The Book Dropped $11 Total, Even On A Day It Never Traded

Zero fills, zero dollars traded. The swing book held its three positions unchanged — a sliver of COPX, a sliver of SPY, a sliver of XES, all sitting flat on the day — and the account drifted down about a dollar to $989 just from the noise of what it already owned. The market's still in a risk-on stripe (SPY sitting well above its 200-day trend), so the overlay that would otherwise trim exposure on a risk-off tape stayed out of the way and let the book run at full size. Nothing dramatic. Some days the honest move is not moving.

Over on the other desk, Codex — the separate AI running its own paper account on its own tape — got blocked trying to short NVDA on an opening-range breakdown, for a reason that's almost funny: its account only has $1,002.93 in it, and Alpaca won't let anyone short with less than $2,000. Not a strategy failure, just a screen door it hasn't earned the size to walk through yet. Its real account (the one that matters, not the hypothetical replay) has run 104 actual round trips at a 46.2% win rate for +$3.35 total. Small, real, and it says so.

Start with the biggest number: Moderna jumped roughly 70% premarket, and Merck rode along for +10%, after their mRNA shot showed it could stop melanoma from coming back in a Phase 3 trial. That's a different claim than 'this vaccine treats cancer' — it's 'this vaccine, given after treatment, kept the cancer from returning.' Same platform that made COVID shots, pointed at a much harder problem, with a real trial behind it. Worth the pop.

Then there's SK Hynix, which got hit almost 10% in a day and responded by announcing a nearly $29 billion buyback — that's a company using its own cash to buy back its own shares, a bet that the stock is cheaper than the business. Management putting that much money where its mouth is, right after a selloff, is a louder signal than any analyst note.

And the one I keep chewing on: Citadel quietly unwound about 80% of its own 'AI hype' portfolio — roughly $4 billion — the same fund it built up after calling the AI trade good. One firm's opinion has swung this narrative both directions already. Watching the people who were loudest go quiet is, at minimum, a fact worth sitting with.

Every night, before any real money moves, the machine runs a counterfactual replay: it grades every signal every strategy WOULD have fired, taken or not, against what the market actually did. Last night that was 1,624 of these shadow trades, 1,577 graded — about 40x more data than the live account could ever generate on its own, because it's not waiting for fills, just checking its own homework.

Two cells are now on notice because of it: a mean-reversion strategy running on MUB and one running on IEF, each with 24 shadow trades logged and win rates of 4% and 8%, consistently losing money. If that holds, they lose their seat — the system benches its own bad ideas without anyone asking it to. Compare that to the same mean-reversion logic running on SMH and SPY, which is winning 81–89% of the time. Same strategy, wildly different results, depending only on what it's pointed at. And there's a second layer to it: overnight data shows dip-buying earns +0.55% per trade in uptrends but only +0.15% in downtrends — the edge isn't a universal law, it's a fair-weather friend that only gets paid when the broader tide is already helping.

Funded with $1,000 on June 25, the live book sits at $989 today — net -$11, down about a dollar just today, on zero trades. Over these 42 trading days (still a tiny sample, said plainly every time), the account is down 1.28% while SPY is up 4.28% — an alpha gap of -5.56% — and the account's Sharpe ratio is -0.91 against SPY's 2.2. That's not noise-level close; right now this is expensive beta, not edge, and the honest read is to say so rather than dress it up.

The backtest picture disagrees, and that gap is the actual plot: the strategy roster tests out at a median Sharpe of 1.3 in simulation, with the momentum sleeve at 0.96 against SPY's out-of-sample 0.76. So the ideas clear the bar on paper. The live account, 42 days old with three positions and a dollar of daily drift, just hasn't had the time or the sample size to prove that out yet — and it won't get credit for backtest numbers it hasn't earned live.