Had another AI advisor stress-test worries about the network-watchdog program — both turned out to be unfounded — and caught one more real safety-layer bug on the live-trading prep side
This is the English version of a post originally written in Korean for my algorithmic trading system devlog(new tab).
Today felt close to the final stretch of live-trading prep. Several different threads of verification wrapped up at once.
Two suspicions I'd been carrying, both turned out wrong
Yesterday I wrote about the network-watchdog program(new tab). Today brought a bit of a twist.
I asked a separate AI advisor to give the program a full review before trusting it in production. Two suspicions the assistant helping with development had been carrying both turned out to be wrong.
The first was "what if a reboot damages an analysis job that happens to be running." It turned out to be the opposite — that job was already designed to resume from where it left off, so a reboot actually helps rather than hurts.
The second was "what if there's only one channel for checking network status, and everything breaks together if it fails." It turned out there were already three independently-checking channels, so one going wrong wouldn't take the others down with it.
Both worries turned out to be unnecessary. I decided to keep the current level of alertness as is.
A safeguard meant to prevent double orders almost caused one
On the live-trading prep side, a code audit turned up one more real bug — this time a case where the safeguard itself nearly worked backwards.
Because the same account gets used for multiple purposes, I need a marker that distinguishes "did I place this order or not." One spot in the code that was supposed to leave that marker was skipping it, which meant the "list of orders I placed" was always coming back empty.
Judging "how many orders have I placed" off an empty list meant the safeguard existed but couldn't recognize orders it had already placed — creating real risk of a duplicate order. I fixed the two spots that were supposed to leave the marker.
Logged the reason a number suddenly changed
After changing how data gets fed into the model recently, I noticed the distribution of the resulting scores had shifted noticeably.
I dug in worried something had broken, but it turned out to be a natural consequence of the method change. The problem was that months from now, when someone looks back at this log, nobody would remember why things suddenly changed at that point.
So I added something that permanently records the fact itself — "the input method changed starting this date." The point is to make sure that if something looks off later, at least one candidate cause is immediately findable.
Re-measured news sentiment over two years — still no signal
I also finally closed out a question I'd been putting off for a long time: whether news sentiment signals actually help returns. This time I re-measured it against more than two years of data.
The result matched the earlier small-sample check. Effectively no relationship — and with a much larger sample this time, even the doubt of "maybe it's just not enough data" went away.
I've fully decided not to use this signal. It's a disappointing result, but I think a solid "no" is a valuable conclusion in its own right.
Also today
- Moved several background jobs the bot used to run itself (checking unfilled orders, logging balances, checking positions) out to a standalone always-on service. Pending approval requests now get written to disk instead of memory, so they survive a bot restart. Also tested that monitoring alerts actually arrive.
- An interesting calculation came out of the measurement simulator(new tab) work: capturing a very small effect (5% annualized) with statistical confidence would take roughly a century of data. A numeric confirmation that "just wait and you'll eventually find out" isn't actually an answer.
- One piece of live-trading prep remains — deliberately saved for last because it's the highest-risk part. Plan to start on it after market close tomorrow.
Looking back, today was a day where two worries disappeared, one real bug got added to the list, and one long-standing question closed for good. Altogether, the remaining distance to live trading shrank noticeably from yesterday.
Top comments (0)