[Oct 22] First Live Day of Early Entry, and a Two-Tier Structure Stalled on Funding

Early entry — freezing the ranking ahead of time and pulling the first trading round forward — went live, while yesterday's proposed two-tier structure stalled for lack of capital to fund it.

This is the English version of a post originally written in Korean for my algorithmic trading system devlog(new tab).

Early entry went live

A structure that freezes the ranking ahead of time and pulls the first trading round's execution time earlier was designed, built, and reviewed by an AI advisor overnight, then deployed this morning.

A few observation-only shadow runs went up alongside it too — an extra intraday model evaluation pass and a sub-bar auxiliary signal. Neither affects actual orders yet; they're purely for watching.

Watching from the first round in the morning through the evening, the freeze, the execution, and the follow-up rounds all stayed within expected bounds.

Yesterday's two-tier structure — the code is done, but there's no money

The execution code for yesterday's proposed "index-replicating core + small active slice" structure got finished and merged today. It's still sitting there inactive, though.

The work split into several pieces: a planner for buying the index-tracking ETF and rebalancing it quarterly, logic that aggregates holdings and cash across a group of accounts, and the actual execution entry point.

When I asked the advisor how much capital to put in and where it should come from, the answer that came back was: fund it with a fresh deposit, buy a set fraction right away, then scale up based on a year-end verdict.

In the afternoon I corrected a premise — there's no fresh deposit coming in, and no existing holdings are being sold either. That killed the proposal, and the core-plus-active structure is on hold for now. The code is already in place, though, so it can switch on as soon as funding is settled.

Six pre-registered hypotheses, all failed

The pre-registered hypothesis testing mentioned yesterday — set up specifically to prevent fitting the criteria to the results after the fact — actually ran today. None of the six hypotheses showed a statistically meaningful signal.

Looking at the results from another angle turned up a bug in one of the implementations. Fixing it and rerunning didn't change the conclusion, and on top of that, the testing design itself was judged too loose to reliably catch an effect even if one existed.

So instead of testing hypotheses one at a time, the approach got replaced with a longer observation window and a stricter, pre-committed bar. The verdict on that won't come until early next year.

The turnover cap advisory had a wrong number, and I caught it

The way the trading turnover cap gets calculated didn't quite fit the early-entry structure anymore. An advisory came in this morning proposing to fold the per-round limit into a single daily limit.

I pushed back — if a large share of rejected sell orders apparently turned into losses, why not just remove the cap entirely? Asking the advisor to recheck turned up a real error: some of the trades blamed on the turnover cap had actually been rejected by a different limit (the per-stock cap), not the turnover cap at all.

Once the attribution got sorted out, the recommendation narrowed to: don't remove the cap outright, fold it into a daily limit, and raise the cap itself somewhat. I approved that version and it's now applied, taking effect early next week.

Why this matters

The thread running through today was keeping "the code is ready" and "the conditions to run it are ready" as two separate questions.

The core-plus-active structure had its code finished but stalled for lack of funding. The hypothesis testing didn't get forced through just because it failed to produce a result — it went straight into a redesign instead.

The turnover-cap episode is a reminder that an AI advisor can be wrong too. Catching the bad attribution only happened because the numbers got pushed back on instead of taken at face value.

Early entry went from design to deployment in a single night, but what actually mattered was watching today's first live round all the way through to confirm it behaved as expected.