Japan, AI Management

Big Week

By now you’ve probably seen many detailed recounts of the Japanese carry trade that rocked markets this week. Too many people put too much money into the same trade. The problem with overcrowded trades is there are always leveraged buyers who end up as forced sellers. That’s why panic selling looks so brutal, investors are unloading assets at whatever price people will accept because they have to.

The specifics of the JPY carry trade are pretty detailed, it’s not something you can access on Robinhood, yet… Maybe some day. It involves earning a spread between JPY interest rates at .25% and US interest rates at 5.5%. That spread gets narrower when the JPY rate increases and when the US rate decreases, and both of those things are finally happening. The JPY rate was increased .25% on Friday and the US is teetering on the edge of its first cut. Normally these JPY details don’t matter, but when too much money unwinds, it vacuums up available market liquidity and prices can flash crash. Often the assets that get hit most are the ones that have done the best and are sitting at elevated multiples like the tech sector which we documented last week. A decrease in multiples to healthier levels is fine, it’s not a crash. One issue is that markets can spook themselves into further declines, known as reflexivity (Soros). There’s probably more shocks to come in the near future as participants digest what just happened and as that spread between JPY rates and USD rates continues to shrink. Hopefully traders unwind in an orderly fashion knowing what’s coming because if there is an immediate US rate cut and or an immediate JPY hike, it will spark another flash crash. People need time to get out of that trade as there is only so much selling a market can absorb before it starts to impact every other asset.

To add to the JPY drama, Warren Buffett sold 50% of his Apple shares unwinding a significant portion of the most profitable trade in investment history. I speculated it was due to historically high Apple multiples, 33 PE vs when Buffett started buying at a 10 PE. Others say he’s taking profits now in case corporate taxes increase under a Harris presidency. Either way, it doesn’t matter, but the timing was not great as it gives CNBC more fodder to stir up emotions during times when volatility is elevated.

AI In Investing

As some of the AI momentum starts to take a breather, I thought it’d be a good time to write about my experience with AI’s impact on investing to date.

It’s not worth doing your own AI modeling unless you have a top tier team and the ability to dedicate 100% of your time and focus solely to that activity. The competition and the stakes are far too high to make it a casual endeavor. That said, I’ve tested and researched models that other people have created, some paid and some free. So far, this is what I’ve seen.

• Certain summaries are starting to get very good. You can feed the same earnings report into Meta.ai, Claude, ChatGPT, Gemini and Perplexity and get vastly different key takeaways. Gemini consistently gives the least helpful answers. It tends to apologize about the difficulty of the question and include many hedging statements. There are paid and free tools that are optimized specifically for earnings reports and market related news. Just like people, not all AI’s are good at everything, but the ones that are good have become surprisingly effective at sorting, filtering and summarizing especially when they provide links to the source material.
• Complex dynamic exponential equations. This is the biggest problem. Markets are complex (of course), dynamic (relationships are always changing) and exponential (highly sensitive to both extremely large and extremely small numbers). The best analogy is weather forecasting, where famously a tiny change in the model changes the local prediction from a sunny day to hurricane winds. The problem with the models that I’ve used so far is, just like the weather models, they often arrive at dramatically different conclusions. What’s more interesting is when there’s consensus among models that tend to disagree. It’s also interesting to see how models react to new data which can sometimes dramatically change the opinion (the delta). They are also optimizing for different priorities like short term technical signals or certain attractive fundamentals.
  The better news is these models almost always have some layer of visibility into their supporting data so you can see how stock specific data contributes to a particular ranking. Unlike a lot of AI experiences, at least you generally know why the models like what they like.
• All markets. Equities are the low hanging fruit for AI models because they are large, volatile and available, but they’re just one asset class. The entirety of the investment opportunity set includes many other asset classes and their derivatives markets which are equally if not far more important from a trading perspective. There can always be a better risk and reward trade somewhere else that the model is not trained to look. Until we have AGI ingesting all information, AI will be used to analyze narrow bands of market opportunity like a flashlight beam sweeping through the dark looking for an edge. That makes it potentially highly potent in some areas but not something that’s anywhere close to generally “solving” the market in its entirety.