Week three: 1342 round trips, two rollbacks, and a lesson in humility

The week ran from September 28 to October 4, our eighteenth day live, and it was a week of the market doing what this pair does best: going back and forth. Over 168 measured hours the pair wandered across a range of 200 ticks with a net drift of minus 129, which is a polite way of saying it slid slowly while wiggling the whole way. Sixty-three of those hours were quiet enough that the range never left six ticks. That mix, drift with wiggle, is exactly the terrain a quoting desk likes, and the numbers show it: 1342 round trips, 820 of them won, 247 break-even, 275 lost, a 61.1 percent win rate, and a median hold of just thirty minutes per lot.

The rhythm of the week was familiar. Monday was the busiest day at 316 round trips, and the pace eased toward the weekend, down to 93 on Saturday and 89 on Sunday. Quiet hours are harder to work with, so the autopilot spent much of the week testing ideas aimed specifically at them: adaptive spacing, the volatility brake, the exit and break-even timers, all measured on paper next to a control copy of the desk running the old settings.

The interesting story is what the autopilot did with its confidence. Early in the week it promoted two changes to live: adaptive spacing with its reference value, and a new lot size. Both had looked good in testing. Both were then judged against their control copies in real conditions, and both were rolled back. Seven other candidates, including the sell-target distance, the number of resting buy orders, and the float of lots quoted above the market, never made it past the control at all. Nine tests started, seven decided, zero survivors. That is not a failure of the system; that is the system working. A paper win that melts when it meets the live market is exactly the kind of thing you want caught by a rollback instead of discovered by a bad week.

Meanwhile the machinery hummed along at 99.7 percent uptime, sending 6690 orders with only 13 refused. The guardian found 25 issues during the week and every single one resolved on its own, with a median of seventeen minutes; five were cured automatically and all 25 were diagnosed by the AI analyst, which also wrote 33 insights. The one thing worth remembering from this week: proving a change wrong cheaply is a win too.