. The position cannot exceed one clip, so the tail cannot form.
Three independent signals confirmed it works: taker exits fell from 1.42 to 0.89 per hour, runs of three or more clips fell from 7.5% to 4.1%, and live observation showed the pattern it was built to stop: three buys in a row followed by a taker exit at minus 6.6 basis points.
It has a real price, and it is honest to state it: turnover drops, because half the time only one leg is quoting. We once switched the lock off for four days on the strength of two measurements showing no benefit. The measurements were right about their own number and wrong about the question: the lock's cost is around 7.7 cost-per-million, which is invisible against hourly noise of plus or minus 29.
- The lock caps the position at one clip by construction.
- Measured: fewer taker exits, fewer multi-clip runs.
- It costs turnover, and its benefit is smaller than hourly noise, so it needs a mechanism argument, not just a measurement.
Exit policy: patient with time, impatient with direction
The instinct is to hold losers and cut winners; a maker needs the opposite discipline in a specific form. Give the exit a long passive window, so it can leave as a maker rather than paying to cross. But cut a position whose price keeps drifting against you long before that window expires.
On our gold book those two numbers are a 30-minute maker-only grace period and a 6 basis point adverse-drift limit. The previous limit of 14 basis points effectively never fired: typical adverse drift on that market runs 0.4 to 1.8 basis points, so a 14 bp ceiling was above the sky.
- Long passive exit window: leaving as a maker is much cheaper than crossing.
- Tight adverse-drift limit: a position that keeps moving against you is not going to come back on schedule.
- Check that your limits fire. A threshold no market ever reaches is not protection.
Questions
What is inventory risk in market making?
The risk carried by the position you are left holding after a fill. You did not choose its direction and you do not want it; every second you hold it, the market can move against it. It is the second-largest cost in market making after adverse selection, and the two compound.
How long should a market maker hold a position?
As briefly as the exit allows without paying to cross the spread. In practice that means a long passive window paired with a tight limit on adverse drift: patient about time, impatient about direction.
Should I cap my inventory?
Yes, and mechanically rather than by intention. The reliable form is to switch off the side that would grow the position while it exists, so the cap holds even when the market is moving fast and your attention is elsewhere.
Why did my stop-loss keep firing?
Most likely it is sized for a faster strategy. A mode that holds inventory for up to ninety minutes will show unrealised swings during that hold. In our first days running such a mode, every session with a five-dollar stop halted within hours, each breaching by a few cents, while sessions with a wider stop ran for days.