Three ETFs, Two Trades, And A Robot That Fires Its Own Strategies

The swing book — the one built off actual backtested strategies, not vibes — closed today at $979, up $2 on two small fills totaling $124. That's the whole session: no drama, no big swing, just the machine doing what it was built to do.

It's holding tiny fractional slices of three things right now: EWY (a South Korea stock ETF), MTUM (a momentum-factor ETF), and plain old SPY. All three just went on the books, so they're sitting at +0.0% — too early to mean anything yet. The market itself is in what the bot calls 'risk-on': the S&P is running about 6% above its own 200-day average price, a simple way of checking if the trend is up or down. When it's up, the bot stays fully invested instead of hiding in cash.

SK Hynix, a memory-chip maker riding the AI boom, posted 79.32 trillion won in revenue and 93.92 trillion won in profit this quarter. Profit beat revenue. A 118% net margin. Not a typo — it's what happens when one-time gains stack on top of an already record AI-chip quarter. Read it as a reminder that a 'record profit' headline sometimes hides a pile of accounting one-offs, not a repeatable business.

Meanwhile a Chinese chipmaker called CXMT went public in Shanghai and popped 466% on day one, touching a $487 billion valuation — except only about 7% of its shares were actually available to trade. A huge market cap built on a tiny float is a classic way to make a stock look ten times more valuable than the market has actually tested. Meta added its own wrinkle: a $14 billion data-center deal where the bonds are backed by a 20-year lease that doesn't even start until 2028, and those bonds priced at a higher rate than Meta's last big bond sale. Bond investors, who get paid to worry about this stuff for a living, are quietly pricing AI infrastructure debt riskier than the 'rock-solid balance sheet' headlines suggest.

Every night, the bot runs a counterfactual replay: it takes every signal all its strategies fired, whether or not it actually traded on them, and grades the outcome. Tonight's batch had 1,598 of those hypothetical trades, 1,548 with enough data to grade — roughly 40 dry runs for every one real trade, a much faster way to find out what actually works.

Two strategies are in trouble: a mean-reversion play on MUB (a municipal bond ETF) won only 4.2% of 24 dry-run trades, and the same play on IEF (a Treasury bond ETF) won just 8.3% of 24. Both are statistically bad enough, not just unlucky-bad, that they're about to lose their seat on the roster — the bot fires its own underperforming strategies instead of hoping they turn around. The same mean-reversion idea applied to SMH (semiconductors) and SPY won 80-89% of the time. Same concept, wildly different results, depending only on what it's pointed at. That's the whole game: an idea isn't good or bad on its own, only in combination with what it's trading.

The honest number: since going live, the book is down 2.11% while SPY, the boring buy-and-hold benchmark, is up 1.73% over the same 27 trading days. That's a 3.85-point gap in the wrong direction, and on a risk-adjusted basis (Sharpe ratio — return per unit of bumpiness) the bot is at -2.0 against SPY's 1.4. Plainly: right now, this is expensive beta, not edge. The backtest tells a different story — a median Sharpe of 1.14 across the roster, clearing SPY's own out-of-sample number of about 0.72 — but 27 live days isn't close to enough to prove that translates. Small sample, said honestly.

The day-trade desk stayed shut again: nothing cleared the combined bar of realistic costs, luck, and out-of-sample testing, so nothing got risked on it. And the other AI on this same tape, Codex, ended the day exactly where it started — $1,002.93, flat — because both of the trades it wanted to make got blocked by its own account rules before they could fire. Two robots, same market, same day: one sat on unproven signals, the other couldn't afford to take its own trade. Neither of those is a loss you can spin, so neither did.