USD, JPY, semis and MU

Mailbag

Two bits of feedback I received about my piece yesterday.

1. If you want to trade CAD crosses, should you not vol-weight each leg like a normal RV trader would? Yes! That’s exactly correct. If you want to do a real CAD trade vs. EUR, for example, you need an unequal amount of USDCAD and EURUSD. You can weight them using historical volatility. There are various less and more complicated ways to do this, but I simply use a spreadsheet with four years of historical data to calculate the position size in USDCAD that will spit out the same P&L (average, median, STDEV) as 10,000,000 EURUSD. The answer is 14,000,000 USDCAD. Both of those position sizes create about $50,000 of standard deviation in P&L each day.

My methodology is admittedly simple, but further sophistication usually leads to a similar answer. You can use implied vol, exponential weighting (so the recent regime gets more weight), beta-weighting, PCA / factor decomposition models, or covariance-based models. Those are all more sophisticated but again, you’ll get similar answers.

2. If I am bearish USD, is short CHFJPY a problem given its relationship to moves in the dollar? The answer is yes. And it was a mistake for me to not at least mention that when I put the trade on. The statistics are stark, although it’s worth noting that any study you do of CHFJPY from 2000 to now is going to yield a lot of bullish results because the pair has moved from 65 to 200 in that period!

That table is not great news for my CHFJPY short in the short run, but my USD view could always change, and the play was to wait for an idiosyncratic change in flows as Katayama’s urging finally makes the big pension fund repatriation story come to life. So far, not so good, but we are a few days into a two-month trade.

If GPIF and friends decide to start rotating into JGBs, we should see some evidence of that in the Japan time zone. Let’s see what it looks like so far. The “seeking to encourage GPIF to invest at home” story came out July 10.

This grid shows the hour-by-hour change in USDJPY over the past 20 days. No strong evidence of anything happening yet, though if you stare hard enough for long enough, you see more blue in NY than Japan.

The historical pattern was always that NY buys USDJPY and Japan sells, but Japan’s selling flatlined after COVID due to a combination of factors: Oil importers buying more USD due to higher oil prices, NISA expansion, Fed hiking rates much faster than BOJ, and more. This next chart shows historical performance by time zone. I drew red lines on the Japanese time zone to show you the changing slope of USDJPY price action in Japan over time.

---

Do or die time for Semis

It’s easy to be bearish semis right now as the bears are in control, new issues are flooding the market, and the boom and bust nature of the sector makes it fun to short. There are pretty much two theories:

1. The traditional-cycle hypothesis is that AI has created a new and more powerful version of the same old semiconductor boom. High prices and shortages encourage producers to expand capacity, while customers double-down and build inventory to protect themselves. Eventually, supply arrives just as spending slows, AI returns disappoint, or heavily-indebted technology companies become less willing to fund new data centers. Memory is an interchangeable product, so even a small surplus can cause prices and profits to collapse. Under this interpretation, claims that “it’s not cyclical anymore” are, ironically, a classic late-cycle signal themselves: investors are extrapolating peak demand and peak margins too far into the future. The cure for high prices is high prices. All these new fabs will finally get built in 2027 and 2028 and they will be selling semis and memory into a massive glut.

Or

2. The structural-boom hypothesis says AI has permanently changed the memory business, or at least changed it for many years to come. Large models require enormous high-bandwidth memory and conventional DRAM, while hyperscalers are committing hundreds of billions of dollars to data centers. They are signing multi-year supply agreements. Because memory capacity is expensive and slow to build, demand will stay ahead of supply for years. People on this side of the fence believe the old memory cycle has been weakened: longer contracts, more customized products, and persistent AI infrastructure spending support strong demand, stable pricing, and high margins for SK Hynix, Micron, and Samsung.

Honestly, I feel like #1 is the easiest one to believe due to history and the tendency of humans to keep on repeating the same mistakes over and over, but #2 makes decent sense to me, too. To believe #2 certainly requires more imagination than a belief in #1. I’m not a medium-term trader or long-term investor, so what I want to note here is that if you are a believer in #2, we are now at some incredible entry points in MU, DRAM, and other high-vol momentum footballs. Exhibit A:

Exhibit B:

Exhibit C:

Long MU with a stop below 800 is positive EV, in my opinion. I don’t generally put single name stock plays in the sidebar, but I do like this enough that if I did, I would. I don’t think the story has changed enough to justify a complete collapse yet. META earnings will be a big tell as they are the most likely firm to be first mover and say something like “we have done enough capex, guys!” But that’s not until July 29, so there is time for a bounce in semis and memory first.

---

Final thoughts

1. Yesterday, EURUSD followed rate differentials with a 5-hour lag. People just absolutely hate euro still and have to plug their noses before buying. I can find clients who will buy CAD, MXN, GBP, or JPY calls. But not EUR.

2. GBPUSD is working well so far. My plan is to hedge 30% in the 1.3600/40 area and let the rest run.
3. The most disappointing and concerning aspect of the USD selloff so far is that gold and silver refuse to participate. That makes me a tiny bit nervous. It seems increasingly clear that Warsh went peak hawkish at the worst possible time vis-à-vis the data and Waller piled on and now yields have turned in their face and gold and silver should be rallying on that. But they are not. Hmm.

Have a ridonkulous day.

I am beefalo

Developed to combine the cold-hardiness and lean meat of bison with the docile temper and high-yield meat of cattle, a registered Beefalo is specifically bred to be 3/8 bison and 5/8 bovine.

Estimated global populations:

Cattle (cows): ~1.5 billion

Water buffalo: ~200 million

African buffalo: ~400,000–900,000

American bison: ~500,000

Beefalo (bison-cattle hybrids): ~ten thousand

good luck ⇅ be nimble