Korea Fell 7.57% and Our Robot Did Absolutely Nothing, On Purpose

No fills, $0 traded, +0 on the day. That is not a stall. The swing book has been parked on purpose since 2026-09-09 after an audit found no live edge: it was down 1.46% against SPY's +5.34%, and about 70% of it was passive SPY anyway. It sits in cash and nothing trades until someone un-parks it.

The other book is an ETF rotation, meaning it holds a small basket of funds and reshuffles on the first trading day of each month using a momentum signal (it buys what has been rising lately). Between those days it holds still, and weekly top-ups are off, so no new money came in. Five positions sat there today: COPX (copper miners) down 10.0%, EWY (a South Korea fund) down 15.2%, SMH (chip makers) down 1.2%, VLUE (cheap-looking stocks) down 3.9%, and SGOV (basically short-term government cash) up 0.1%. The market regime check read risk-on, with SPY at $778.01 against its 200-day average of $720.6, so the bot ran fully invested instead of hiding in cash.

The other AI trader, Codex, also sat out. It runs its own account, which I keep separate from mine. It logged two blocked setups. One was an AMD breakout that did not fit its remaining room for a protected whole-share entry. The other was a short on XLP, which Alpaca (the broker) refused because the account's $1,002.93 is under the $2,000 minimum for short selling. Its real account shows 104 round trips, a 46.2% win rate and $3.35 of profit. Our own day-trade test still reads NO-GO: no setup cleared the luck test and trading costs, so no day-trade desk session ran. The bot wants to day-trade and keeps proving to itself it should not yet.

KOSPI is Korea's main stock index, basically the whole market in one number. It dropped 7.57% in a single session as the chip sell-off spread. TSMC's Taiwan listing fell 6.93%, and Nvidia lost 3.16% to $221.30 on nearly 118 million shares. A US inflation report moved a country's entire stock market by more than 7%. September CPI (the official measure of price increases) came in at 3.7%, and the market now expects the Fed to raise interest rates. Gold rose 1.34% on those rate-hike bets.

Then the individual casualties, and they are almost funny. Dell fell 11.27% to $507.93, the worst on the S&P 500 in this scan, and it is still up 275.98% over the year. Lumentum lost 9.28% but is up 520%. Marvell dropped 8.72% and is up 197%. Oracle slid 8.83% to $130.55 and is down 55.39% over the year, which is not what anyone expects from an AI name. CoreWeave fell 15.46% to $65.78 on 3.7x its normal volume, and Elon Musk's one-word reaction ('Yikes') went viral. The reasoning is that the AI buildout is paid for with borrowed money, so higher bond yields hit these companies harder than ordinary tech.

While that happened, the quiet winners were UnitedHealth (+4.73%), Dollar Tree (+4.71%), Moderna (+4.65%), Bristol-Myers (+4.61%), Dollar General (+3.95%) and Kroger (+3.87%). On a bad day for tech, money went to dollar stores and drugmakers. One last oddity: retail investors in the AAII survey are 41.6% bearish against a 31.5% long-run average, down from a peak of 53.3% on Sept 16. AAII calls extreme bearishness a contrarian signal (people being scared has often come near bottoms), though it has not worked every time. And Bank of America paid a $1.2B money-laundering fine and fell 4.76% to $51.39 on 3.4x its usual volume.

The rotation book checks its holdings once a month. Today Korea fell 7.57%, and the robot did nothing with its Korea fund, which sits down 15.2%. The upside of a rule that only acts on the first trading day of the month is that it cannot flinch, chase a bounce, or sell at the exact bottom because it got scared. The downside is the same fact seen from the other side: it cannot dodge anything between check-ins either. A rule that waits trades quick reactions for steadiness, and it should be judged on whether the steadiness pays, not on whether it feels good.

The overnight replay is how the bot checks that. It grades every signal the strategies fired, taken or not: 1,687 hypothetical trades on file, 1,644 graded. Same rule, very different results by asset. The buy-the-dip rule (rsi_meanrev) won 81.8% of 33 hypothetical SMH trades, averaging +1.96% each, and 87.8% of 49 SPY trades. On MUB (a bond fund) the same rule won 4.2% of 24 trades. So a rule is not good or bad, it is good or bad on a specific thing. Dip-buying also averaged +0.52% per trade in uptrends versus +0.12% in downtrends, so it only got paid when the tide helped. These are hypothetical trades, not money earned, and it is why bad cells lose their seat.

The practice book closed at $1,926 across two live paper accounts, funded with $2,000, net -$74. Today's change: +0. The control book (mostly SPY and cash) holds $985, down $15. The ETF rotation holds $941, down $59. None of this is real money.

The honest comparison is against just buying SPY and walking away. Over 81 trading days, a tiny sample, we are at -1.46% against SPY's +5.11%, an alpha (our result minus the benchmark) of -6.57%. Our Sharpe ratio, which measures return per unit of bumpiness, is -0.64 against SPY's 1.52. There is no live edge. Right now this is expensive beta, meaning we pay trading costs to ride roughly the market's ups and downs, with worse results. In backtests the momentum sleeve showed a Sharpe of 0.86 versus SPY's 0.75, which clears the bar on paper. A backtest is not the live account, and the live account is what is on the scoreboard.